Afrikaans
Akan
Albanian
Amharic
Arabic
Armenian
Azerbaijani
Basque
Belarusian
Bemba
Bengali
Bihari
Bosnian
Breton
Bulgarian
Cambodian
Catalan
Cebuano
Cherokee
Chichewa
Chinese (Traditional)
Corsican
Croatian
Czech
Danish
Dutch
English
Esperanto
Estonian
Ewe
Faroese
Filipino
Finnish
French
Frisian
Ga
Galician
Georgian
German
Greek
Guarani
Gujarati
Haitian Creole
Hausa
Hawaiian
Hebrew
Hindi
Hmong
Hungarian
Icelandic
Igbo
Indonesian
Interlingua
Irish
Italian
Japanese
Javanese
Kannada
Kazakh
Kinyarwanda
Kirundi
Kongo
Korean
Krio (Sierra Leone)
Kurdish
Kurdish (Soranรฎ)
Kyrgyz
Laothian
Latin
Latvian
Lingala
Lithuanian
Lozi
Luganda
Luo
Luxembourgish
Macedonian
Malagasy
Malay
Malayalam
Maltese
Maori
Marathi
Mauritian Creole
Moldavian
Mongolian
Myanmar (Burmese)
Montenegrin
Nepali
Nigerian Pidgin
Northern Sotho
Norwegian
Norwegian (Nynorsk)
Occitan
Oriya
Oromo
Pashto
Persian
Polish
Portuguese (Brazil)
Portuguese (Portugal)
Punjabi
Quechua
Romansh
Runyakitara
Russian
Samoan
Scots Gaelic
Serbian
Serbo-Croatian
Sesotho
Setswana
Seychellois Creole
Shona
Sindhi
Sinhalese
Slovak
Slovenian
Somali
Spanish
Spanish (Latin American)
Sundanese
Swahili
Swedish
Tajik
Tamil
Tatar
Telugu
Thai
Tigrinya
Tonga
Tshiluba
Tumbuka
Turkish
Turkmen
Twi
Uighur
Ukrainian
Urdu
Uzbek
Vietnamese
Welsh
Wolof
Xhosa
Yiddish
Yoruba
Zulu
all right okay so welcome back everyone
all right okay so welcome back everyone
all right okay so welcome back everyone so today it's my pleasure to welcome you
so today it's my pleasure to welcome you
so today it's my pleasure to welcome you Hong who will be talking about deep
Hong who will be talking about deep
Hong who will be talking about deep learning for uh critical transations
learning for uh critical transations
learning for uh critical transations yeah thank you okay uh thank you and
yeah thank you okay uh thank you and
yeah thank you okay uh thank you and thank you for your invitation and yeah
thank you for your invitation and yeah
thank you for your invitation and yeah I'm very happy here uh to talk about our
I'm very happy here uh to talk about our
I'm very happy here uh to talk about our recent work about uh uh using uh the
recent work about uh uh using uh the
recent work about uh uh using uh the planning to predict R steing Points uh
planning to predict R steing Points uh
planning to predict R steing Points uh yeah I'm a post doctor uh from uh Tech
yeah I'm a post doctor uh from uh Tech
yeah I'm a post doctor uh from uh Tech Technical University of Munich and this
Technical University of Munich and this
Technical University of Munich and this work is actually a co
work is actually a co
work is actually a co cooporation uh study with my colleagues
cooporation uh study with my colleagues
cooporation uh study with my colleagues Sebastian uh Peter and uh Nicholas
Sebastian uh Peter and uh Nicholas
Sebastian uh Peter and uh Nicholas and
and
and yeah uh firstly I want to give a very
yeah uh firstly I want to give a very
yeah uh firstly I want to give a very brief uh intro introduction about the
brief uh intro introduction about the
brief uh intro introduction about the Tipping points especially uh especially
Tipping points especially uh especially
Tipping points especially uh especially in the
in the
in the climate uh field yeah but as made maybe
climate uh field yeah but as made maybe
climate uh field yeah but as made maybe most of uh the audience yeah you already
most of uh the audience yeah you already
most of uh the audience yeah you already know uh a lot about uh tipping points um
know uh a lot about uh tipping points um
know uh a lot about uh tipping points um but uh here I would start from the Earth
but uh here I would start from the Earth
but uh here I would start from the Earth energ Balance model this is a very uh
energ Balance model this is a very uh
energ Balance model this is a very uh very important also very famous uh basic
very important also very famous uh basic
very important also very famous uh basic uh uh uh physical uh physical uh
uh uh uh physical uh physical uh
uh uh uh physical uh physical uh conceptual model for the uh Earth
conceptual model for the uh Earth
conceptual model for the uh Earth climate um understanding and also
climate um understanding and also
climate um understanding and also sometimes for the simulation but if we
sometimes for the simulation but if we
sometimes for the simulation but if we perform the uh simulation on this energ
perform the uh simulation on this energ
perform the uh simulation on this energ balance model using different solar
balance model using different solar
balance model using different solar radiative forcing this is a the
radiative forcing this is a the
radiative forcing this is a the so-called control parameter Ro and if
so-called control parameter Ro and if
so-called control parameter Ro and if way uh change uh uh increase or decrease
way uh change uh uh increase or decrease
way uh change uh uh increase or decrease the control parameter and uh along the
the control parameter and uh along the
the control parameter and uh along the time then we could
time then we could
time then we could observe we can we will observe the
observe we can we will observe the
observe we can we will observe the abrupt transition of the earth climate
abrupt transition of the earth climate
abrupt transition of the earth climate from a from a warm warm climate St to a
from a from a warm warm climate St to a
from a from a warm warm climate St to a cold warm climate state but it's also
cold warm climate state but it's also
cold warm climate state but it's also possible uh just to tr Transit back from
possible uh just to tr Transit back from
possible uh just to tr Transit back from uh back to the warm climate but uh just
uh back to the warm climate but uh just
uh back to the warm climate but uh just along the different pathway uh but we
along the different pathway uh but we
along the different pathway uh but we notice here the transition along the
notice here the transition along the
notice here the transition along the time is uh is very uh is at just
time is uh is very uh is at just
time is uh is very uh is at just suddenly transition uh not gradual
suddenly transition uh not gradual
suddenly transition uh not gradual transition but here I would like firstly
transition but here I would like firstly
transition but here I would like firstly I would like to use a very simple uh
I would like to use a very simple uh
I would like to use a very simple uh example from the
example from the
example from the realistic uh our realistic life to uh
realistic uh our realistic life to uh
realistic uh our realistic life to uh explain the ab transition here I take
explain the ab transition here I take
explain the ab transition here I take the I I just noticed some pictures for
the I I just noticed some pictures for
the I I just noticed some pictures for the uh s greenw classier uh maybe over
the uh s greenw classier uh maybe over
the uh s greenw classier uh maybe over uh 20 uh uh 250 50 years ago uh some uh
uh 20 uh uh 250 50 years ago uh some uh
uh 20 uh uh 250 50 years ago uh some uh from some pictures over that time we
from some pictures over that time we
from some pictures over that time we already we noticed there is a lot of
already we noticed there is a lot of
already we noticed there is a lot of glassier during summer uh even during
glassier during summer uh even during
glassier during summer uh even during summer uh over the glassier uh over
summer uh over the glassier uh over
summer uh over the glassier uh over greart um Mountain Foods uh even until
greart um Mountain Foods uh even until
greart um Mountain Foods uh even until 20 years ago uh the mountain Foods uh
20 years ago uh the mountain Foods uh
20 years ago uh the mountain Foods uh there is still a lot there was still a
there is still a lot there was still a
there is still a lot there was still a lot of glassier and ice and Ice there
lot of glassier and ice and Ice there
lot of glassier and ice and Ice there during summer but until most recent 10
during summer but until most recent 10
during summer but until most recent 10 years uh this glassier uh this ice over
years uh this glassier uh this ice over
years uh this glassier uh this ice over the mountain food disappear Suddenly
the mountain food disappear Suddenly
the mountain food disappear Suddenly It's very fast transition from the ice
It's very fast transition from the ice
It's very fast transition from the ice to just to the board uh B B
to just to the board uh B B
to just to the board uh B B rocks
rocks
rocks but yeah uh and and the local people
but yeah uh and and the local people
but yeah uh and and the local people also already noticed this uh this
also already noticed this uh this
also already noticed this uh this phenomena and they really want to
phenomena and they really want to
phenomena and they really want to protect uh this uh beautiful uh glassier
protect uh this uh beautiful uh glassier
protect uh this uh beautiful uh glassier over the F Mountain food uh but this is
over the F Mountain food uh but this is
over the F Mountain food uh but this is picture I take I uh from my camera this
picture I take I uh from my camera this
picture I take I uh from my camera this uh the past uh September I noticed that
uh the past uh September I noticed that
uh the past uh September I noticed that the local people really tried a lot uh
the local people really tried a lot uh
the local people really tried a lot uh for example they built some damp they
for example they built some damp they
for example they built some damp they they are trying to build some uh
they are trying to build some uh
they are trying to build some uh construct some bad rocks for this glass
construct some bad rocks for this glass
construct some bad rocks for this glass here and protect them but actually this
here and protect them but actually this
here and protect them but actually this does not does not really work for for
does not does not really work for for
does not does not really work for for the uh for pro uh avoid the shrink of
the uh for pro uh avoid the shrink of
the uh for pro uh avoid the shrink of the glassier this means maybe in one
the glassier this means maybe in one
the glassier this means maybe in one generation in our whole generation it's
generation in our whole generation it's
generation in our whole generation it's really hard to recover the uh this glass
really hard to recover the uh this glass
really hard to recover the uh this glass here this is also true a truth for the
here this is also true a truth for the
here this is also true a truth for the Tipping points this also this could also
Tipping points this also this could also
Tipping points this also this could also uh happen for a lot of climate sub
uh happen for a lot of climate sub
uh happen for a lot of climate sub components uh uh components for example
components uh uh components for example
components uh uh components for example like the uh Antarctic uh ocean
like the uh Antarctic uh ocean
like the uh Antarctic uh ocean circulation also the uh Greenland SE
circulation also the uh Greenland SE
circulation also the uh Greenland SE green L land SE ice also for some
green L land SE ice also for some
green L land SE ice also for some vegetation uh like the Amazon uh Forest
vegetation uh like the Amazon uh Forest
vegetation uh like the Amazon uh Forest this such a transition a bre transition
this such a transition a bre transition
this such a transition a bre transition and E reversible transition would be uh
and E reversible transition would be uh
and E reversible transition would be uh very possible for this uh nonlinear
very possible for this uh nonlinear
very possible for this uh nonlinear system
system
system so uh that a central Target of our uh
so uh that a central Target of our uh
so uh that a central Target of our uh part of our climate people is to uh
part of our climate people is to uh
part of our climate people is to uh anticipate the Tipping points and access
anticipate the Tipping points and access
anticipate the Tipping points and access access the Tipping points risks uh on
access the Tipping points risks uh on
access the Tipping points risks uh on behind the uh a specific climate syst
behind the uh a specific climate syst
behind the uh a specific climate syst subsystem uh actually here uh I I would
subsystem uh actually here uh I I would
subsystem uh actually here uh I I would uh we notice we could have s s methods
uh we notice we could have s s methods
uh we notice we could have s s methods to uh anticipate tiing points from the
to uh anticipate tiing points from the
to uh anticipate tiing points from the current condition but uh usually uh the
current condition but uh usually uh the
current condition but uh usually uh the physical uh physical based researcher
physical uh physical based researcher
physical uh physical based researcher would consider maybe uh we could uh find
would consider maybe uh we could uh find
would consider maybe uh we could uh find find out the physical equations of a
find out the physical equations of a
find out the physical equations of a specific um subsystem and then we could
specific um subsystem and then we could
specific um subsystem and then we could recover the dynamical trajectory uh or
recover the dynamical trajectory uh or
recover the dynamical trajectory uh or the face uh face base of the
the face uh face base of the
the face uh face base of the um D the this dynamical system then we
um D the this dynamical system then we
um D the this dynamical system then we can construct such a um verification
can construct such a um verification
can construct such a um verification diagram for the for the interested
diagram for the for the interested
diagram for the for the interested system then we can directly observe the
system then we can directly observe the
system then we can directly observe the um the critical threshold for the tpping
um the critical threshold for the tpping
um the critical threshold for the tpping points like like the uh here we can we
points like like the uh here we can we
points like like the uh here we can we can find usually for a by stable State a
can find usually for a by stable State a
can find usually for a by stable State a bable equ equili uh equilibriums States
bable equ equili uh equilibriums States
bable equ equili uh equilibriums States we could find two t create uh critical
we could find two t create uh critical
we could find two t create uh critical threshold for the forcing values uh to
threshold for the forcing values uh to
threshold for the forcing values uh to then we can estimate uh if the system
then we can estimate uh if the system
then we can estimate uh if the system will approach to the uh tipping points
will approach to the uh tipping points
will approach to the uh tipping points but actually but usually this uh this
but actually but usually this uh this
but actually but usually this uh this method does not really work well because
method does not really work well because
method does not really work well because uh it's really hard to recover the whole
uh it's really hard to recover the whole
uh it's really hard to recover the whole equations or dynamics of the realistic
equations or dynamics of the realistic
equations or dynamics of the realistic system is especially like the um climate
system is especially like the um climate
system is especially like the um climate system it's really hard so uh most
system it's really hard so uh most
system it's really hard so uh most recent years uh scientists also find
recent years uh scientists also find
recent years uh scientists also find some uh alternative way it's called the
some uh alternative way it's called the
some uh alternative way it's called the critical slowing down maybe uh maybe you
critical slowing down maybe uh maybe you
critical slowing down maybe uh maybe you already heard about it and that is we
already heard about it and that is we
already heard about it and that is we can uh we can hold the we can have
can uh we can hold the we can have
can uh we can hold the we can have really have the observational time
really have the observational time
really have the observational time series of the
series of the
series of the interested uh system for example here is
interested uh system for example here is
interested uh system for example here is a
a
a Greenland
Greenland
Greenland uh uh say surface temperature uh from
uh uh say surface temperature uh from
uh uh say surface temperature uh from the P climate records and uh here for
the P climate records and uh here for
the P climate records and uh here for based on the time series um usually we
based on the time series um usually we
based on the time series um usually we could we could make a linear assumption
could we could make a linear assumption
could we could make a linear assumption when it's when the system is around its
when it's when the system is around its
when it's when the system is around its equilibrium a fix point then we can we
equilibrium a fix point then we can we
equilibrium a fix point then we can we can take a a a this uh uh this appro
can take a a a this uh uh this appro
can take a a a this uh uh this appro approximation equation then um based on
approximation equation then um based on
approximation equation then um based on the damping uh parameter Lambda we could
the damping uh parameter Lambda we could
the damping uh parameter Lambda we could estimate if it's were uh under it were
estimate if it's were uh under it were
estimate if it's were uh under it were if it were uh undergo at tipping points
if it were uh undergo at tipping points
if it were uh undergo at tipping points the Lambda D dumping Lambda dumping
the Lambda D dumping Lambda dumping
the Lambda D dumping Lambda dumping parameter Lambda when it's smaller than
parameter Lambda when it's smaller than
parameter Lambda when it's smaller than zero it means the system will be stable
zero it means the system will be stable
zero it means the system will be stable but if Lambda the increase to an
but if Lambda the increase to an
but if Lambda the increase to an increase Anor to zero it will approach
increase Anor to zero it will approach
increase Anor to zero it will approach to its critical threshold then when if
to its critical threshold then when if
to its critical threshold then when if Lambda continue increase this also means
Lambda continue increase this also means
Lambda continue increase this also means the external forcing will increase and
the external forcing will increase and
the external forcing will increase and the system could be uh unstable will be
the system could be uh unstable will be
the system could be uh unstable will be unstable and the Tipping points will
unstable and the Tipping points will
unstable and the Tipping points will happens but usually for the Lambda
happens but usually for the Lambda
happens but usually for the Lambda estimation we could make use of some
estimation we could make use of some
estimation we could make use of some statist statistic Matrix like the uh
statist statistic Matrix like the uh
statist statistic Matrix like the uh Auto coration of the time serious
Auto coration of the time serious
Auto coration of the time serious segments and also usally the variance of
segments and also usally the variance of
segments and also usally the variance of the time serious uh segments for example
the time serious uh segments for example
the time serious uh segments for example here uh based on this Greenland
here uh based on this Greenland
here uh based on this Greenland uh uh say surface temperature records uh
uh uh say surface temperature records uh
uh uh say surface temperature records uh every time when there is abrupt
every time when there is abrupt
every time when there is abrupt transition we can observe a uh increase
transition we can observe a uh increase
transition we can observe a uh increase in the
in the
in the autocorrelation and also the variance uh
autocorrelation and also the variance uh
autocorrelation and also the variance uh of the time seral
of the time seral
of the time seral this is a precursory signal uh of the uh
this is a precursory signal uh of the uh
this is a precursory signal uh of the uh bation tipping
bation tipping
bation tipping points uh of course recent years we also
points uh of course recent years we also
points uh of course recent years we also have the third way uh from the data
have the third way uh from the data
have the third way uh from the data science perspective uh based by using
science perspective uh based by using
science perspective uh based by using the de planning Naro networks we can
the de planning Naro networks we can
the de planning Naro networks we can also uh construct the early warning
also uh construct the early warning
also uh construct the early warning signal of the Tipping points and here
signal of the Tipping points and here
signal of the Tipping points and here based on the Thomas Barry's uh first
based on the Thomas Barry's uh first
based on the Thomas Barry's uh first paper about about this direction uh
paper about about this direction uh
paper about about this direction uh where we can find that uh critical
where we can find that uh critical
where we can find that uh critical slowing down indicator and both critical
slowing down indicator and both critical
slowing down indicator and both critical slowing down and the de planning
slowing down and the de planning
slowing down and the de planning indicator can make prediction for the
indicator can make prediction for the
indicator can make prediction for the bation Tipping points uh very well and
bation Tipping points uh very well and
bation Tipping points uh very well and this is uh this is good for for our
this is uh this is good for for our
this is uh this is good for for our tipping points and also climate climate
tipping points and also climate climate
tipping points and also climate climate uh prediction uh communities
uh prediction uh communities
uh prediction uh communities but uh uh most most recently um some
but uh uh most most recently um some
but uh uh most most recently um some research some scientists just noticed
research some scientists just noticed
research some scientists just noticed some different thing uh it's it's uh
some different thing uh it's it's uh
some different thing uh it's it's uh it's based on a modeling research uh on
it's based on a modeling research uh on
it's based on a modeling research uh on the ATL
the ATL
the ATL Atlantic um o ocean circulation
Atlantic um o ocean circulation
Atlantic um o ocean circulation simulation experiments usally we can
simulation experiments usally we can
simulation experiments usally we can Define some safe uh space uh like when
Define some safe uh space uh like when
Define some safe uh space uh like when we per from uh how can when we we will
we per from uh how can when we we will
we per from uh how can when we we will know we have the we know we have the uh
know we have the we know we have the uh
know we have the we know we have the uh global warming uh as a external forcing
global warming uh as a external forcing
global warming uh as a external forcing to uh to the ocean circulation but when
to uh to the ocean circulation but when
to uh to the ocean circulation but when the external forcing increase but U if
the external forcing increase but U if
the external forcing increase but U if it's it has not reach uh a critical
it's it has not reach uh a critical
it's it has not reach uh a critical threshold then this means our uh climb
threshold then this means our uh climb
threshold then this means our uh climb system is safe is stable uh tipping
system is safe is stable uh tipping
system is safe is stable uh tipping points will not happens and uh the
points will not happens and uh the
points will not happens and uh the usually we can define a distin distinct
usually we can define a distin distinct
usually we can define a distin distinct uh boundary for the safe operation uh
uh boundary for the safe operation uh
uh boundary for the safe operation uh space between the safe the uh stable
space between the safe the uh stable
space between the safe the uh stable climate and also and the and the bation
climate and also and the and the bation
climate and also and the and the bation Tipping transation uh but something
Tipping transation uh but something
Tipping transation uh but something different was noticed when uh yanas
different was noticed when uh yanas
different was noticed when uh yanas Luman uh
Luman uh
Luman uh research uh when they researched on the
research uh when they researched on the
research uh when they researched on the uh aor simulation uh usually we can we
uh aor simulation uh usually we can we
uh aor simulation uh usually we can we can construct such a bable equilibrium
can construct such a bable equilibrium
can construct such a bable equilibrium State uh diagram and we can understand
State uh diagram and we can understand
State uh diagram and we can understand it using the traditional bation theory
it using the traditional bation theory
it using the traditional bation theory but um when uh when yanas uh and Peter
but um when uh when yanas uh and Peter
but um when uh when yanas uh and Peter when they perform the different foring
when they perform the different foring
when they perform the different foring uh rates uh on the AAR simulation that
uh rates uh on the AAR simulation that
uh rates uh on the AAR simulation that is the fresh water forcing from the SE
is the fresh water forcing from the SE
is the fresh water forcing from the SE Ice uh when they use different
Ice uh when they use different
Ice uh when they use different increasing rate of the forcing um
increasing rate of the forcing um
increasing rate of the forcing um specifically when they increase the
specifically when they increase the
specifically when they increase the forcing rate like uh a very fast foring
forcing rate like uh a very fast foring
forcing rate like uh a very fast foring rate uh then they notice the t uh the
rate uh then they notice the t uh the
rate uh then they notice the t uh the clim the Amar uh state would uh reach
clim the Amar uh state would uh reach
clim the Amar uh state would uh reach the uh tipping points much earlier than
the uh tipping points much earlier than
the uh tipping points much earlier than avocation tipping points this is
avocation tipping points this is
avocation tipping points this is uh this is out of out of blue and uh
uh this is out of out of blue and uh
uh this is out of out of blue and uh yeah this is caused by the uh rting the
yeah this is caused by the uh rting the
yeah this is caused by the uh rting the Tipping and uh this also could happen
Tipping and uh this also could happen
Tipping and uh this also could happen for a lot of uh other nonlinear
for a lot of uh other nonlinear
for a lot of uh other nonlinear system B basically uh around 10 years
system B basically uh around 10 years
system B basically uh around 10 years ago uh um some studies already from the
ago uh um some studies already from the
ago uh um some studies already from the mathematicians they already realized
mathematicians they already realized
mathematicians they already realized maybe there will be three kinds of
maybe there will be three kinds of
maybe there will be three kinds of tipping for the ninar systems uh
tipping for the ninar systems uh
tipping for the ninar systems uh usually the the usual one we talk about
usually the the usual one we talk about
usually the the usual one we talk about is the bation IND tipping points this
is the bation IND tipping points this
is the bation IND tipping points this means under the uh under the slowing
means under the uh under the slowing
means under the uh under the slowing changing uh external forcing the
changing uh external forcing the
changing uh external forcing the attractor Bas of the uh system would
attractor Bas of the uh system would
attractor Bas of the uh system would have a would change its shape and the uh
have a would change its shape and the uh
have a would change its shape and the uh the equilibrium state would the the the
the equilibrium state would the the the
the equilibrium state would the the the dynamical state would shift it would be
dynamical state would shift it would be
dynamical state would shift it would be shifted also uh by just with external
shifted also uh by just with external
shifted also uh by just with external forcing then it could change to another
forcing then it could change to another
forcing then it could change to another state this is bation tipping points but
state this is bation tipping points but
state this is bation tipping points but uh there is another another two two uh
uh there is another another two two uh
uh there is another another two two uh kinds of tipping that is noise IND
kinds of tipping that is noise IND
kinds of tipping that is noise IND tipping that is uh there's no very
tipping that is uh there's no very
tipping that is uh there's no very strong external forcing but just some
strong external forcing but just some
strong external forcing but just some noise perations or noise uh for things
noise perations or noise uh for things
noise perations or noise uh for things but basically the uh the trct bre at
but basically the uh the trct bre at
but basically the uh the trct bre at trct the Basin of the system could keep
trct the Basin of the system could keep
trct the Basin of the system could keep its shape and uh and just because the uh
its shape and uh and just because the uh
its shape and uh and just because the uh noise p uh noise forcing could drive the
noise p uh noise forcing could drive the
noise p uh noise forcing could drive the uh system States from one equ
uh system States from one equ
uh system States from one equ equilibrium state to another equilibrium
equilibrium state to another equilibrium
equilibrium state to another equilibrium State this is the noising tipping but
State this is the noising tipping but
State this is the noising tipping but for the rting Tipping I mentioned in
for the rting Tipping I mentioned in
for the rting Tipping I mentioned in last slide it's uh it's uh caused by uh
last slide it's uh it's uh caused by uh
last slide it's uh it's uh caused by uh the fast forcing specifically a fast
the fast forcing specifically a fast
the fast forcing specifically a fast forcing that means the when the forcing
forcing that means the when the forcing
forcing that means the when the forcing is very fast and uh comparably the
is very fast and uh comparably the
is very fast and uh comparably the attractor Basin maybe it shape has not
attractor Basin maybe it shape has not
attractor Basin maybe it shape has not been changed a lot but the equilibri but
been changed a lot but the equilibri but
been changed a lot but the equilibri but the uh System state cannot uh attract
the uh System state cannot uh attract
the uh System state cannot uh attract its original equilibrium State anymore
its original equilibrium State anymore
its original equilibrium State anymore then it can just very easily shift uh
then it can just very easily shift uh
then it can just very easily shift uh shift to another equilibrium state or
shift to another equilibrium state or
shift to another equilibrium state or just collapse this is the r IND tipping
just collapse this is the r IND tipping
just collapse this is the r IND tipping and they are the noise and R IND tipping
and they are the noise and R IND tipping
and they are the noise and R IND tipping are just out of blue of the bation teing
are just out of blue of the bation teing
are just out of blue of the bation teing diagram and uh what does this mean for
diagram and uh what does this mean for
diagram and uh what does this mean for our climate system uh we know for the
our climate system uh we know for the
our climate system uh we know for the future scenarios uh people always uh and
future scenarios uh people always uh and
future scenarios uh people always uh and the society always worry about uh the um
the society always worry about uh the um
the society always worry about uh the um global warming gas emissions and they
global warming gas emissions and they
global warming gas emissions and they already designed different uh emission
already designed different uh emission
already designed different uh emission Pathways trajectories Pathways this
Pathways trajectories Pathways this
Pathways trajectories Pathways this means uh we could have various uh very
means uh we could have various uh very
means uh we could have various uh very different forcing rates in the future
different forcing rates in the future
different forcing rates in the future scenarios and but this means uh our
scenarios and but this means uh our
scenarios and but this means uh our climate could have much more complex
climate could have much more complex
climate could have much more complex tipping risks from the noise indu
tipping risks from the noise indu
tipping risks from the noise indu tipping and R IND tipping uh diagram
tipping and R IND tipping uh diagram
tipping and R IND tipping uh diagram maybe uh this means even where we don't
maybe uh this means even where we don't
maybe uh this means even where we don't we have not reach a bation tipping
we have not reach a bation tipping
we have not reach a bation tipping points a bation tipping points threshold
points a bation tipping points threshold
points a bation tipping points threshold in the future scenario but our forcing
in the future scenario but our forcing
in the future scenario but our forcing is very fast then uh it's also possible
is very fast then uh it's also possible
is very fast then uh it's also possible for our climate lose its uh
stability uh
stability uh
stability uh so we really want to uh also make some
so we really want to uh also make some
so we really want to uh also make some prediction for uh make some improvement
prediction for uh make some improvement
prediction for uh make some improvement uh for the prediction of rting De
uh for the prediction of rting De
uh for the prediction of rting De tipping this out of blue
tipping this out of blue
tipping this out of blue tipping points and but but this is
tipping points and but but this is
tipping points and but but this is really challenging because based on
really challenging because based on
really challenging because based on recent most recent uh studies a series
recent most recent uh studies a series
recent most recent uh studies a series of studies we already realized it's
of studies we already realized it's
of studies we already realized it's really challenging firstly uh the
really challenging firstly uh the
really challenging firstly uh the dynamical equation usually the dynamical
dynamical equation usually the dynamical
dynamical equation usually the dynamical equation of our realistic system is
equation of our realistic system is
equation of our realistic system is unknown and also uh we then we we can
unknown and also uh we then we we can
unknown and also uh we then we we can really know the critical treshold holds
really know the critical treshold holds
really know the critical treshold holds for uh for the Tipping rting tipping
for uh for the Tipping rting tipping
for uh for the Tipping rting tipping secondly uh we Al uh we also already
secondly uh we Al uh we also already
secondly uh we Al uh we also already know critical slowing down does not
know critical slowing down does not
know critical slowing down does not necessarily work for uh uh for the noise
necessarily work for uh uh for the noise
necessarily work for uh uh for the noise and rating the stepping because um
and rating the stepping because um
and rating the stepping because um basically uh the critical slowing down
basically uh the critical slowing down
basically uh the critical slowing down is based on the linear uh assumption
is based on the linear uh assumption
is based on the linear uh assumption around the
around the
around the equilibrium but in this Cas in the in
equilibrium but in this Cas in the in
equilibrium but in this Cas in the in the in those tipping case this does not
the in those tipping case this does not
the in those tipping case this does not necessarily
necessarily
necessarily uh hold uh certainly and usually for the
uh hold uh certainly and usually for the
uh hold uh certainly and usually for the realistic system the rate IND tipping
realistic system the rate IND tipping
realistic system the rate IND tipping and noise IND tipping like the weather
and noise IND tipping like the weather
and noise IND tipping like the weather perations also like some uh internal
perations also like some uh internal
perations also like some uh internal varability of climate system it
varability of climate system it
varability of climate system it could the usually happens together just
could the usually happens together just
could the usually happens together just work in a concert to uh to our climate
work in a concert to uh to our climate
work in a concert to uh to our climate system uh fourthly um it's also very
system uh fourthly um it's also very
system uh fourthly um it's also very hard to General uh to uh get um get some
hard to General uh to uh get um get some
hard to General uh to uh get um get some experience from the emble simulations of
experience from the emble simulations of
experience from the emble simulations of climate models then make prediction on
climate models then make prediction on
climate models then make prediction on some individual trajectories and in the
some individual trajectories and in the
some individual trajectories and in the uh following slides I will explain this
uh following slides I will explain this
uh following slides I will explain this two the uh the third one and the fourth
two the uh the third one and the fourth
two the uh the third one and the fourth one uh first points in very
one uh first points in very
one uh first points in very detail uh what what I mean the um the
detail uh what what I mean the um the
detail uh what what I mean the um the rting de tiing here uh it's uh
rting de tiing here uh it's uh
rting de tiing here uh it's uh uh it could be explained by this three
uh it could be explained by this three
uh it could be explained by this three uh these three panels firstly when we
uh these three panels firstly when we
uh these three panels firstly when we perform very slow uh slowly changing
perform very slow uh slowly changing
perform very slow uh slowly changing forcing on the system perhaps the system
forcing on the system perhaps the system
forcing on the system perhaps the system um the uh the dynamical state as a red
um the uh the dynamical state as a red
um the uh the dynamical state as a red dots it will it it will always attract
dots it will it it will always attract
dots it will it it will always attract it's uh equilibrium State fix point and
it's uh equilibrium State fix point and
it's uh equilibrium State fix point and the Tipping points will not happens but
the Tipping points will not happens but
the Tipping points will not happens but if way shift the attracted Basin of the
if way shift the attracted Basin of the
if way shift the attracted Basin of the system very fast uh potentially the this
system very fast uh potentially the this
system very fast uh potentially the this red bar it could not um really track its
red bar it could not um really track its
red bar it could not um really track its original attractor Basin and just tip to
original attractor Basin and just tip to
original attractor Basin and just tip to another uh another state this is
another uh another state this is
another uh another state this is possible but in this two case
possible but in this two case
possible but in this two case two cases uh there's there are no noise
two cases uh there's there are no noise
two cases uh there's there are no noise there there are no noise
there there are no noise
there there are no noise pations uh but if we uh cons in in
pations uh but if we uh cons in in
pations uh but if we uh cons in in additional cons in addition consider uh
additional cons in addition consider uh
additional cons in addition consider uh the noise pations like some weather uh
the noise pations like some weather uh
the noise pations like some weather uh pations and some extreme extreme ex
pations and some extreme extreme ex
pations and some extreme extreme ex internal viabilities in this case even
internal viabilities in this case even
internal viabilities in this case even though the um uh the forcing shift rate
though the um uh the forcing shift rate
though the um uh the forcing shift rate is not so fast but uh the r de tipping
is not so fast but uh the r de tipping
is not so fast but uh the r de tipping can still happens this is what I mean
can still happens this is what I mean
can still happens this is what I mean the um conert of the noise and the
the um conert of the noise and the
the um conert of the noise and the rating D
rating D
rating D tipping this could be very realistic for
tipping this could be very realistic for
tipping this could be very realistic for um more closer to uh this could be
um more closer to uh this could be
um more closer to uh this could be closer to our realistic cases like the
closer to our realistic cases like the
closer to our realistic cases like the vegetation and also climate
vegetation and also climate
vegetation and also climate systems and uh here we just take s uh uh
systems and uh here we just take s uh uh
systems and uh here we just take s uh uh we just take three
we just take three
we just take three uh typical uh dynamical system the sad
uh typical uh dynamical system the sad
uh typical uh dynamical system the sad no and bolting and the compos boom uh
no and bolting and the compos boom uh
no and bolting and the compos boom uh system and perform perform the fast
system and perform perform the fast
system and perform perform the fast forcing uh external forcing on on this
forcing uh external forcing on on this
forcing uh external forcing on on this three systems in uh spef uh respectively
three systems in uh spef uh respectively
three systems in uh spef uh respectively and also perform at uh also add some
and also perform at uh also add some
and also perform at uh also add some noise probations to them and uh then we
noise probations to them and uh then we
noise probations to them and uh then we can get different simulations even for
can get different simulations even for
can get different simulations even for example if we uh first have a look at
example if we uh first have a look at
example if we uh first have a look at the S node system simulations we always
the S node system simulations we always
the S node system simulations we always keep the same forcing the same forcing
keep the same forcing the same forcing
keep the same forcing the same forcing profile and also the same forcing rate
profile and also the same forcing rate
profile and also the same forcing rate forcing uh same uh forcing TR trajectory
forcing uh same uh forcing TR trajectory
forcing uh same uh forcing TR trajectory time series um in this case we if we
time series um in this case we if we
time series um in this case we if we have different white noise perturbations
have different white noise perturbations
have different white noise perturbations on this system and the trajectory of the
on this system and the trajectory of the
on this system and the trajectory of the system time series will be very
system time series will be very
system time series will be very different
different
different and here we notice some of them were
and here we notice some of them were
and here we notice some of them were were TP but some of them were not and we
were TP but some of them were not and we
were TP but some of them were not and we also make a statistic for the Tipping uh
also make a statistic for the Tipping uh
also make a statistic for the Tipping uh tip time like the when this red lines
tip time like the when this red lines
tip time like the when this red lines when this red time Ser
when this red time Ser
when this red time Ser largely significant diverge from the uh
largely significant diverge from the uh
largely significant diverge from the uh nontipping
nontipping
nontipping category uh envelope then we can Define
category uh envelope then we can Define
category uh envelope then we can Define The Tipping time and you you'll see here
The Tipping time and you you'll see here
The Tipping time and you you'll see here the tiing time is uh the uncertainty of
the tiing time is uh the uncertainty of
the tiing time is uh the uncertainty of the Tipping time is really large we
the Tipping time is really large we
the Tipping time is really large we cannot we cannot really estimate uh
cannot we cannot really estimate uh
cannot we cannot really estimate uh estimates The Tipping time simply from
estimates The Tipping time simply from
estimates The Tipping time simply from the equations even though even though
the equations even though even though
the equations even though even though now we already know the uh physical
now we already know the uh physical
now we already know the uh physical equation mathematical uh equation of
equation mathematical uh equation of
equation mathematical uh equation of this set no system but we can steer with
this set no system but we can steer with
this set no system but we can steer with steer cannot really estimate the uh
steer cannot really estimate the uh
steer cannot really estimate the uh forcing the thresholds of the tpping r
forcing the thresholds of the tpping r
forcing the thresholds of the tpping r IND tpping and this uh this could this
IND tpping and this uh this could this
IND tpping and this uh this could this would be similar uh for both for the
would be similar uh for both for the
would be similar uh for both for the other two systems and of course this is
other two systems and of course this is
other two systems and of course this is similar uh similar phenomenon uh are
similar uh similar phenomenon uh are
similar uh similar phenomenon uh are already observed by Johannes and uh
already observed by Johannes and uh
already observed by Johannes and uh Peter in the simul amok
simulations so uh here we we just want
simulations so uh here we we just want
simulations so uh here we we just want to make some pred prediction for for
to make some pred prediction for for
to make some pred prediction for for this tipping points we just know I
this tipping points we just know I
this tipping points we just know I mentioned three main methods uh the
mentioned three main methods uh the
mentioned three main methods uh the first methods it's yeah it does not work
first methods it's yeah it does not work
first methods it's yeah it does not work here the F found are um threshold values
here the F found are um threshold values
here the F found are um threshold values but it does not work here and uh then uh
but it does not work here and uh then uh
but it does not work here and uh then uh we take the uh critical slowing down
we take the uh critical slowing down
we take the uh critical slowing down indicators the auto
indicators the auto
indicators the auto cation and and also the um variance of
cation and and also the um variance of
cation and and also the um variance of the time serus here firstly if we uh uh
the time serus here firstly if we uh uh
the time serus here firstly if we uh uh if we uh classify the uh simply Cate
if we uh classify the uh simply Cate
if we uh classify the uh simply Cate make a category two categories for the
make a category two categories for the
make a category two categories for the ring scenario and noning scenario for
ring scenario and noning scenario for
ring scenario and noning scenario for our emble simulations time series or
our emble simulations time series or
our emble simulations time series or samples then uh from the time serious
samples then uh from the time serious
samples then uh from the time serious trajectory before the ring happens we
trajectory before the ring happens we
trajectory before the ring happens we cannot uh we cannot find find out the uh
cannot uh we cannot find find out the uh
cannot uh we cannot find find out the uh distinct the tping time Ser in visual
distinct the tping time Ser in visual
distinct the tping time Ser in visual observation and then if we calculate the
observation and then if we calculate the
observation and then if we calculate the uh alation and the variance of the two
uh alation and the variance of the two
uh alation and the variance of the two categories then we will find there there
categories then we will find there there
categories then we will find there there is no diff almost no vual difference for
is no diff almost no vual difference for
is no diff almost no vual difference for the r tping and the nontipping
the r tping and the nontipping
the r tping and the nontipping categories uh but of course uh for the
categories uh but of course uh for the
categories uh but of course uh for the uh deping methods uh we can distinguish
uh deping methods uh we can distinguish
uh deping methods uh we can distinguish them very well but I will introduce uh
them very well but I will introduce uh
them very well but I will introduce uh the Third indicator
the Third indicator
the Third indicator later and yeah but still we we still
later and yeah but still we we still
later and yeah but still we we still believe maybe we we can find some find
believe maybe we we can find some find
believe maybe we we can find some find out some uh uh find out some evidence of
out some uh uh find out some evidence of
out some uh uh find out some evidence of the potential PR predictability of the
the potential PR predictability of the
the potential PR predictability of the ring from the raw data that means and
ring from the raw data that means and
ring from the raw data that means and here we have the time serious or we
here we have the time serious or we
here we have the time serious or we already have the time series of tpping
already have the time series of tpping
already have the time series of tpping and non- tipping categories then uh
and non- tipping categories then uh
and non- tipping categories then uh focus on the specific uh leading time uh
focus on the specific uh leading time uh
focus on the specific uh leading time uh before the Tipping happens for example
before the Tipping happens for example
before the Tipping happens for example here uh if if we take the 100 time steps
here uh if if we take the 100 time steps
here uh if if we take the 100 time steps before the Tipping and make a make a
before the Tipping and make a make a
before the Tipping and make a make a slides for the uh for the the two
slides for the uh for the the two
slides for the uh for the the two envelopes then we can get a uh PDF and a
envelopes then we can get a uh PDF and a
envelopes then we can get a uh PDF and a probability distribution of these two
probability distribution of these two
probability distribution of these two categories and we can visually we can
categories and we can visually we can
categories and we can visually we can really observe some difference between
really observe some difference between
really observe some difference between these two uh categories but uh uh but in
these two uh categories but uh uh but in
these two uh categories but uh uh but in um practice if we really want to make
um practice if we really want to make
um practice if we really want to make prediction based on this uh based on the
prediction based on this uh based on the
prediction based on this uh based on the diff visual difference from the uh from
diff visual difference from the uh from
diff visual difference from the uh from the probability distribution then it
the probability distribution then it
the probability distribution then it will uh will make you uh disappointed
will uh will make you uh disappointed
will uh will make you uh disappointed because but but I will introduce the
because but but I will introduce the
because but but I will introduce the prediction results based on this uh
prediction results based on this uh
prediction results based on this uh probability distribution later but yeah
probability distribution later but yeah
probability distribution later but yeah I can I can tell here it does not work
I can I can tell here it does not work
I can I can tell here it does not work well and also we we observe some we also
well and also we we observe some we also
well and also we we observe some we also have a look at the uh distribution
have a look at the uh distribution
have a look at the uh distribution difference uh on the based on the autoc
difference uh on the based on the autoc
difference uh on the based on the autoc coration indicator and also the variance
coration indicator and also the variance
coration indicator and also the variance uh indicator and indeed we can observe
uh indicator and indeed we can observe
uh indicator and indeed we can observe some difference in the uh
some difference in the uh
some difference in the uh uh in in the PDF but uh uh but still it
uh in in the PDF but uh uh but still it
uh in in the PDF but uh uh but still it cannot really help us to make prediction
cannot really help us to make prediction
cannot really help us to make prediction for a uh individual tra trajectories uh
for a uh individual tra trajectories uh
for a uh individual tra trajectories uh that means basically we can if we
that means basically we can if we
that means basically we can if we observe the and we already know the um
observe the and we already know the um
observe the and we already know the um envelopes of the Tipping and nontipping
envelopes of the Tipping and nontipping
envelopes of the Tipping and nontipping trajectory but um hopefully for
trajectory but um hopefully for
trajectory but um hopefully for individual trajectory time series we we
individual trajectory time series we we
individual trajectory time series we we can also make the prediction but
can also make the prediction but
can also make the prediction but uh yeah this is what we want uh just um
uh yeah this is what we want uh just um
uh yeah this is what we want uh just um get some experience from the emble
get some experience from the emble
get some experience from the emble simulations and then make the prediction
simulations and then make the prediction
simulations and then make the prediction for a specific TR uh time series uh this
for a specific TR uh time series uh this
for a specific TR uh time series uh this is also usually what we did for the uh
is also usually what we did for the uh
is also usually what we did for the uh climate in the climate uh
climate in the climate uh
climate in the climate uh studies uh so here uh let's uh let's
studies uh so here uh let's uh let's
studies uh so here uh let's uh let's move to to the uh the third indicator
move to to the uh the third indicator
move to to the uh the third indicator the third prediction method uh the deing
the third prediction method uh the deing
the third prediction method uh the deing method here uh we we just take the time
method here uh we we just take the time
method here uh we we just take the time serious from The Tipping and nontipping
serious from The Tipping and nontipping
serious from The Tipping and nontipping categories and then for example here we
categories and then for example here we
categories and then for example here we uh just pick uh pick up a time serious
uh just pick uh pick up a time serious
uh just pick uh pick up a time serious segment uh just with a with a certain
segment uh just with a with a certain
segment uh just with a with a certain Focus lead time to the tiing points then
Focus lead time to the tiing points then
Focus lead time to the tiing points then fed it fed them into the Neal networks
fed it fed them into the Neal networks
fed it fed them into the Neal networks here uh and then hopefully we uh we hope
here uh and then hopefully we uh we hope
here uh and then hopefully we uh we hope the expect the uh neon networks can give
the expect the uh neon networks can give
the expect the uh neon networks can give us uh can give us the a classification
us uh can give us the a classification
us uh can give us the a classification results it's basically the ring
results it's basically the ring
results it's basically the ring probability and the nontipping
probability and the nontipping
probability and the nontipping probability of this time series uh here
probability of this time series uh here
probability of this time series uh here are the Naro Network for the Naro
are the Naro Network for the Naro
are the Naro Network for the Naro Network architecture we take we took a
Network architecture we take we took a
Network architecture we take we took a very simple architecture based on the
very simple architecture based on the
very simple architecture based on the well known uh s architecture uh Neal
well known uh s architecture uh Neal
well known uh s architecture uh Neal Network here we take two S Neal Network
Network here we take two S Neal Network
Network here we take two S Neal Network and uh then uh and and then uh combine
and uh then uh and and then uh combine
and uh then uh and and then uh combine them with a fully connected uh Naro
them with a fully connected uh Naro
them with a fully connected uh Naro Network this is very simple because uh
Network this is very simple because uh
Network this is very simple because uh we uh we want to yeah we expect if the
we uh we want to yeah we expect if the
we uh we want to yeah we expect if the very s very simple Naro Network
very s very simple Naro Network
very s very simple Naro Network architectures can work for the T WR
architectures can work for the T WR
architectures can work for the T WR those tipping prediction then hopefully
those tipping prediction then hopefully
those tipping prediction then hopefully the more much more complicated
the more much more complicated
the more much more complicated architecture will also works um so here
architecture will also works um so here
architecture will also works um so here let's say the uh prediction uh results
let's say the uh prediction uh results
let's say the uh prediction uh results here uh firstly let's have a look at the
here uh firstly let's have a look at the
here uh firstly let's have a look at the S node system time Ser here uh I only
S node system time Ser here uh I only
S node system time Ser here uh I only present two time Ser of this from this
present two time Ser of this from this
present two time Ser of this from this system and here the the red one one is
system and here the the red one one is
system and here the the red one one is for the Tipping time serus blue one is
for the Tipping time serus blue one is
for the Tipping time serus blue one is the nontipping case and here if um by
the nontipping case and here if um by
the nontipping case and here if um by fit them into the trended uh Naro
fit them into the trended uh Naro
fit them into the trended uh Naro networks then the Naro Network could
networks then the Naro Network could
networks then the Naro Network could give uh provide as uh the indicator ex
give uh provide as uh the indicator ex
give uh provide as uh the indicator ex exactly the um RTP probability of this
exactly the um RTP probability of this
exactly the um RTP probability of this time seral at different forecast lead
time seral at different forecast lead
time seral at different forecast lead time and when this when the time is
time and when this when the time is
time and when this when the time is approach to the uh tipping uh tipping
approach to the uh tipping uh tipping
approach to the uh tipping uh tipping points then for the non- Tipping case
points then for the non- Tipping case
points then for the non- Tipping case it's tipping probability where
it's tipping probability where
it's tipping probability where continuous increase until 100% but for
continuous increase until 100% but for
continuous increase until 100% but for the non- Tipping case uh its predict
the non- Tipping case uh its predict
the non- Tipping case uh its predict probability tipping probability will
probability tipping probability will
probability tipping probability will decree continuously uh decrease and they
decree continuously uh decrease and they
decree continuously uh decrease and they just this two this indicators for this
just this two this indicators for this
just this two this indicators for this two times Ser just develop alone uh two
two times Ser just develop alone uh two
two times Ser just develop alone uh two opposite directions but for the uh autoc
opposite directions but for the uh autoc
opposite directions but for the uh autoc colation and the uh Rance
colation and the uh Rance
colation and the uh Rance indicators yeah they both give the
indicators yeah they both give the
indicators yeah they both give the increase uh increase signal but we
increase uh increase signal but we
increase uh increase signal but we cannot really yeah infer if there will
cannot really yeah infer if there will
cannot really yeah infer if there will be a tipping points for uh for the blue
be a tipping points for uh for the blue
be a tipping points for uh for the blue or red uh time series before the Tipping
or red uh time series before the Tipping
or red uh time series before the Tipping happens uh this similar thing also
happens uh this similar thing also
happens uh this similar thing also similar phenomena conclusion can be also
similar phenomena conclusion can be also
similar phenomena conclusion can be also find from the uh boting system and
find from the uh boting system and
find from the uh boting system and compost boom uh
compost boom uh
compost boom uh system yeah here talk a bit about the
system yeah here talk a bit about the
system yeah here talk a bit about the training data that's used for the model
training data that's used for the model
training data that's used for the model what kind of training data was used yeah
what kind of training data was used yeah
what kind of training data was used yeah and the training data basically it's uh
and the training data basically it's uh
and the training data basically it's uh it's here uh if we have a look at time
it's here uh if we have a look at time
it's here uh if we have a look at time series in this slide uh the first panel
series in this slide uh the first panel
series in this slide uh the first panel here uh the uh wait yeah
here uh the uh wait yeah
here uh the uh wait yeah [Music]
[Music]
[Music] and this envelope is is made from the a
and this envelope is is made from the a
and this envelope is is made from the a lot of time serious is emble simulations
lot of time serious is emble simulations
lot of time serious is emble simulations this means uh here we perform the same
this means uh here we perform the same
this means uh here we perform the same external forcing on the same system but
external forcing on the same system but
external forcing on the same system but using different White Noise
using different White Noise
using different White Noise realizations then we can get
realizations then we can get
realizations then we can get different uh we can that get different
different uh we can that get different
different uh we can that get different simulations this is just like the uh
simulations this is just like the uh
simulations this is just like the uh climate also like the climate
climate also like the climate
climate also like the climate simulations we we give the different uh
simulations we we give the different uh
simulations we we give the different uh pations then we can get different uh
pations then we can get different uh
pations then we can get different uh trajectories this is then yeah our data
trajectories this is then yeah our data
trajectories this is then yeah our data sets uh source and the the the training
sets uh source and the the the training
sets uh source and the the the training training set is uh basically we only
training set is uh basically we only
training set is uh basically we only have two this is a uh classifier Naro
have two this is a uh classifier Naro
have two this is a uh classifier Naro network is a classifier uh we only fed
network is a classifier uh we only fed
network is a classifier uh we only fed two kinds of data one is tipping data n
two kinds of data one is tipping data n
two kinds of data one is tipping data n uh second is nontipping uh nontipping
uh second is nontipping uh nontipping
uh second is nontipping uh nontipping data and uh and we also labor labor them
data and uh and we also labor labor them
data and uh and we also labor labor them as zero and one respectively then we can
as zero and one respectively then we can
as zero and one respectively then we can get the uh tipping probab the uh
get the uh tipping probab the uh
get the uh tipping probab the uh classifier probability for these two
classifier probability for these two
classifier probability for these two category two
category two
category two classifications yeah I hope this is
classifications yeah I hope this is
classifications yeah I hope this is clear to you yeah
clear to you yeah
clear to you yeah so it's climate simulation data that's
so it's climate simulation data that's
so it's climate simulation data that's used for for training um ENT so
used for for training um ENT so
used for for training um ENT so essentially this model is capturing the
essentially this model is capturing the
essentially this model is capturing the the way in which those climate models uh
the way in which those climate models uh
the way in which those climate models uh identify the r typing
identify the r typing
identify the r typing yes okay yeah thanks okay yeah you're
yes okay yeah thanks okay yeah you're
yes okay yeah thanks okay yeah you're welome and but yeah of course here in
welome and but yeah of course here in
welome and but yeah of course here in current stage uh the model which
current stage uh the model which
current stage uh the model which actually train three kinds of model um
actually train three kinds of model um
actually train three kinds of model um each kind of model specifically for a
each kind of model specifically for a
each kind of model specifically for a specific a certain system for example
specific a certain system for example
specific a certain system for example for S no system which and the Neo
for S no system which and the Neo
for S no system which and the Neo networks but for the boting system we
networks but for the boting system we
networks but for the boting system we will retrain a we will train a new uh
will retrain a we will train a new uh
will retrain a we will train a new uh deing model for its prediction and uh
deing model for its prediction and uh
deing model for its prediction and uh but but in in the uh last slides uh in
but but in in the uh last slides uh in
but but in in the uh last slides uh in the later uh following slides I will
the later uh following slides I will
the later uh following slides I will also present uh the uh inter model
also present uh the uh inter model
also present uh the uh inter model interstem
interstem
interstem prediction uh but here it's we can OB
prediction uh but here it's we can OB
prediction uh but here it's we can OB everything is clear and we can
everything is clear and we can
everything is clear and we can understand the understand the system and
understand the understand the system and
understand the understand the system and de planing very Behavior Uh better in
de planing very Behavior Uh better in
de planing very Behavior Uh better in this way yeah uh here I yeah I would
this way yeah uh here I yeah I would
this way yeah uh here I yeah I would yeah make a simple make some simple U
yeah make a simple make some simple U
yeah make a simple make some simple U comparation between the S tipping uh the
comparation between the S tipping uh the
comparation between the S tipping uh the between the say critical slowing down
between the say critical slowing down
between the say critical slowing down indicator and say uh de planning
indicator and say uh de planning
indicator and say uh de planning indicator comparation here we we notice
indicator comparation here we we notice
indicator comparation here we we notice for the auto coralation when we perform
for the auto coralation when we perform
for the auto coralation when we perform the uh prediction on all the uh time
the uh prediction on all the uh time
the uh prediction on all the uh time serial samples then we can observe the
serial samples then we can observe the
serial samples then we can observe the uh for the uh for the uh deing indicator
uh for the uh for the uh deing indicator
uh for the uh for the uh deing indicator uh they can it can always help uh it can
uh they can it can always help uh it can
uh they can it can always help uh it can always help us distinguish The Tipping
always help us distinguish The Tipping
always help us distinguish The Tipping and tipping and nontipping
and tipping and nontipping
and tipping and nontipping uh
uh
uh category with a considerable uh forecast
category with a considerable uh forecast
category with a considerable uh forecast lead
lead
lead time of course here uh we we also make a
time of course here uh we we also make a
time of course here uh we we also make a estimate uh evaluation for the um uh for
estimate uh evaluation for the um uh for
estimate uh evaluation for the um uh for the de planning prediction actually it's
the de planning prediction actually it's
the de planning prediction actually it's it's a classif it's just a classifier uh
it's a classif it's just a classifier uh
it's a classif it's just a classifier uh for the Tipping and nontipping uh but
for the Tipping and nontipping uh but
for the Tipping and nontipping uh but let's firstly have a look at the bottom
let's firstly have a look at the bottom
let's firstly have a look at the bottom panels bottom panels means the a
panels bottom panels means the a
panels bottom panels means the a different col the blue and red colors
different col the blue and red colors
different col the blue and red colors represents the uh testing data set and
represents the uh testing data set and
represents the uh testing data set and the training data set uh
the training data set uh
the training data set uh uh overall uh the the training and the
uh overall uh the the training and the
uh overall uh the the training and the testing
testing
testing accuracy uh curve are over almost
accuracy uh curve are over almost
accuracy uh curve are over almost overlapped this means our model is
overlapped this means our model is
overlapped this means our model is robust and the and and and
robust and the and and and
robust and the and and and we don't need to worry about the
we don't need to worry about the
we don't need to worry about the overfitting problem but of course we
overfitting problem but of course we
overfitting problem but of course we also perform the cross validation uh a
also perform the cross validation uh a
also perform the cross validation uh a lot of a com
lot of a com
lot of a com extensive uh evaluation on to check if
extensive uh evaluation on to check if
extensive uh evaluation on to check if there is any
there is any
there is any um overfitting uh issue for the model
um overfitting uh issue for the model
um overfitting uh issue for the model but uh the results also show the
but uh the results also show the
but uh the results also show the overfitting almost very weak here and
overfitting almost very weak here and
overfitting almost very weak here and and also where uh in the very beginning
and also where uh in the very beginning
and also where uh in the very beginning of our when we have a look at the time
of our when we have a look at the time
of our when we have a look at the time serious T tipping and nontipping time
serious T tipping and nontipping time
serious T tipping and nontipping time serious I mentioned we can uh actually
serious I mentioned we can uh actually
serious I mentioned we can uh actually make a very simple prediction based on
make a very simple prediction based on
make a very simple prediction based on the uh probability distribution of the
the uh probability distribution of the
the uh probability distribution of the tpping and noning and uh yeah we write
tpping and noning and uh yeah we write
tpping and noning and uh yeah we write uh we uh create some very simple
uh we uh create some very simple
uh we uh create some very simple algorithm to realize the probability uh
algorithm to realize the probability uh
algorithm to realize the probability uh distribution based prediction
distribution based prediction
distribution based prediction uh method and uh and we compare it with
uh method and uh and we compare it with
uh method and uh and we compare it with our deing indicator here and same uh
our deing indicator here and same uh
our deing indicator here and same uh deing indicator uh have better um
deing indicator uh have better um
deing indicator uh have better um accuracy uh than the uh than the visual
accuracy uh than the uh than the visual
accuracy uh than the uh than the visual visual observation on the um probability
visual observation on the um probability
visual observation on the um probability distribution yeah and uh in the
distribution yeah and uh in the
distribution yeah and uh in the following I would like to uh pres uh
following I would like to uh pres uh
following I would like to uh pres uh introduce some very interesting
introduce some very interesting
introduce some very interesting phenomena based on our analyze on our uh
phenomena based on our analyze on our uh
phenomena based on our analyze on our uh trended def uh deing model uh firstly we
trended def uh deing model uh firstly we
trended def uh deing model uh firstly we uh just uh we want to discuss about out
uh just uh we want to discuss about out
uh just uh we want to discuss about out out of sample prediction of dep planning
out of sample prediction of dep planning
out of sample prediction of dep planning uh because this is this is very
uh because this is this is very
uh because this is this is very important um for uh for the data science
important um for uh for the data science
important um for uh for the data science uh usually we train our model on a
uh usually we train our model on a
uh usually we train our model on a specific uh some specific data sets but
specific uh some specific data sets but
specific uh some specific data sets but but
but
but afterwards uh when we met some new data
afterwards uh when we met some new data
afterwards uh when we met some new data set data resource uh our
set data resource uh our
set data resource uh our model uh it it would be some uh it would
model uh it it would be some uh it would
model uh it it would be some uh it would be dangerous for a model to make some
be dangerous for a model to make some
be dangerous for a model to make some some U make prediction on some unseen a
some U make prediction on some unseen a
some U make prediction on some unseen a previously unseen data resource uh so
previously unseen data resource uh so
previously unseen data resource uh so here we want to discuss about this point
here we want to discuss about this point
here we want to discuss about this point uh firstly we test our model for the out
uh firstly we test our model for the out
uh firstly we test our model for the out of sample foring rates prediction uh uh
of sample foring rates prediction uh uh
of sample foring rates prediction uh uh prediction uh that is just now I
prediction uh that is just now I
prediction uh that is just now I mentioned uh the model the model I me
mentioned uh the model the model I me
mentioned uh the model the model I me trended model I mentioned just now is
trended model I mentioned just now is
trended model I mentioned just now is all uh is trended based on a fixed
all uh is trended based on a fixed
all uh is trended based on a fixed forcing rate for example here uh
forcing rate for example here uh
forcing rate for example here uh actually for the set node system
actually for the set node system
actually for the set node system basically we can design different
basically we can design different
basically we can design different forcing forcing rates for it but when we
forcing forcing rates for it but when we
forcing forcing rates for it but when we train our deing model uh we always use a
train our deing model uh we always use a
train our deing model uh we always use a fixed forcing rates
fixed forcing rates
fixed forcing rates the foring rate is just the diff uh the
the foring rate is just the diff uh the
the foring rate is just the diff uh the um it can be controlled uh controlled by
um it can be controlled uh controlled by
um it can be controlled uh controlled by the parameter Lambda in the S node
the parameter Lambda in the S node
the parameter Lambda in the S node system and uh yeah here we take the only
system and uh yeah here we take the only
system and uh yeah here we take the only take a fixed forcing rate it's
take a fixed forcing rate it's
take a fixed forcing rate it's 1.25 like like this dash line
1.25 like like this dash line
1.25 like like this dash line sh um uh when we TR trained on after we
sh um uh when we TR trained on after we
sh um uh when we TR trained on after we trained our model we have fixed
trained our model we have fixed
trained our model we have fixed everything uh unchanged and just let
everything uh unchanged and just let
everything uh unchanged and just let uh uh just let's
uh uh just let's
uh uh just let's uh just use the trended deing model to
uh just use the trended deing model to
uh just use the trended deing model to make prediction for other uh time serus
make prediction for other uh time serus
make prediction for other uh time serus uh the other time serus are generated
uh the other time serus are generated
uh the other time serus are generated from the different uh foring rates and
from the different uh foring rates and
from the different uh foring rates and these forcing rates are unseen during
these forcing rates are unseen during
these forcing rates are unseen during the training Pro uh training phase and
the training Pro uh training phase and
the training Pro uh training phase and here it seems uh it's still work for
here it seems uh it's still work for
here it seems uh it's still work for other foring rate uh but and uh it seems
other foring rate uh but and uh it seems
other foring rate uh but and uh it seems it could work for the out of sample
it could work for the out of sample
it could work for the out of sample forcing rate so in the following way we
forcing rate so in the following way we
forcing rate so in the following way we further uh train a new uh deing model
further uh train a new uh deing model
further uh train a new uh deing model for the set node system prediction here
for the set node system prediction here
for the set node system prediction here we just take another fixed forcing rate
we just take another fixed forcing rate
we just take another fixed forcing rate like the uh panel d uh in the panel D we
like the uh panel d uh in the panel D we
like the uh panel d uh in the panel D we use another forcing rate and generates
use another forcing rate and generates
use another forcing rate and generates the data and the Train the deping model
the data and the Train the deping model
the data and the Train the deping model then um your the trended model predicts
then um your the trended model predicts
then um your the trended model predicts the unse forcing rates on for other time
the unse forcing rates on for other time
the unse forcing rates on for other time series like here panel D and here if we
series like here panel D and here if we
series like here panel D and here if we compare panel a and panel d uh the deing
compare panel a and panel d uh the deing
compare panel a and panel d uh the deing model are trended on different forcing
model are trended on different forcing
model are trended on different forcing rate but uh the accuracy pattern uh are
rate but uh the accuracy pattern uh are
rate but uh the accuracy pattern uh are very similar for this uh the two uh uh
very similar for this uh the two uh uh
very similar for this uh the two uh uh trended trended models and uh
trended trended models and uh
trended trended models and uh uh for the pan in the panel G uh basic
uh for the pan in the panel G uh basic
uh for the pan in the panel G uh basic actually this is we did a additional
actually this is we did a additional
actually this is we did a additional experiment where just fed uh fed uh fed
experiment where just fed uh fed uh fed
experiment where just fed uh fed uh fed all the forcing rate time seral into the
all the forcing rate time seral into the
all the forcing rate time seral into the uh deing model then in this case uh
uh deing model then in this case uh
uh deing model then in this case uh there will be no unseen previous
there will be no unseen previous
there will be no unseen previous previously unseen foring rates for the
previously unseen foring rates for the
previously unseen foring rates for the deing model then we compare the
deing model then we compare the
deing model then we compare the prediction accurate pattern uh from
prediction accurate pattern uh from
prediction accurate pattern uh from panel a d g and they looks uh very
panel a d g and they looks uh very
panel a d g and they looks uh very consistent and this is for the set node
consistent and this is for the set node
consistent and this is for the set node system and here also did the same thing
system and here also did the same thing
system and here also did the same thing for other uh the other two typical uh uh
for other uh the other two typical uh uh
for other uh the other two typical uh uh systems uh for the compost boom system
systems uh for the compost boom system
systems uh for the compost boom system uh we use different forcing rates and uh
uh we use different forcing rates and uh
uh we use different forcing rates and uh we can get the same the same conclusion
we can get the same the same conclusion
we can get the same the same conclusion as the S node system but for the boting
as the S node system but for the boting
as the S node system but for the boting system
system
system yeah honestly here um it's it's it's
yeah honestly here um it's it's it's
yeah honestly here um it's it's it's really different for panel B and panel E
really different for panel B and panel E
really different for panel B and panel E when we use different uh forcing rate to
when we use different uh forcing rate to
when we use different uh forcing rate to train the model and uh yeah for this uh
train the model and uh yeah for this uh
train the model and uh yeah for this uh point for this phenomena uh we we
point for this phenomena uh we we
point for this phenomena uh we we thought is because the boting system
thought is because the boting system
thought is because the boting system basically it's very it's uh a lot
basically it's very it's uh a lot
basically it's very it's uh a lot of it's uh it's it has has a lot of uh
of it's uh it's it has has a lot of uh
of it's uh it's it has has a lot of uh complicated uh chaotic ulations in in
complicated uh chaotic ulations in in
complicated uh chaotic ulations in in itself so in this case um like in panel
itself so in this case um like in panel
itself so in this case um like in panel B when we train the train train the
B when we train the train train the
B when we train the train train the depending model on fix forcing rate for
depending model on fix forcing rate for
depending model on fix forcing rate for other forcing rat the chaotic
other forcing rat the chaotic
other forcing rat the chaotic occilations behavior is really different
occilations behavior is really different
occilations behavior is really different we thought this is uh this uh chaotic
we thought this is uh this uh chaotic
we thought this is uh this uh chaotic occilations will disturb the uh uh will
occilations will disturb the uh uh will
occilations will disturb the uh uh will disturb the prediction of the deing
disturb the prediction of the deing
disturb the prediction of the deing model and but in if we fit all the time
model and but in if we fit all the time
model and but in if we fit all the time Ser forcing race time series into the
Ser forcing race time series into the
Ser forcing race time series into the mod train the model and make prediction
mod train the model and make prediction
mod train the model and make prediction on the testing data set then uh it can
on the testing data set then uh it can
on the testing data set then uh it can give us a a looks uh much um much more
give us a a looks uh much um much more
give us a a looks uh much um much more comfortable uh pattern uh in its
comfortable uh pattern uh in its
comfortable uh pattern uh in its accuracy
accuracy
accuracy and yeah here is this means uh when we
and yeah here is this means uh when we
and yeah here is this means uh when we Trend this is also important for uh for
Trend this is also important for uh for
Trend this is also important for uh for us uh in in the future if we want to
us uh in in the future if we want to
us uh in in the future if we want to train a deing model we should consider
train a deing model we should consider
train a deing model we should consider uh for a specific uh dynamical system
uh for a specific uh dynamical system
uh for a specific uh dynamical system its internal Dynamics could Al could
its internal Dynamics could Al could
its internal Dynamics could Al could potentially disturb the deing uh
potentially disturb the deing uh
potentially disturb the deing uh indicator this need to be uh careful
indicator this need to be uh careful
indicator this need to be uh careful yeah uh and then second uh second
yeah uh and then second uh second
yeah uh and then second uh second analyze is for the uh prediction of the
analyze is for the uh prediction of the
analyze is for the uh prediction of the out of sample dynamical systems uh that
out of sample dynamical systems uh that
out of sample dynamical systems uh that means um for the three dynamical system
means um for the three dynamical system
means um for the three dynamical system set node Bing and composed boom uh
set node Bing and composed boom uh
set node Bing and composed boom uh the here we plot the autocorrelation
the here we plot the autocorrelation
the here we plot the autocorrelation function of the time serious uh
function of the time serious uh
function of the time serious uh respectively from this this three system
respectively from this this three system
respectively from this this three system systems and we can observe of the um
systems and we can observe of the um
systems and we can observe of the um the uh there is some uh time scale
the uh there is some uh time scale
the uh there is some uh time scale separation between this uh s uh systems
separation between this uh s uh systems
separation between this uh s uh systems this means the dynamical dynamical
this means the dynamical dynamical
this means the dynamical dynamical Dynamics of the three systems are
Dynamics of the three systems are
Dynamics of the three systems are largely different from one another uh
largely different from one another uh
largely different from one another uh but here we still perform a uh
but here we still perform a uh
but here we still perform a uh interesting experiment for example for
interesting experiment for example for
interesting experiment for example for the sod system we train the S train the
the sod system we train the S train the
the sod system we train the S train the depending model on I fixed uh foring
depending model on I fixed uh foring
depending model on I fixed uh foring rate uh a a certain foring rate uh from
rate uh a a certain foring rate uh from
rate uh a a certain foring rate uh from the set not system then make prediction
the set not system then make prediction
the set not system then make prediction for other foring rates and also for the
for other foring rates and also for the
for other foring rates and also for the other foring rates in the uh uh in the
other foring rates in the uh uh in the
other foring rates in the uh uh in the Bing system and the composed system
Bing system and the composed system
Bing system and the composed system Let's uh let's look at a panel a b c uh
Let's uh let's look at a panel a b c uh
Let's uh let's look at a panel a b c uh this means
this means
this means uh and and we can still observe some uh
uh and and we can still observe some uh
uh and and we can still observe some uh some uh some skill for prediction but
some uh some skill for prediction but
some uh some skill for prediction but also it's very close to the Tipping
also it's very close to the Tipping
also it's very close to the Tipping happen uh uh tipping times but but still
happen uh uh tipping times but but still
happen uh uh tipping times but but still this means uh hopefully this uh this uh
this means uh hopefully this uh this uh
this means uh hopefully this uh this uh deing could make some out of sample
deing could make some out of sample
deing could make some out of sample prediction on make some prediction scale
prediction on make some prediction scale
prediction on make some prediction scale for prediction under out of sample
for prediction under out of sample
for prediction under out of sample dynamical systems and uh here we we also
dynamical systems and uh here we we also
dynamical systems and uh here we we also did the same thing for the for the other
did the same thing for the for the other
did the same thing for the for the other uh two dynamical systems and it seems
uh two dynamical systems and it seems
uh two dynamical systems and it seems it's very interesting like the uh B for
it's very interesting like the uh B for
it's very interesting like the uh B for the boting system the depending model
the boting system the depending model
the boting system the depending model trended based on boting system it can
trended based on boting system it can
trended based on boting system it can work very well for the composed boom
work very well for the composed boom
work very well for the composed boom system and uh this pattern if we have
system and uh this pattern if we have
system and uh this pattern if we have looked at the pattern F and I the Bing
looked at the pattern F and I the Bing
looked at the pattern F and I the Bing the boting systems de planing model can
the boting systems de planing model can
the boting systems de planing model can make very well prediction for the
make very well prediction for the
make very well prediction for the compost boom system and the pattern are
compost boom system and the pattern are
compost boom system and the pattern are very similar to each other uh this means
very similar to each other uh this means
very similar to each other uh this means although the dynamic Dynamics are
although the dynamic Dynamics are
although the dynamic Dynamics are different but it can still make skill
different but it can still make skill
different but it can still make skill for prediction sometimes and uh uh yeah
for prediction sometimes and uh uh yeah
for prediction sometimes and uh uh yeah uh finally we also uh did a uh uh we
uh finally we also uh did a uh uh we
uh finally we also uh did a uh uh we also trans the TR our deping model trans
also trans the TR our deping model trans
also trans the TR our deping model trans deping model in include all the forcing
deping model in include all the forcing
deping model in include all the forcing rates and all the uh time the time Ser
rates and all the uh time the time Ser
rates and all the uh time the time Ser from all the system all the three
from all the system all the three
from all the system all the three systems and then the this is we call it
systems and then the this is we call it
systems and then the this is we call it a comprehensive deing model then
a comprehensive deing model then
a comprehensive deing model then eventually this comprehensive deing
eventually this comprehensive deing
eventually this comprehensive deing model can predict the uh yeah can
model can predict the uh yeah can
model can predict the uh yeah can predict the uh uh wrting those tipping
predict the uh uh wrting those tipping
predict the uh uh wrting those tipping very very well for different the three
very very well for different the three
very very well for different the three different systems
yeah um um yeah uh it's very close to uh
yeah um um yeah uh it's very close to uh
yeah um um yeah uh it's very close to uh to the final slides but yeah um here I I
to the final slides but yeah um here I I
to the final slides but yeah um here I I want to also want to introduce some
want to also want to introduce some
want to also want to introduce some interesting thing uh it's a perspective
interesting thing uh it's a perspective
interesting thing uh it's a perspective from the explainable AI but uh yeah
from the explainable AI but uh yeah
from the explainable AI but uh yeah um basically the explainable AI we could
um basically the explainable AI we could
um basically the explainable AI we could regard we could associate it with is R
regard we could associate it with is R
regard we could associate it with is R regression for example consider if we
regression for example consider if we
regression for example consider if we have a r regression uh we use a r
have a r regression uh we use a r
have a r regression uh we use a r regression model to fit uh based on a
regression model to fit uh based on a
regression model to fit uh based on a lot of exp explainable variables then
lot of exp explainable variables then
lot of exp explainable variables then for each variables explainable variables
for each variables explainable variables
for each variables explainable variables we could have the feature importance
we could have the feature importance
we could have the feature importance based on the fitting uh efficience of
based on the fitting uh efficience of
based on the fitting uh efficience of the r regression model uh here for the s
the r regression model uh here for the s
the r regression model uh here for the s model uh it's the same we can here we
model uh it's the same we can here we
model uh it's the same we can here we just take the uh layer uh layer
just take the uh layer uh layer
just take the uh layer uh layer relevance propagation algorithm to
relevance propagation algorithm to
relevance propagation algorithm to extract the uh feature importance uh of
extract the uh feature importance uh of
extract the uh feature importance uh of each uh input data uh this could yeah
each uh input data uh this could yeah
each uh input data uh this could yeah just regard it just associate it uh with
just regard it just associate it uh with
just regard it just associate it uh with the r regression featureing importance
the r regression featureing importance
the r regression featureing importance uh this would be very similar diagram
uh this would be very similar diagram
uh this would be very similar diagram but here uh it's a prediction for the
but here uh it's a prediction for the
but here uh it's a prediction for the set node system and just 30 time steps
set node system and just 30 time steps
set node system and just 30 time steps before the RTP happens then if we
before the RTP happens then if we
before the RTP happens then if we compare the uh uh layerwise propagation
compare the uh uh layerwise propagation
compare the uh uh layerwise propagation lrp score uh before uh for different
lrp score uh before uh for different
lrp score uh before uh for different time series uh for different uh
time series uh for different uh
time series uh for different uh different point of the time Ser along
different point of the time Ser along
different point of the time Ser along the time series then uh we expect to
the time series then uh we expect to
the time series then uh we expect to find out some precursor signal uh heed
find out some precursor signal uh heed
find out some precursor signal uh heed in in the time seral segments at
in in the time seral segments at
in in the time seral segments at different time lead time uh forecast
different time lead time uh forecast
different time lead time uh forecast lead time for example here if we compare
lead time for example here if we compare
lead time for example here if we compare the uh compare aring and nontipping
the uh compare aring and nontipping
the uh compare aring and nontipping categories of the for the lrp score we
categories of the for the lrp score we
categories of the for the lrp score we can find it's very different and
can find it's very different and
can find it's very different and specifically we find two uh difference
specifically we find two uh difference
specifically we find two uh difference uh fature feature importance one is very
uh fature feature importance one is very
uh fature feature importance one is very close to the Tipping uh the end of the
close to the Tipping uh the end of the
close to the Tipping uh the end of the time series segment this is uh
time series segment this is uh
time series segment this is uh comparably easy to understand because
comparably easy to understand because
comparably easy to understand because this means the tping uh the r tping r
this means the tping uh the r tping r
this means the tping uh the r tping r tping probability are highly related to
tping probability are highly related to
tping probability are highly related to the uh the current time step the current
the uh the current time step the current
the uh the current time step the current time point of the system State for the
time point of the system State for the
time point of the system State for the system state but Al also we find a
system state but Al also we find a
system state but Al also we find a fingerprint uh it's a fixed fingerprint
fingerprint uh it's a fixed fingerprint
fingerprint uh it's a fixed fingerprint uh
uh
uh uh around around a uh time window uh
uh around around a uh time window uh
uh around around a uh time window uh that is uh um 50 times dep to 100 times
that is uh um 50 times dep to 100 times
that is uh um 50 times dep to 100 times dep uh before the arping happens and we
dep uh before the arping happens and we
dep uh before the arping happens and we notice this peak and for this peak it's
notice this peak and for this peak it's
notice this peak and for this peak it's very interesting where uh we realize
very interesting where uh we realize
very interesting where uh we realize it's uh it there must be some dynamical
it's uh it there must be some dynamical
it's uh it there must be some dynamical features can be uh already captured by
features can be uh already captured by
features can be uh already captured by the deing model then there is a peak
the deing model then there is a peak
the deing model then there is a peak here so basically for the uh uh for the
here so basically for the uh uh for the
here so basically for the uh uh for the model for the deping uh for the RTP uh
model for the deping uh for the RTP uh
model for the deping uh for the RTP uh uh for the RTP here uh there are two
uh for the RTP here uh there are two
uh for the RTP here uh there are two elements uh for the RTP one is the
elements uh for the RTP one is the
elements uh for the RTP one is the forcing rate second second uh second one
forcing rate second second uh second one
forcing rate second second uh second one is the uh is the noise pations but for
is the uh is the noise pations but for
is the uh is the noise pations but for the noise pations we checked it but not
the noise pations we checked it but not
the noise pations we checked it but not but we didn't find any special thing
but we didn't find any special thing
but we didn't find any special thing related to this fingerprint but for the
related to this fingerprint but for the
related to this fingerprint but for the foring rate we found some interesting
foring rate we found some interesting
foring rate we found some interesting thing here uh because the time serious
thing here uh because the time serious
thing here uh because the time serious uh the time series here they have
uh the time series here they have
uh the time series here they have different uh maximal uh foring rates
different uh maximal uh foring rates
different uh maximal uh foring rates maximal that is the uh uh just now here
maximal that is the uh uh just now here
maximal that is the uh uh just now here for the forcing profile if we calculate
for the forcing profile if we calculate
for the forcing profile if we calculate the
the
the difference uh difference time series of
difference uh difference time series of
difference uh difference time series of the the foring profile then we can get
the the foring profile then we can get
the the foring profile then we can get the instant forcing rate also yeah it's
the instant forcing rate also yeah it's
the instant forcing rate also yeah it's instant shift rate also then we
instant shift rate also then we
instant shift rate also then we can we just make a composition and find
can we just make a composition and find
can we just make a composition and find uh some early R tping and The Late Late
uh some early R tping and The Late Late
uh some early R tping and The Late Late RTP that means the early or late is
RTP that means the early or late is
RTP that means the early or late is comparably defined by the Yeah the
comparably defined by the Yeah the
comparably defined by the Yeah the maximum of the instant forcing
maximum of the instant forcing
maximum of the instant forcing rate for example like the early RTP The
rate for example like the early RTP The
rate for example like the early RTP The Tipping tipping time is very close to
Tipping tipping time is very close to
Tipping tipping time is very close to the maximal uh maximal forcing rate uh
the maximal uh maximal forcing rate uh
the maximal uh maximal forcing rate uh usually for the uh in previous studies
usually for the uh in previous studies
usually for the uh in previous studies people would think would s would think
people would think would s would think
people would think would s would think when the uh for external forcing reach
when the uh for external forcing reach
when the uh for external forcing reach is maximum forcing rate this is most
is maximum forcing rate this is most
is maximum forcing rate this is most dangerous for the uh most dangerous
dangerous for the uh most dangerous
dangerous for the uh most dangerous period for the system to undergo to have
period for the system to undergo to have
period for the system to undergo to have a tipping points uh but here we with
a tipping points uh but here we with
a tipping points uh but here we with find that it's relatively late to the
find that it's relatively late to the
find that it's relatively late to the maximum forcing uh but uh we also find
maximum forcing uh but uh we also find
maximum forcing uh but uh we also find another uh some other uh samples uh the
another uh some other uh samples uh the
another uh some other uh samples uh the Tipping time is much much more late much
Tipping time is much much more late much
Tipping time is much much more late much later than the maximal forcing uh rate
later than the maximal forcing uh rate
later than the maximal forcing uh rate and uh here if we the forcing rate uh if
and uh here if we the forcing rate uh if
and uh here if we the forcing rate uh if we have looked at forcing rate it's
we have looked at forcing rate it's
we have looked at forcing rate it's exactly uh sh the C
exactly uh sh the C
exactly uh sh the C time window with the fingerprint
time window with the fingerprint
time window with the fingerprint here yeah and uh uh finally we also uh
here yeah and uh uh finally we also uh
here yeah and uh uh finally we also uh have a look we also want to talk about
have a look we also want to talk about
have a look we also want to talk about something about the selection of the uh
something about the selection of the uh
something about the selection of the uh Naro Network architecture because I
Naro Network architecture because I
Naro Network architecture because I mentioned uh we just use a very simple
mentioned uh we just use a very simple
mentioned uh we just use a very simple thing and and somebody would argue oh
thing and and somebody would argue oh
thing and and somebody would argue oh why why just use a very simple thing and
why why just use a very simple thing and
why why just use a very simple thing and it's uh it's a old dog but but I have to
it's uh it's a old dog but but I have to
it's uh it's a old dog but but I have to say the Naro Network s is very powerful
say the Naro Network s is very powerful
say the Naro Network s is very powerful but it's simple uh if we select a simple
but it's simple uh if we select a simple
but it's simple uh if we select a simple naral Network architecture then this
naral Network architecture then this
naral Network architecture then this could be uh helpful for the community to
could be uh helpful for the community to
could be uh helpful for the community to understand it but here we also uh uh we
understand it but here we also uh uh we
understand it but here we also uh uh we also compel it way compare different
also compel it way compare different
also compel it way compare different just take took different Naro Network
just take took different Naro Network
just take took different Naro Network architectures like the LM and also the
architectures like the LM and also the
architectures like the LM and also the uh MLP or self attention mechanism and
uh MLP or self attention mechanism and
uh MLP or self attention mechanism and then makes the prediction this is the
then makes the prediction this is the
then makes the prediction this is the our curve of the prediction results uh
our curve of the prediction results uh
our curve of the prediction results uh it seems uh s in lstm or the combination
it seems uh s in lstm or the combination
it seems uh s in lstm or the combination of LM CN uh this is exactly Thomas Barry
of LM CN uh this is exactly Thomas Barry
of LM CN uh this is exactly Thomas Barry ever used in his P paper and also
ever used in his P paper and also
ever used in his P paper and also compare with and also the self attention
compare with and also the self attention
compare with and also the self attention they have the comp terrible prediction
they have the comp terrible prediction
they have the comp terrible prediction ability uh maybe for the boting system
ability uh maybe for the boting system
ability uh maybe for the boting system uh it's chaotic uh Dynamics uh really
uh it's chaotic uh Dynamics uh really
uh it's chaotic uh Dynamics uh really dist uh uh disturb the performance of
dist uh uh disturb the performance of
dist uh uh disturb the performance of the MLP but here yeah we uh anyway we
the MLP but here yeah we uh anyway we
the MLP but here yeah we uh anyway we could we could conclude that uh PR de
could we could conclude that uh PR de
could we could conclude that uh PR de planing indicator actually it's
planing indicator actually it's
planing indicator actually it's independent on as specific naron net nwk
independent on as specific naron net nwk
independent on as specific naron net nwk architecture we can in the future we can
architecture we can in the future we can
architecture we can in the future we can even consider a more uh comprehensive or
even consider a more uh comprehensive or
even consider a more uh comprehensive or complicated architecture like
complicated architecture like
complicated architecture like Transformer or or generative uh deing
Transformer or or generative uh deing
Transformer or or generative uh deing model to yeah to get maybe can get
model to yeah to get maybe can get
model to yeah to get maybe can get better results but here the same thing
better results but here the same thing
better results but here the same thing is also works very well and it's
is also works very well and it's
is also works very well and it's efficient in
efficient in
efficient in computation uh yeah yeah and and I
computation uh yeah yeah and and I
computation uh yeah yeah and and I I I also want to uh highlight some some
I I also want to uh highlight some some
I I also want to uh highlight some some useful thing for the depending uh uh
useful thing for the depending uh uh
useful thing for the depending uh uh indicator for the we know for the
indicator for the we know for the
indicator for the we know for the critical slowing down it's always
critical slowing down it's always
critical slowing down it's always calculated based on the uh time serious
calculated based on the uh time serious
calculated based on the uh time serious articulation and variance and this means
articulation and variance and this means
articulation and variance and this means we usually even though we for a specific
we usually even though we for a specific
we usually even though we for a specific system like Arctic uh like the an an
system like Arctic uh like the an an
system like Arctic uh like the an an Atlantic illation uh illation o ocean
Atlantic illation uh illation o ocean
Atlantic illation uh illation o ocean illation we we could have a lot of
illation we we could have a lot of
illation we we could have a lot of multiple um we could have a multiple
multiple um we could have a multiple
multiple um we could have a multiple time Ser for for it but usually we can
time Ser for for it but usually we can
time Ser for for it but usually we can only calculate credit cost Lo Down based
only calculate credit cost Lo Down based
only calculate credit cost Lo Down based one of the time series and but here for
one of the time series and but here for
one of the time series and but here for the deing model we can just simply add
the deing model we can just simply add
the deing model we can just simply add the uh add more uh explainable uh more
the uh add more uh explainable uh more
the uh add more uh explainable uh more dynamical variables into the CM just at
dynamical variables into the CM just at
dynamical variables into the CM just at simply at one additional Channel then we
simply at one additional Channel then we
simply at one additional Channel then we could make the prediction uh
could make the prediction uh
could make the prediction uh for by by involving more uh dynamical
for by by involving more uh dynamical
for by by involving more uh dynamical variables but here it seems uh for the
variables but here it seems uh for the
variables but here it seems uh for the boting system compos boom system they
boting system compos boom system they
boting system compos boom system they also have a multiple multiple dynamical
also have a multiple multiple dynamical
also have a multiple multiple dynamical variables but if we include them them
variables but if we include them them
variables but if we include them them into the uh de planning prediction it
into the uh de planning prediction it
into the uh de planning prediction it seems It's comparable to the uh yeah a
seems It's comparable to the uh yeah a
seems It's comparable to the uh yeah a single variable based uh
single variable based uh
single variable based uh indicator yeah this is the last slide
indicator yeah this is the last slide
indicator yeah this is the last slide and uh yeah it's a short summary but uh
and uh yeah it's a short summary but uh
and uh yeah it's a short summary but uh I really want to talk something uh here
I really want to talk something uh here
I really want to talk something uh here this could be yeah this could be useful
this could be yeah this could be useful
this could be yeah this could be useful uh for our future work firstly we know
uh for our future work firstly we know
uh for our future work firstly we know the uh de planning uh the planning can
the uh de planning uh the planning can
the uh de planning uh the planning can advance our ability to predict raing de
advance our ability to predict raing de
advance our ability to predict raing de tping and noise induc tping and uh then
tping and noise induc tping and uh then
tping and noise induc tping and uh then we can just uh get make use of the
we can just uh get make use of the
we can just uh get make use of the experience of emble realizations and
experience of emble realizations and
experience of emble realizations and just make you uh just apply the them
just make you uh just apply the them
just make you uh just apply the them into the influence on individual time
into the influence on individual time
into the influence on individual time series and also we find uh the deing
series and also we find uh the deing
series and also we find uh the deing model can have good transferability
model can have good transferability
model can have good transferability across uh across different out of sample
across uh across different out of sample
across uh across different out of sample forcing rates and out of sample
forcing rates and out of sample
forcing rates and out of sample dynamical system
dynamical system
dynamical system but of course if we train all the models
but of course if we train all the models
but of course if we train all the models together
together
together overall we must can we can must get a
overall we must can we can must get a
overall we must can we can must get a very uh comprehensive deing model and
very uh comprehensive deing model and
very uh comprehensive deing model and very powerful but in the future um it's
very powerful but in the future um it's
very powerful but in the future um it's promising uh for us uh to predict
promising uh for us uh to predict
promising uh for us uh to predict tipping uh across the critical forcing
tipping uh across the critical forcing
tipping uh across the critical forcing value this is bation tipping diagram and
value this is bation tipping diagram and
value this is bation tipping diagram and also in addition the uh critical forcing
also in addition the uh critical forcing
also in addition the uh critical forcing rates and uh then we can yeah we
rates and uh then we can yeah we
rates and uh then we can yeah we hopefully
hopefully
hopefully um we hopefully compare to the critical
um we hopefully compare to the critical
um we hopefully compare to the critical slowing down we can have much more
slowing down we can have much more
slowing down we can have much more comprehensive early warning indicator
comprehensive early warning indicator
comprehensive early warning indicator also make of the explan explainable AI
also make of the explan explainable AI
also make of the explan explainable AI algorithm we could discover more
algorithm we could discover more
algorithm we could discover more Fingerprints of the Tipping points for
Fingerprints of the Tipping points for
Fingerprints of the Tipping points for some specific um systems like the um
some specific um systems like the um
some specific um systems like the um Antarctic ocean o
Antarctic ocean o
Antarctic ocean o occilations this is very Wishful and we
occilations this is very Wishful and we
occilations this is very Wishful and we really want to um uh further focus on
really want to um uh further focus on
really want to um uh further focus on this points but of course uh there is
this points but of course uh there is
this points but of course uh there is still open question here because just
still open question here because just
still open question here because just now uh
now uh
now uh the the the Tipping uh early warning uh
the the the Tipping uh early warning uh
the the the Tipping uh early warning uh indicator is based on the Tipping probab
indicator is based on the Tipping probab
indicator is based on the Tipping probab probability it's basically it's it's
probability it's basically it's it's
probability it's basically it's it's aesthetic a matric uh but uh of course
aesthetic a matric uh but uh of course
aesthetic a matric uh but uh of course alter alternatively uh we could also
alter alternatively uh we could also
alter alternatively uh we could also have a tipping trajectory prediction
have a tipping trajectory prediction
have a tipping trajectory prediction that is uh stand on current climate we
that is uh stand on current climate we
that is uh stand on current climate we want we can we maybe we want also want
want we can we maybe we want also want
want we can we maybe we want also want to predict the time seral the uh EV time
to predict the time seral the uh EV time
to predict the time seral the uh EV time uh evolutional uh trajectory in the
uh evolutional uh trajectory in the
uh evolutional uh trajectory in the future scenario and this is the teing
future scenario and this is the teing
future scenario and this is the teing trajectory but currently in in our de
trajectory but currently in in our de
trajectory but currently in in our de planning model architecture this cannot
planning model architecture this cannot
planning model architecture this cannot is is not un achievable but in a very
is is not un achievable but in a very
is is not un achievable but in a very recent paper published in uh physical
recent paper published in uh physical
recent paper published in uh physical review uh uh research and also uh last
review uh uh research and also uh last
review uh uh research and also uh last just last out of out in last month and
just last out of out in last month and
just last out of out in last month and uh they they already tried to do
uh they they already tried to do
uh they they already tried to do something along this direction but we
something along this direction but we
something along this direction but we really hope this this all the progress
really hope this this all the progress
really hope this this all the progress could help us to better to help us in uh
could help us to better to help us in uh
could help us to better to help us in uh serve for the um tpping points uh
serve for the um tpping points uh
serve for the um tpping points uh Community yeah and this is my email if
Community yeah and this is my email if
Community yeah and this is my email if you are interested and uh just feel free
you are interested and uh just feel free
you are interested and uh just feel free to contact me and also uh last week our
to contact me and also uh last week our
to contact me and also uh last week our paper was uh p uh was published out uh
paper was uh p uh was published out uh
paper was uh p uh was published out uh in nature machine intelligence and uh if
in nature machine intelligence and uh if
in nature machine intelligence and uh if you want to know more details yeah you
you want to know more details yeah you
you want to know more details yeah you could also yeah find more in in that
could also yeah find more in in that
could also yeah find more in in that paper yeah thank you much I think I'm
paper yeah thank you much I think I'm
paper yeah thank you much I think I'm done thank you for your time yeah thank
done thank you for your time yeah thank
done thank you for your time yeah thank you very much thank you any questions or
comments yes I
comments yes I
comments yes I yeah I'm assuming you mean me yeah yeah
yeah I'm assuming you mean me yeah yeah
yeah I'm assuming you mean me yeah yeah sorry sorry I me Lou yeah sorry okay
sorry sorry I me Lou yeah sorry okay
sorry sorry I me Lou yeah sorry okay yeah yeah interesting talk uh I'm aware
yeah yeah interesting talk uh I'm aware
yeah yeah interesting talk uh I'm aware of a recent paper by yining live the
of a recent paper by yining live the
of a recent paper by yining live the Arizona State University and his group
Arizona State University and his group
Arizona State University and his group looking at uh the critical point at
looking at uh the critical point at
looking at uh the critical point at which uh a crisis happens in a chaotic
which uh a crisis happens in a chaotic
which uh a crisis happens in a chaotic attractor as you vary a parameter in the
attractor as you vary a parameter in the
attractor as you vary a parameter in the system yes and they did this they did
system yes and they did this they did
system yes and they did this they did this um learning of that using a
this um learning of that using a
this um learning of that using a reservoir computer and they did a pretty
reservoir computer and they did a pretty
reservoir computer and they did a pretty good job of telling where that was going
good job of telling where that was going
good job of telling where that was going to happen uh unlike your there was no
to happen uh unlike your there was no
to happen uh unlike your there was no noise on it uh I recall anyway uh
noise on it uh I recall anyway uh
noise on it uh I recall anyway uh however they did Miss uh they could not
however they did Miss uh they could not
however they did Miss uh they could not predict a window in the chaotic system
predict a window in the chaotic system
predict a window in the chaotic system I'm wondering uh if you have tried a
I'm wondering uh if you have tried a
I'm wondering uh if you have tried a chaotic system something like that and
chaotic system something like that and
chaotic system something like that and tried to predict a u a bifurcation and
tried to predict a u a bifurcation and
tried to predict a u a bifurcation and whether you've looked into using
whether you've looked into using
whether you've looked into using Reservoir computers rather than a neural
Reservoir computers rather than a neural
Reservoir computers rather than a neural network yeah yeah thank you uh I think
network yeah yeah thank you uh I think
network yeah yeah thank you uh I think this uh the paper you mentioned exactly
this uh the paper you mentioned exactly
this uh the paper you mentioned exactly I exactly the paper here okay yeah I
I exactly the paper here okay yeah I
I exactly the paper here okay yeah I point here yeah we also already noticed
point here yeah we also already noticed
point here yeah we also already noticed it is very it's very amazing um yeah
it is very it's very amazing um yeah
it is very it's very amazing um yeah firstly for your uh for your question
firstly for your uh for your question
firstly for your uh for your question let's focus on your questions uh firstly
let's focus on your questions uh firstly
let's focus on your questions uh firstly um way uh we for our de planning model
um way uh we for our de planning model
um way uh we for our de planning model CM based indicator we have not tried for
CM based indicator we have not tried for
CM based indicator we have not tried for the bation tipping but yeah the reason
the bation tipping but yeah the reason
the bation tipping but yeah the reason is this is already done by Thomas bur uh
is this is already done by Thomas bur uh
is this is already done by Thomas bur uh he he has a p publication on pns it's um
he he has a p publication on pns it's um
he he has a p publication on pns it's um 2021 and uh he already demonstrates
2021 and uh he already demonstrates
2021 and uh he already demonstrates using a lstm plus c architecture here uh
using a lstm plus c architecture here uh
using a lstm plus c architecture here uh can achieve very nice prediction results
can achieve very nice prediction results
can achieve very nice prediction results for the bation Tipping even for the uh
for the bation Tipping even for the uh
for the bation Tipping even for the uh yeah I mean I I know maybe you mean a
yeah I mean I I know maybe you mean a
yeah I mean I I know maybe you mean a specific chaotic system like the bation
specific chaotic system like the bation
specific chaotic system like the bation in the
in the
in the Lawrence 63 system maybe it's yeah I
Lawrence 63 system maybe it's yeah I
Lawrence 63 system maybe it's yeah I noticed in uh in ch Li group they also
noticed in uh in ch Li group they also
noticed in uh in ch Li group they also uh have the yeah have a paper about this
uh have the yeah have a paper about this
uh have the yeah have a paper about this point about this direction but yeah
point about this direction but yeah
point about this direction but yeah thank you uh we will consider about this
thank you uh we will consider about this
thank you uh we will consider about this a specific CL uh chaotic system but uh
a specific CL uh chaotic system but uh
a specific CL uh chaotic system but uh also we have not tried it but uh I I
also we have not tried it but uh I I
also we have not tried it but uh I I trust I I I would believe uh San could
trust I I I would believe uh San could
trust I I I would believe uh San could work for it because uh because here uh
work for it because uh because here uh
work for it because uh because here uh we have the boting system the boting
we have the boting system the boting
we have the boting system the boting system itself is already very C have a
system itself is already very C have a
system itself is already very C have a very uh a lot of have a various uh
very uh a lot of have a various uh
very uh a lot of have a various uh chaotic oscillations in this boting
chaotic oscillations in this boting
chaotic oscillations in this boting system
system
system so here our model works for boting
so here our model works for boting
so here our model works for boting system then I yeah I would believe it
system then I yeah I would believe it
system then I yeah I would believe it can also work for the LA
can also work for the LA
can also work for the LA system thank you very much but yeah
system thank you very much but yeah
system thank you very much but yeah thank you yeah but also maybe for
thank you yeah but also maybe for
thank you yeah but also maybe for another question uh maybe your related
another question uh maybe your related
another question uh maybe your related to your question uh is the like the uh
to your question uh is the like the uh
to your question uh is the like the uh in uh inance group uh they uh your
in uh inance group uh they uh your
in uh inance group uh they uh your Reservoir computer is also very
Reservoir computer is also very
Reservoir computer is also very efficient in computation and uh uh it
efficient in computation and uh uh it
efficient in computation and uh uh it can predict the trajectory tipping
can predict the trajectory tipping
can predict the trajectory tipping trajectory and uh uh but but based on
trajectory and uh uh but but based on
trajectory and uh uh but but based on the uh recent two paper
the uh recent two paper
the uh recent two paper Publications uh we have a very close uh
Publications uh we have a very close uh
Publications uh we have a very close uh look at the paper uh they yeah and they
look at the paper uh they yeah and they
look at the paper uh they yeah and they they already did a lot of nice work on
they already did a lot of nice work on
they already did a lot of nice work on the bation tipping but since they have
the bation tipping but since they have
the bation tipping but since they have not done uh test if it also works for
not done uh test if it also works for
not done uh test if it also works for the WR tipping but yeah I'm also very
the WR tipping but yeah I'm also very
the WR tipping but yeah I'm also very curious if it were yeah what will
curious if it were yeah what will
curious if it were yeah what will happens when using Reservoir computer
happens when using Reservoir computer
happens when using Reservoir computer predicts the Tipping trajectory uh in
predicts the Tipping trajectory uh in
predicts the Tipping trajectory uh in rping diagram this could be very amazing
rping diagram this could be very amazing
rping diagram this could be very amazing yeah yeah thank you
much all right uh other
questions all right seems not yeah thank
questions all right seems not yeah thank
questions all right seems not yeah thank you very much you for coming and thanks
you very much you for coming and thanks
you very much you for coming and thanks again for the nice talk and thanks all
again for the nice talk and thanks all
again for the nice talk and thanks all for for coming yeah see you next week
for for coming yeah see you next week
for for coming yeah see you next week yeah byebye yeah byebye
Can't find what you're looking for?
Get subtitles in any language from opensubtitles.com, and translate them here.