15 March 2023…
GFDL Lunchtime Seminar Series (Presentation): Using explainable AI to identify key regions of climate change in GFDL SPEAR large ensembles, Princeton, NJ.
References...
Labe, Z.M., E.A. Barnes, and J.W. Hurrell (2023). Identifying the regional emergence of climate patterns in the ARISE-SAI-1.5 simulations. https://doi.org/10.31223/X5394Z (submitted)
Human & Veterinary Respiratory Physilogy_DR.E.Muralinath_Associate Professor....
Using explainable AI to identify key regions of climate change in GFDL SPEAR large ensembles
1. Using explainable AI to identify
key regions of climate change in
GFDL SPEAR large ensembles
https://zacklabe.com/ @ZLabe
Zachary Labe
Postdoc in Seasonal-to-Decadal (S2D) Variability and Predictability Division
with Nathaniel Johnson and Thomas Delworth
15 March 2023
GFDL Lunchtime Seminar
3. 2
1) Where do we
go from here?
2) How do we
disentangle
internal climate
variability?
Feb/Mar 2016
4. 3
3) How do we
account for
regional
patterns of
change?
5. Explainable machine learning can
distinguish between regional patterns
of time-evolving climate change
in GFDL models.
4
SIGNIFICANCE
6. Machine Learning
is not new!
“A Bayesian Neural Network for
Severe-Hail Prediction (2000)”
“Classification of Convective Areas
Using Decision Trees (2009)”
“A Neural Network for Damaging
Wind Prediction (1998)”
“Generative Additive Models versus
Linear Regression in Generating
Probabilistic MOS Forecasts of
Aviation Weather Parameters (1995)”
”A Neural Network for
Tornado Prediction
Based on Doppler
Radar-Derived
Attributes (1996)”
”The Diagnosis of
Upper-Level Humidity
(1968)”
5
7. “An adaptive data processing system for weather forecasting”
It’s a neural network!
[Hu and Root (1964), APME]
6
10. Do it better
e.g., parameterizations in climate models are not
perfect, use ML to make them more accurate
Do it faster
e.g., code in climate models is very slow (but we
know the right answer) - use ML methods to speed
things up
Do something new
• e.g., go looking for non-linear relationships you
didn’t know were there
WHY ELSE SHOULD WE CONSIDER
MACHINE LEARNING?
9
11. Do it better
e.g., parameterizations in climate models are not
perfect, use ML to make them more accurate
Do it faster
e.g., code in climate models is very slow (but we
know the right answer) - use ML methods to speed
things up
Do something new
• e.g., go looking for non-linear relationships you
didn’t know were there
Very relevant for
research: may be
slower and worse,
but can still learn
something
10
WHY ELSE SHOULD WE CONSIDER
MACHINE LEARNING?
12. Machine learning for meteorology
IDENTIFYING SEVERE THUNDERSTORMS
Molina et al. 2021
Martin et al. 2022
CLASSIFYING PHASE OF MADDEN-JULLIAN OSCILLATION
DETECTING CONVECTION FROM SATELLITES
Lee et al. 2021
LOCATING COLD FRONTS
Dagon et al. 2022
11
13. Machine learning for oceanography
CLASSIFYING ARCTIC OCEAN ACIDIFICATION
Krasting et al. 2022
LARGE-SCALE OCEAN CIRCULATION
Clare et al. 2022
ESTIMATING OCEAN SURFACE CURRENTS
Sinha and Abernathey, 2021
12
14. Machine learning for climate
PHYSICAL DRIVERS OF ENSO DYNAMICS
Shin et al. 2022
IDENTIFYING DECADAL STATE DEPENDENCE
Gordon and Barnes, 2022
INTERNAL/EXTERNAL CLIMATE FORCING
Po-Chedley et al. 2022
13
21. Artificial Neural Networks [ANN]
X1
X2
W1
W2
∑
INPUTS
NODE
Linear regression with non-linear
mapping by an “activation function”
Training of the network is merely
determining the weights “w” and
bias/offset “b"
= factivation(X1W1+ X2W2 + b)
20
24. Complexity and nonlinearities of the ANN allow it to learn
many different pathways of predictable behavior
Once trained, you have an array of weights and biases
which can be used for prediction on new data
INPUT
[DATA]
PREDICTION
Artificial Neural Networks [ANN]
23
25. What is the annual mean temperature of Earth?
24
27. What is the annual mean temperature of Earth?
THE REAL WORLD
(Observations)
Let’s run a
climate model
One ensemble member
2022
1930 2050
26
Data
from
SPEAR_M
ED
28. What is the annual mean temperature of Earth?
THE REAL WORLD
(Observations)
Let’s run a
climate model
again!
Two ensemble members
27
Data
from
SPEAR_M
ED
29. What is the annual mean temperature of Earth?
THE REAL WORLD
(Observations)
Let’s run a
climate model
again & again!
Three ensemble members
28
Data
from
SPEAR_M
ED
30. What is the annual mean temperature of Earth?
THE REAL WORLD
(Observations)
CLIMATE MODEL
LARGE ENSEMBLE
30 ensemble
members in
GFDL SPEAR
29
31. What is the annual mean temperature of Earth?
Mean of ensembles
= forced response (climate change)
THE REAL WORLD
(Observations)
CLIMATE MODEL
LARGE ENSEMBLE
30
ensemble members
In GFDL SPEAR
30
32. What is the annual mean temperature of Earth?
Mean of ensembles
= forced response (climate change)
THE REAL WORLD
(Observations)
CLIMATE MODEL
LARGE ENSEMBLE
30
ensemble members
In GFDL SPEAR
Range of ensembles
= internal variability (noise)
Mean of ensembles
= forced response (climate change)
31
33. But let’s remove
climate change…
Climate Change Signal
(ensemble mean)
Observations
Ensemble
Members
32
41. Yearly Maps of T2M
Yearly Maps of T2M
Neural
Network
Classify
Climate
Scenario
Artificial
Neural
Network
Output
=
5
Classes
Yearly Maps of T2M
Neural
Network
Binary Output Binary Output
40
Step #1
Read in gridded maps of a
climate variable from
SPEAR simulations
42. Yearly Maps of T2M
Yearly Maps of T2M
Neural
Network
Classify
Climate
Scenario
Artificial
Neural
Network
Output
=
5
Classes
Yearly Maps of T2M
Neural
Network
Binary Output Binary Output
41
Step #2
Feed data into an
artificial neural network
with three hidden layers
43. Yearly Maps of T2M
Yearly Maps of T2M
Neural
Network
Classify
Climate
Scenario
Artificial
Neural
Network
Output
=
5
Classes
Yearly Maps of T2M
Neural
Network
Binary Output Binary Output
42
Step #3
Classify which climate
scenario (n=5) is
associated with each map
44. Yearly Maps of T2M
Yearly Maps of T2M
Neural
Network
Classify
Climate
Scenario
Artificial
Neural
Network
Output
=
5
Classes
Yearly Maps of T2M
Neural
Network
Binary Output Binary Output
43
Step #4
Why? à XAI
45. WHY
LAYER-WISE RELEVANCE PROPAGATION (LRP)
Volcano
Timber
Wolf
Image Classification LRP
LRP heatmaps show regions
of “relevance” that
contribute to the neural
network’s decision-making
process for a sample
belonging to a particular
output category
Neural Network
WHY
Backpropagation – LRP
https://heatmapping.org/ 44
46. WHY
LAYER-WISE RELEVANCE PROPAGATION (LRP)
Volcano
Timber
Wolf
Image Classification LRP
LRP heatmaps show regions
of “relevance” that
contribute to the neural
network’s decision-making
process for a sample
belonging to a particular
output category
Neural Network
WHY
Backpropagation – LRP
https://heatmapping.org/ 45
51. Neural
Network
[0] La Niña [1] El Niño
Input a map of sea surface temperatures
50
[Toms et al. 2020, JAMES]
52. Visualizing something we already know…
Input maps of sea surface
temperatures (SST) to
identify El Niño or La Niña
Use ‘LRP’ to see how the
neural network is making
its decision
[Toms et al. 2020, JAMES]
Layer-wise Relevance Propagation
Composite SST Observations
LRP [Relevance]
SST Anomaly [°C]
0.00 0.75
0.0 1.5
-1.5
51
Warmer
Colder
High
Low
62. 61
What if we start
mitigation
10 years earlier?
Transition from
SSP5-8.5 to SSP2-4.5
Transition from
SSP5-8.5 to SSP2-4.5
Transition from
SSP2-4.5 to SSP1-1.9
2015 2060 2100
2015 2060 2100
63. 62
What climate
patterns are
associated with
these transitions?
Transition from
SSP5-8.5 to SSP2-4.5
Transition from
SSP5-8.5 to SSP2-4.5
Transition from
SSP2-4.5 to SSP1-1.9
Rapid mitigation
Rapid mitigation
64. 63
Difficult to distinguish
the patterns
associated with
scenario transitions
Composites of
relevance maps for
the mitigation
predictions
SSP5-3.4OS example for 2015 to 2100
Nearer to predicted scenario
Further from predicted scenario
66. 65
XAI composites of years associated with the transition from SSP5-8.5 to SSP2-4.5
(a) approx. 2055-2060 (b) approx. 2040-2045
67. 66
North Atlantic is an
important indictor
region for climate
signals related to
identifying from
SSP5-8.5 to SSP2-4.5
Future Climate Change Rapid Mitigation
68. 67
Framework can be
applied to different
geographic regions
and climate variables
Parallel approach for
detecting climate
intervention scenarios
Labe, Z.M., E.A. Barnes, and J.W. Hurrell (2023). Identifying the
regional emergence of climate patterns in the ARISE-SAI-1.5
simulations. EarthArXiv, DOI: 10.31223/X5394Z
69. 1) Explainable AI methods identify
regions of climate change that
indicate the emission scenario
2) Framework for outlining benefits
from mitigation by testing out of
sample ensembles from SSP5-3.4OS
3) Patterns of fast and slow responses
to rapid climate mitigation are
revealed by explainable AI
KEY FINDINGS
• Indicator
patterns of
climate
change
• Quantify
benefits of
rapid climate
mitigation
• Regions of
rapid or slow
response to
mitigation
Labe, Z.M., N.C. Johnson, and T.L. Delworth (2023), A data-driven approach to
identifying key regions of change associated with rapid climate mitigation, in prep
70. 1) Explainable AI methods identify
regions of climate change that
indicate the emission scenario
2) Framework for outlining benefits
from mitigation by testing out of
sample ensembles from SSP5-3.4OS
3) Patterns of fast and slow responses
to rapid climate mitigation are
revealed by explainable AI
KEY FINDINGS
• Indicator
patterns of
climate
change
• Quantify
benefits of
rapid climate
mitigation
• Regions of
rapid or slow
response to
mitigation
Labe, Z.M., N.C. Johnson, and T.L. Delworth (2023), A data-driven approach to
identifying key regions of change associated with rapid climate mitigation, in prep
71. 1) Explainable AI methods identify
regions of climate change that
indicate the emission scenario
2) Framework for outlining benefits
from mitigation by testing out of
sample ensembles from SSP5-3.4OS
3) Patterns of fast and slow responses
to rapid climate mitigation are
revealed by explainable AI
KEY FINDINGS
• Indicator
patterns of
climate
change
• Quantify
benefits of
rapid climate
mitigation
• Regions of
rapid or slow
response to
mitigation
Labe, Z.M., N.C. Johnson, and T.L. Delworth (2023), A data-driven approach to
identifying key regions of change associated with rapid climate mitigation, in prep
76. Temperature anomalies [ °C ] relative to 1981-2010
Observations from NClimGrid
Climate model data from GFDL SPEAR_MED
United States – Summer
75
1920 2020
80. We know some metadata…
+ What year is it?
+ Where did it come from?
[Labe and Barnes, 2022; ESS]
TEMPERATURE
79
81. TEMPERATURE
Neural network learns nonlinear
combinations of forced climate
patterns to identify the year
We know some metadata…
+ What year is it?
+ Where did it come from?
[Labe and Barnes, 2022; ESS]
80
82. ----ANN----
2 Hidden Layers
10 Nodes each
Ridge Regularization
Early Stopping
[e.g., Barnes et al. 2019, 2020]
[e.g., Labe and Barnes, 2021]
TIMING OF EMERGENCE
(COMBINED VARIABLES)
RESPONSES TO
EXTERNAL CLIMATE
FORCINGS
PATTERNS OF
CLIMATE INDICATORS
[e.g., Rader et al. 2022]
Surface Temperature Map Precipitation Map
+
TEMPERATURE
We know some metadata…
+ What year is it?
+ Where did it come from?
[Labe and Barnes, 2022; ESS]
81
92. June – August – Timing of Emergence (ToE) for observations
91
June – August – skill for observations
93. How is the neural network able to detect the year prior to ~1990?
Temperature anomalies [ °C ] relative to 1981-2010
Machine learning predictions GFDL SPEAR_MED simulation
92
99. First year that the 10-year running-mean
temperature exceeds and stays above the
mean 1921–1950 reference temperature by
more than two standard deviations
Most Areas = 2000s
98
108. TRENDS FROM 1921 TO 1950
Fully-Coupled [Historical + SSP5-8.5] SPEAR_MED, but NO anthropogenic aerosols SPEAR_MED, but NO anthropogenic forcings
107
Warmer
Colder
109. TRENDS FROM 1921 TO 1950
Fully-Coupled [Historical + SSP5-8.5] SPEAR_MED, but NO anthropogenic aerosols SPEAR_MED, but NO anthropogenic forcings
108
Warmer
Colder
110. TRENDS FROM 1921 TO 1950
Fully-Coupled [Historical + SSP5-8.5] SPEAR_MED, but NO anthropogenic aerosols SPEAR_MED, but NO anthropogenic forcings
109
Warmer
Colder
111. TRENDS IN EVAPORATION
Fully-Coupled [Historical + SSP5-8.5] SPEAR_MED, but NO anthropogenic aerosols SPEAR_MED, but NO anthropogenic forcings
110
Increase
Decrease
115. 1. Increasing spatial resolution
improves the neural network skill
for identifying climate signals
2. Externally-forced temperature
signals have emerged in
observations in the United States
3. Trends in western United States
land surface fields are linked to
timing of emergence in SPEAR
KEY FINDINGS
• Machine
learning
prediction
skill
• ToE in
observations
• Physical
drivers of
climate
signals
Labe, Z.M., N.C. Johnson, and T.L. Delworth (2023), Forced signals in United States summer
temperatures revealed by explainable neural networks in climate models and observations, in prep
116. 1. Increasing spatial resolution
improves the neural network skill
for identifying climate signals
2. Externally-forced temperature
signals have emerged in
observations in the United States
3. Trends in western United States
land surface fields are linked to
timing of emergence in SPEAR
KEY FINDINGS
• Machine
learning
prediction
skill
• ToE in
observations
• Physical
drivers of
climate
signals
Labe, Z.M., N.C. Johnson, and T.L. Delworth (2023), Forced signals in United States summer
temperatures revealed by explainable neural networks in climate models and observations, in prep
117. 1. Increasing spatial resolution
improves the neural network skill
for identifying climate signals
2. Externally-forced temperature
signals have emerged in
observations in the United States
3. Trends in western United States
land surface fields are linked to
timing of emergence in SPEAR
KEY FINDINGS
• Machine
learning
prediction
skill
• ToE in
observations
• Physical
drivers of
climate
signals
Labe, Z.M., N.C. Johnson, and T.L. Delworth (2023), Forced signals in United States summer
temperatures revealed by explainable neural networks in climate models and observations, in prep
119. XAI can identify regional patterns of
climate change & variability
in GFDL large ensembles.
1)
118
120. Method can identify differences in
time-evolving climate signals between
other climate model large ensembles.
2)
119
121. Framework can be adapted for
monitoring and predicting patterns
of climate change in observations.
3)
120
122. DECADAL CLIMATE PREDICTION
Explainable machine learning for improving prediction skill
and identifying physical drivers
DETECTION AND ATTRIBUTION
Classification neural network for monitoring extreme events
in climate models and observations
FUTURE DIRECTIONS AT GFDL
DIAGNOSTIC TOOL FOR MODEL BIASES
Method for extracting forced climate signals across Earth system models
123. TAKEAWAYS
1. XAI can identify regional patterns of climate change & variability in GFDL large ensembles.
2. Method can identify differences in time-evolving forced climate signals between other
climate model large ensembles.
3. Framework can be adapted for monitoring and predicting patterns of climate change in
observations.
Zack Labe
zachary.labe@noaa.gov
122
15 March 2023
GFDL Lunchtime Seminar