26 March 2024…
GFDL Polar Climate Interest Group (Presentation): An intro to explainable AI for polar climate science, NOAA GFDL, Princeton, NJ.
References:
Labe, Z.M. and E.A. Barnes (2022), Comparison of climate model large ensembles with observations in the Arctic using simple neural networks. Earth and Space Science, DOI:10.1029/2022EA002348, https://doi.org/10.1029/2022EA002348
Labe, Z.M. and E.A. Barnes (2021), Detecting climate signals using explainable AI with single-forcing large ensembles. Journal of Advances in Modeling Earth Systems, DOI:10.1029/2021MS002464, https://agupubs.onlinelibrary.wiley.com/doi/abs/10.1029/2021MS002464
Analytical Profile of Coleus Forskohlii | Forskolin .pptx
An intro to explainable AI for polar climate science
1. An intro to
explainable AI for
polar climate science
Zachary M. Labe
Postdoc in Seasonal-to-Decadal (S2D) Variability and Predictability Division
with Elizabeth A. Barnes (CSU), Thomas L. Delworth (GFDL), and Nathaniel C. Johnson (GFDL)
26 March 2024
Polar Climate
Interest Group Meeting
https://zacklabe.com/ @ZLabe
2. 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?
3. 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?
Very relevant for
research: may be
slower and worse,
but can still learn
something
5. STANDARD EVALUATION OF
CLIMATE MODELS
Pattern correlation
RMSE
EOFs
Trends, anomalies, mean state
Climate modes of variability
Negative Correlation Positive Correlation
PATTERN CORRELATION – T2M
PATTERN CORRELATION : NEAR-SURFACE AIR TEMPERATURE
8. ----ANN----
2 Hidden Layers
10 Nodes each
Ridge Regularization
Early Stopping
TEMPERATURE
We know some metadata…
+ What year is it? (Labe & Barnes, 2021)
+ Where did it come from?
9. TEMPERATURE
We know some metadata…
+ What year is it? (Labe & Barnes, 2021)
+ Where did it come from?
Train on data from the
Multi-Model Large
Ensemble Archive
10. TEMPERATURE
We know some metadata…
+ What year is it? (Labe & Barnes, 2021)
+ Where did it come from?
NEURAL NETWORK
CLASSIFICATION TASK
HIDDEN LAYERS
INPUT LAYER
INPUT LAYER
OUTPUT LAYER
HIDDEN LAYERS
15. 2-m Actual Air Temperature (°C)
THERE ARE MANY CLIMATE MODEL LARGE ENSEMBLES…
Annual mean 2-m temperature
7 global climate models
16 ensembles each
ERA5 (observations)
17. EXPLAINABLE AI (XAI)
1. Is the prediction correct for the right reasons?
• Is it consistent with our physical understanding of the climate system?
2. Provide insights for improving the machine learning model
• Is the model overfitting? Can the model be further optimized?
3. Learn new science
• For example, in climate prediction this could be a new forecast of opportunity or teleconnection
https://doi.org/10.1175/AIES-D-22-0001.1
21. EXPLAINABLE AI (XAI)
Sensitivity: refers to how much the value of the output will
change for a unit change in a specific feature
Such as… Gradient (Saliency Maps), Smooth Gradient (first derivative of the output with respect to input)
Signal: all the information in the input that is relevant to the
prediction task (i.e., signal component versus distractor)
Such as… PatternNet
Attribution: refers to the relative contribution of an input
feature to the output
Such as… Input*Gradient, Integrated Gradients, Layer-wise Relevance Propagation (LRP), Deep Taylor, DeepSHAP
https://doi.org/10.1175/AIES-D-22-0012.1
22. ATTRIBUTION-BASED XAI METHODS
Volcano
Great White
Shark
Timber
Wolf
Image Classification XAI
https://heatmapping.org/
XAI 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
WHY
WHY
Backpropagation Rules
30. Neural
Network
[0] La Niña [1] El Niño
Input a map of sea surface temperatures
[Toms et al. 2020, JAMES]
31. 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 (LRP)
Composite SST Observations
LRP [Relevance]
SST Anomaly [°C]
0.00 0.75
0.0 1.5
-1.5
Warmer
Colder
High
Low
32. Visualizing something we already know…
Input maps of sea surface
temperatures (SST) to
identify El Niño or La Niña
Use ‘Backward Optimization’ to
identify synthetic input that
maximizes the neural network’s
confidence of the prediction
[Toms et al. 2020, JAMES]
Backward Optimization
Composite SST Observations
Optimal Input
SST Anomaly [°C]
-1.0 1.0
0.0 1.5
-1.5
Warmer
Colder
Warmer
Colder
0.0
40. 1. Shuffle ensemble member and year
dimensions (bootstrap-like method)
2. Apply true labels (unshuffled years)
3. Apply same ANN architecture and LRP
4. Repeat 500x by using different
combinations of training/testing data and
initialization seeds
5. Compute 95th percentile of the distribution
of LRP at all grid points
Uncertainty for XAI
[Labe and Barnes 2021, JAMES]
41. Uncertainty for XAI
Ultimately, we are trying to
mask noise in the LRP output
Identify robust climate pattern indicators!
[Labe and Barnes 2021, JAMES]
45. Interpretable vs. Explainable
Explainable AI (XAI): method to explain black box after
training model – approximate model behavior
Interpretable AI: model is inherently interpretable and
provides own explanation – degree to which a model can
be understood
https://www.nature.com/articles/s42256-019-0048-x
a priori a posterio
NO CONSENSUS!
46. Adapted from McGovern et al. (2022, EDS) at https://doi.org/10.1017/eds.2022.5
ETHICAL, RESPONSIBLE, TRUSTWORTHY AI
1. Issues related to training data
q Non-representative training data, including lack of geo-diversity
q Training labels are biased or faulty
q Data is affected by adversaries
2. Issues related to AI models
q Model training choices
q Algorithms learns faulty strategies
q AI learns to fake something plausible
q AI model used in inappropriate situations
q Non-trustworthy AI model deployed
q Lack of robustness in the AI model
3. Other issues related to workforce and society
q Globally applicable AI approaches may stymie burgeoning efforts in developing countries
q Lack of input or consent on data collection and model training
q Scientists might feel disenfranchised
q Increase of carbon emissions due to computing
McGovern
et
al.
(2024,
AI)
57. KEY POINTS
1. Artificial neural network can identify which climate model produced an annual mean map of
near-surface temperature in the Arctic
2. Classification network is evaluated using input from atmospheric reanalysis as a method of
comparing climate models and observations
3. XAI method reveals regional temperature patterns the artificial neural network is using to
classify observations with different climate models
Labe, Z. M., & Barnes, E. A. (2022). Comparison of climate model large
ensembles with observations in the Arctic using simple neural networks.
Earth and Space Science, 9(7), e2022EA002348
https://doi.org/10.1029/2022EA002348