The relative roles of individual forcings on large-scale climate variability remain difficult to disentangle within fully-coupled global climate model simulations. Here, we train an artificial neural network (ANN) to classify the climate forcings of a new set of CESM1 initial-condition large ensembles that are forced by different combinations of aerosol (industrial and biomass burning), greenhouse gas, and land-use/land-cover forcings. As a result of learning the regional responses of internal variability to the different external forcings, the ANN is able to successfully classify the dominant forcing for each model simulation. Using recently developed explainable AI methods, such as layerwise relevance propagation, we then compare the patterns of climate variability identified by the ANN between different external climate forcings that are learned by the neural network. Further, we apply this ANN architecture on additional climate simulations from the multi-model large ensemble archive, which include all anthropogenic and natural radiative forcings. From this collection of initial-condition ensembles, the ANN is also able to detect changes in atmospheric internal variability between the 20th and 21st centuries by training on climate fields after the mean forced signal has already been removed. This ANN framework and its associated visualization tools provide a novel approach to extract complex patterns of observable and projected climate variability and trends in Earth system models. (from https://ams.confex.com/ams/101ANNUAL/meetingapp.cgi/Paper/379553)
Discovery of an Accretion Streamer and a Slow Wide-angle Outflow around FUOri...
Disentangling Climate Forcing in Multi-Model Large Ensembles Using Neural Networks
1. DISENTANGLING CLIMATE FORCING
IN SINGLE-FORCING LARGE ENSEMBLES
USING NEURAL NETWORKS
@ZLabe
Zachary M. Labe & Elizabeth A. Barnes
Department of Atmospheric Science at Colorado State University
14 January 2021
20th Conference on Artificial Intelligence for Environmental Science
101st AMS Annual Meeting
3. What is the annual mean temperature of Earth?
THE REAL WORLD
(Observations)
Anomaly is relative to 1951-1980
4. What is the annual mean temperature of Earth?
THE REAL WORLD
(Observations)
Let’s run a
climate model
5. What is the annual mean temperature of Earth?
THE REAL WORLD
(Observations)
Let’s run a
climate model
again
6. What is the annual mean temperature of Earth?
THE REAL WORLD
(Observations)
Let’s run a
climate model
again & again
7. What is the annual mean temperature of Earth?
THE REAL WORLD
(Observations)
CLIMATE MODEL
ENSEMBLES
8. What is the annual mean temperature of Earth?
THE REAL WORLD
(Observations)
Range of ensembles
= internal variability (noise)
Mean of ensembles
= forced response (climate change)
CLIMATE MODEL
ENSEMBLES
9. What is the annual mean temperature of Earth?
• Increasing greenhouse gases (CO2, CH4, N2O)
• Changes in industrial aerosols (SO4, BC, OC)
• Changes in biomass burning (aerosols)
• Changes in land-use & land-cover (albedo)
10. What is the annual mean temperature of Earth?
• Increasing greenhouse gases (CO2, CH4, N2O)
• Changes in industrial aerosols (SO4, BC, OC)
• Changes in biomass burning (aerosols)
• Changes in land-use & land-cover (albedo)
Plus everything else…
(Natural/internal variability)
11. What is the annual mean temperature of Earth?
• Increasing greenhouse gases (CO2, CH4, N2O)
• Changes in industrial aerosols (SO4, BC, OC)
• Changes in biomass burning (aerosols)
• Changes in land-use & land-cover (albedo)
ALL
13. Greenhouse gases fixed to 1920 levels
All forcings (CESM-LE)
Industrial aerosols fixed to 1920 levels
[Deser et al. 2020, JCLI]
Fully-coupled CESM1.1
20 Ensemble Members
Run from 1920-2080
Reanalysis
14. So what?
Greenhouse gases = warming
Aerosols = ?? (though mostly cooling)
What are the relative responses
between greenhouse gas
and aerosol forcing?
17. INPUT LAYER
HIDDEN LAYERS
OUTPUT LAYER
Surface Temperature Map
“2000-2009”
DECADE CLASS
“2070-2079”
“1920-1929”
ARTIFICIAL NEURAL NETWORK (ANN)
18. INPUT LAYER
HIDDEN LAYERS
OUTPUT LAYER
Surface Temperature Map
“2000-2009”
DECADE CLASS
“2070-2079”
“1920-1929”
BACK-PROPAGATE THROUGH NETWORK = EXPLAINABLE AI
ARTIFICIAL NEURAL NETWORK (ANN)
19. INPUT LAYER
HIDDEN LAYERS
OUTPUT LAYER
Layer-wise Relevance Propagation
Surface Temperature Map
“2000-2009”
DECADE CLASS
“2070-2079”
“1920-1929”
BACK-PROPAGATE THROUGH NETWORK = EXPLAINABLE AI
ARTIFICIAL NEURAL NETWORK (ANN)
[Barnes et al. 2020, JAMES]
[Labe and Barnes 2021, submitted]
20. LAYER-WISE RELEVANCE PROPAGATION (LRP)
Volcano
Great White
Shark
Timber
Wolf
Image Classification LRP
https://heatmapping.org/ [Geoscience examples in Toms et al. 2020, JAMES]
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
WHYWHYWHY
Backpropagation – LRP
21. LAYER-WISE RELEVANCE PROPAGATION (LRP)
Volcano
Great White
Shark
Timber
Wolf
Image Classification LRP
https://heatmapping.org/ [Geoscience examples in Toms et al. 2020, JAMES]
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
WHYWHYWHY
Backpropagation – LRP
22. LAYER-WISE RELEVANCE PROPAGATION (LRP)
Volcano
Great White
Shark
Timber
Wolf
Image Classification LRP
https://heatmapping.org/ [Geoscience examples in Toms et al. 2020, JAMES]
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
Backpropagation – LRP
WHYWHYWHY
23. LAYER-WISE RELEVANCE PROPAGATION (LRP)
Image Classification LRP
https://heatmapping.org/ [Geoscience examples in Toms et al. 2020, JAMES]
NOT PERFECTCrock
Pot
Neural Network
Backpropagation – LRP
WHY
24. OUTPUT LAYER
Layer-wise Relevance Propagation
“2000-2009”
DECADE CLASS
“2070-2079”
“1920-1929”
BACK-PROPAGATE THROUGH NETWORK = EXPLAINABLE AI
[Labe and Barnes 2021, submitted]
WHY?= LRP HEAT MAPS
25. 1960-1999: ANNUAL MEAN TEMPERATURE TRENDS
Greenhouse gases fixed
to 1920 levels
[AEROSOLS PREVAIL]
Industrial aerosols fixed
to 1920 levels
[GREENHOUSE GASES PREVAIL]
All forcings
[STANDARD CESM-LE]
DATA
26. 1960-1999: ANNUAL MEAN TEMPERATURE TRENDS
Greenhouse gases fixed
to 1920 levels
[AEROSOLS PREVAIL]
Industrial aerosols fixed
to 1920 levels
[GREENHOUSE GASES PREVAIL]
All forcings
[STANDARD CESM-LE]
DATA
27. 1960-1999: ANNUAL MEAN TEMPERATURE TRENDS
Greenhouse gases fixed
to 1920 levels
[AEROSOLS PREVAIL]
Industrial aerosols fixed
to 1920 levels
[GREENHOUSE GASES PREVAIL]
All forcings
[STANDARD CESM-LE]
DATA
28. 1960-1999: ANNUAL MEAN TEMPERATURE TRENDS
Greenhouse gases fixed
to 1920 levels
[AEROSOLS PREVAIL]
Industrial aerosols fixed
to 1920 levels
[GREENHOUSE GASES PREVAIL]
All forcings
[STANDARD CESM-LE]
DATA
29. CLIMATE MODEL DATA PREDICT THE YEAR FROM MAPS OF TEMPERATURE
[Labe and Barnes 2021, submitted]
37. 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
[Labe and Barnes 2021, submitted]
Uncertainty in LRP
38. Uncertainty in LRP
Ultimately, we are trying to
mask noise in the LRP output
Identify robust climate pattern indicators!
[Labe and Barnes 2021, submitted]
44. KEY POINTS
Zachary Labe
zmlabe@rams.colostate.edu
@ZLabe
1. Using explainable AI methods with artificial neural networks (ANNs) reveals climate patterns in
large ensemble simulations
2. Metric proposed for quantifying the uncertainty of an ANN visualization method that extracts
signals from different external forcings
3. ANN trained using a large ensemble simulation without time-evolving aerosols makes more
accurate predictions of real world data