Presentation for: GFDL/AOS Summer Internship Lecture Series
The popularity of machine learning is rapidly growing in nearly all areas of Earth science. However, there is also some hesitancy in adopting the use of these methods due to concerns about their reliability, reproducibility, and interpretability – thus, they are often described as “black boxes.”
In this talk, I will introduce a few simple examples from climate science that leverage new visualization methods to peer into the machine learning “black box,” which help us to better understand their predictions while also learning new science. These same machine learning visualization tools can be easily adapted for a wide variety of applications and other scientific fields of study.
Labelling Requirements and Label Claims for Dietary Supplements and Recommend...
Learning new climate science by thinking creatively with machine learning
1. Learning new climate science by
thinking creatively with
machine learning
@ZLabe
Zachary M. Labe
Postdoc in Princeton AOS and NOAA GFDL
28 June 2022
Summer Internship Lecture Series
2. Zack Labe
Pioneer Coal Mine Blue Whale of Catoosa Centralia Underground Fire Roadside America Greenland Sea (81°N)
Enjoy: roadside oddities, diners, and horror movies
Hobbies: gardening, #scicomm ( @ZLabe), hiking, traveling to lighthouses
Graduate work: Arctic – midlatitude climate variability (sea-ice changes)
Recent work: using ANNs to detect patterns of climate variability/change
Now at GFDL I am thinking about: detection and attribution of extreme events
Relevance
LENS X-Single Forcing Runs
Hometown – Linglestown,
Pennsylvania
BSc – Atmospheric Science
at Cornell University
PhD – Earth System Science
at UC Irvine
Postdoc – climate variability
using AI at Colorado State U.
Postdoc – climate attribution
with T. Delworth & N. Johnson
he/him
7. Computer Science
Artificial Intelligence
Machine Learning
Deep Learning
Supervised
Learning
Unsupervised
Learning
Labeled data
Classification
Regression
Unlabeled data
Clustering
Dimension reduction
8. • 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
WHY SHOULD WE CONSIDER
MACHINE LEARNING?
11. Adapted from: Kotamarthi, R., Hayhoe, K., Mearns, L., Wuebbles, D., Jacobs, J., & Jurado, J.
(2021). Global Climate Models. In Downscaling Techniques for High-Resolution Climate
Projections: From Global Change to Local Impacts (pp. 19-39). Cambridge: Cambridge University
Press. doi:10.1017/9781108601269.003
CLIMATE MODELS
Horizontal Grid
Vertical Levels
Past/Present/Future
Fully-Coupled System
20-40 Petabytes of data
17. Today’s weather or climate
scientist is far more likely to be
debugging code written in
Python… than to be poring over
satellite images or releasing
radiosondes.
“
D. Irving| Bulletin of the American Meteorological Society| 2016
18. Machine learning for weather
IDENTIFYING SEVERE THUNDERSTORMS
Molina et al. 2021
Toms et al. 2021
CLASSIFYING PHASE OF MADDEN-JULLIAN OSCILLATION
SATELLITE DETECTION
Lee et al. 2021
DETECTING TORNADOES
McGovern et al. 2019
19. Machine learning for climate
FINDING FORECASTS OF OPPORTUNITY
Mayer and Barnes, 2021
PREDICTING CLIMATE MODES OF VARIABILITY
Gordon et al. 2021
TIMING OF CLIMATE CHANGE
Barnes et al. 2019
20. Machine learning for oceanography
CLASSIFYING ARCTIC OCEAN ACIDIFICATION
Krasting et al. 2022
TRACK AND REVEAL DEEP WATER MASSES
Sonnewald and Lguensat, 2021
ESTIMATING OCEAN SURFACE CURRENTS
Sinha and Abernathey, 2021
26. E.g.,
Establish robust,
responsible AI for
severe weather
detection
Tornado Warning
Special Marine Warning
Severe Thunderstorm Warning
Flash Flood Warning
30. What could be a problem
with this type of model fitting?
Question
31. Training Data (1880-1970) Validation Data (1971-2014) Testing Data (2015-2100)
Algorithm learns patterns in the
data to build a prediction model
Tune the hyperparameters to select
the ‘best’ model for your task
Evaluate the skill of your prediction
model on data it has never seen before
Input maps of sea surface temperatures to identify El Niño or La Niña.
Does the machine learning model have the same skill in a future climate?
32. Training Data (1880-1970) Validation Data (1971-2014) Testing Data (2015-2100)
Algorithm learns patterns in the
data to build a prediction model
Tune the hyperparameters to select
the ‘best’ model for your task
Evaluate the skill of your prediction
model on data it has never seen before
Input maps of sea surface temperatures to identify El Niño or La Niña.
Does the machine learning model have the same skill in a future climate?
33. Training Data (1880-1970) Validation Data (1971-2014) Testing Data (2015-2100)
Algorithm learns patterns in the
data to build a prediction model
Tune the hyperparameters to select
the ‘best’ model for your task
Evaluate the skill of your prediction
model on data it has never seen before
Input maps of sea surface temperatures to identify El Niño or La Niña.
Does the machine learning model have the same skill in a future climate?
34. Training Data (1880-1970) Validation Data (1971-2014) Testing Data (2015-2100)
Algorithm learns patterns in the
data to build a prediction model
Tune the hyperparameters to select
the ‘best’ model for your task
Evaluate the skill of your prediction
model on data it has never seen before
Input maps of sea surface temperatures to identify El Niño or La Niña.
Does the machine learning model have the same skill in a future climate?
37. 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)
38. Artificial Neural Networks [ANN]
X1
X2
W1
W2
∑
INPUTS
NODE
= factivation(X1W1+ X2W2 + b)
ReLU Sigmoid Linear
40. 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]
43. We know some metadata…
+ What year is it?
+ Where did it come from?
TEMPERATURE
44. We know some metadata…
+ What year is it?
+ Where did it come from?
TEMPERATURE
Neural network learns nonlinear
combinations of forced climate
patterns to identify the year
45. ----ANN----
2 Hidden Layers
10 Nodes each
Ridge Regularization
Early Stopping
We know some metadata…
+ What year is it?
+ Where did it come from?
[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, in review]
Surface Temperature Map Precipitation Map
+
TEMPERATURE
46. ----ANN----
2 Hidden Layers
10 Nodes each
Ridge Regularization
Early Stopping
We know some metadata…
+ What year is it?
+ Where did it come from?
[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, in review]
Surface Temperature Map Precipitation Map
+
TEMPERATURE
48. What is the annual mean temperature of Earth?
THE REAL WORLD
(Observations)
Anomaly is relative to 1951-1980
49. What is the annual mean temperature of Earth?
THE REAL WORLD
(Observations)
Let’s run a
climate model
50. What is the annual mean temperature of Earth?
THE REAL WORLD
(Observations)
Let’s run a
climate model
again
51. What is the annual mean temperature of Earth?
THE REAL WORLD
(Observations)
Let’s run a
climate model
again & again
52. What is the annual mean temperature of Earth?
THE REAL WORLD
(Observations)
CLIMATE MODEL
ENSEMBLES
53. What is the annual mean temperature of Earth?
THE REAL WORLD
(Observations)
CLIMATE MODEL
ENSEMBLES
Range of ensembles
= internal variability (noise)
Mean of ensembles
= forced response (climate change)
54. What is the annual mean temperature of Earth?
Range of ensembles
= internal variability (noise)
Mean of ensembles
= forced response (climate change)
But let’s remove
climate change…
55. What is the annual mean temperature of Earth?
Range of ensembles
= internal variability (noise)
Mean of ensembles
= forced response (climate change)
After removing the
forced response…
anomalies/noise!
56. 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)
57. 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)
59. 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
Observations
60. So what?
Greenhouse gases = warming
Aerosols = ?? (though mostly cooling)
What are the relative responses
between greenhouse gas
and aerosol forcing?
63. INPUT LAYER
Surface Temperature Map
ARTIFICIAL NEURAL NETWORK (ANN)
Collection of nodes (neurons)
that adjust their weights and
biases across layers in order to
learn signals for making
predictions
Learns nonlinear processes
through selected parameters
in the model
64. INPUT LAYER
HIDDEN LAYERS
OUTPUT LAYER
Surface Temperature Map
“2000-2009”
DECADE CLASS
“2070-2079”
“1920-1929”
ARTIFICIAL NEURAL NETWORK (ANN)
65. 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)
66. 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, JAMES]
67. LAYER-WISE RELEVANCE PROPAGATION (LRP)
Volcano
Great White
Shark
Timber
Wolf
Image Classification LRP
https://heatmapping.org/
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
WHY
WHY
Backpropagation – LRP
68. LAYER-WISE RELEVANCE PROPAGATION (LRP)
Volcano
Great White
Shark
Timber
Wolf
Image Classification LRP
https://heatmapping.org/
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
WHY
WHY
Backpropagation – LRP
69. LAYER-WISE RELEVANCE PROPAGATION (LRP)
Volcano
Great White
Shark
Timber
Wolf
Image Classification LRP
https://heatmapping.org/
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
WHY
WHY
WHY
70. LAYER-WISE RELEVANCE PROPAGATION (LRP)
Image Classification LRP
https://heatmapping.org/
NOT PERFECT
Crock
Pot
Neural Network
Backpropagation – LRP
WHY
72. Neural
Network
[0] La Niña [1] El Niño
[Toms et al. 2020, JAMES]
Input a map of sea surface temperatures
73. Visualizing something we already know…
Input maps of sea surface
temperatures 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 Observations
LRP [Relevance]
SST Anomaly [°C]
0.00 0.75
0.0 1.5
-1.5
74. [Adapted from Adebayo et al., 2020]
EXPLAINABLE AI IS
NOT PERFECT
THERE ARE MANY
METHODS
75. [Adapted from Adebayo et al., 2020]
THERE ARE MANY
METHODS
EXPLAINABLE AI IS
NOT PERFECT
76. OUTPUT LAYER
Layer-wise Relevance Propagation
“2000-2009”
DECADE CLASS
“2070-2079”
“1920-1929”
BACK-PROPAGATE THROUGH NETWORK = EXPLAINABLE AI
WHY?
= LRP HEAT MAPS
[Labe and Barnes 2021, JAMES]
78. Layer-wise Relevance Propagation
BACK-PROPAGATE THROUGH NETWORK = EXPLAINABLE AI
WHY?
= LRP HEAT MAPS
Find regions of “relevance”
that contribute to the
neural network’s
decision-making process
[Labe and Barnes 2021, JAMES]
79. 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
80. 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
81. 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
82. 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
83. CLIMATE MODEL DATA PREDICT THE YEAR FROM MAPS OF TEMPERATURE
AEROSOLS
PREVAIL
GREENHOUSE GASES
PREVAIL
STANDARD
CLIMATE MODEL
[Labe and Barnes 2021, JAMES]
84. OBSERVATIONS PREDICT THE YEAR FROM MAPS OF TEMPERATURE
AEROSOLS
PREVAIL
GREENHOUSE GASES
PREVAIL
STANDARD
CLIMATE MODEL
[Labe and Barnes 2021, JAMES]
85. OBSERVATIONS
SLOPES
PREDICT THE YEAR FROM MAPS OF TEMPERATURE
AEROSOLS
PREVAIL
GREENHOUSE GASES
PREVAIL
STANDARD
CLIMATE MODEL
[Labe and Barnes 2021, JAMES]
93. ----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?
94. 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
96. 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
107. WE CAN LEARN NEW SCIENCE
FROM EXPLAINABLE AI.
3)
108.
109. KEY POINTS
1. Machine learning is just another tool to add to our scientific workflow
2. We can use explainable AI (XAI) methods to peer into the black box of machine learning
3. We can learn new science by using XAI methods in conjunction with existing statistical tools
Zachary Labe
zachary.labe@noaa.gov
@ZLabe
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
Labe, Z.M. and E.A. Barnes (2022), Comparison of climate model large ensembles with observations in
the Arctic using simple neural networks, in review, DOI:10.1002/essoar.10510977.2
110. • Are there any challenges you have experienced in pre/post-processing
Earth science data?
• If traditional statistics already perform well on a prediction task, should
we even bother trying machine learning techniques for it?
• What are some potential issues for using machine learning methods in
your summer research projects?
• What are some of the best ways to get started in creating a machine
learning model and selecting hyperparameters?
Discussion