Successfully reported this slideshow.
Your SlideShare is downloading. ×

Assessing climate variability and change with explainable neural networks

Ad
Ad
Ad
Ad
Ad
Ad
Ad
Ad
Ad
Ad

Check these out next

1 of 108 Ad
Advertisement

More Related Content

Slideshows for you (20)

Advertisement
Advertisement

Assessing climate variability and change with explainable neural networks

  1. 1. ASSESSING CLIMATE VARIABILITY & CHANGE WITH EXPLAINABLE NEURAL NETWORKS @ZLabe Zachary M. Labe with Elizabeth A. Barnes Colorado State University Department of Atmospheric Science 13 October 2021 Lunchtime Seminar GFDL – Princeton AOS
  2. 2. Machine Learning is not new!
  3. 3. Machine Learning is not new!
  4. 4. Artificial Intelligence Machine Learning Deep Learning Computer Science
  5. 5. Computer Science Artificial Intelligence Machine Learning Deep Learning Supervised Learning Unsupervised Learning Labeled data Classification Regression Unlabeled data Clustering Dimension reduction
  6. 6. • 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?
  7. 7. Machine learning for weather IDENTIFYING SEVERE THUNDERSTORMS Molina et al. 2021 Toms et al. 2021 CLASSIFYING PHASE OF MADDEN-JULLIAN OSCILLATION SATELLITE-CONVECTION DETECTION Lee et al. 2021 DETECTING TORNADOES McGovern et al. 2019
  8. 8. Machine learning for climate FINDING FORECASTS OF OPPORTUNITY Mayer and Barnes, 2021 PREDICTING CLIMATE MODES OF VARIABILITY Gordon et al. 2021, ESSOAr TIMING OF EMERGENCE Barnes et al. 2019
  9. 9. INPUT [DATA] PREDICTION Machine Learning
  10. 10. INPUT [DATA] PREDICTION Machine Learning
  11. 11. Artificial Intelligence Machine Learning Deep Learning
  12. 12. X1 X2 INPUTS Artificial Neural Networks [ANN]
  13. 13. Linear regression! Artificial Neural Networks [ANN] X1 X2 W1 W2 ∑ = X1W1+ X2W2 + b INPUTS NODE
  14. 14. 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)
  15. 15. X1 X2 ∑ inputs HIDDEN LAYERS X3 ∑ ∑ ∑ OUTPUT = predictions Artificial Neural Networks [ANN] : : :: INPUTS
  16. 16. 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]
  17. 17. TEMPERATURE
  18. 18. TEMPERATURE We know some metadata… + What year is it? + Where did it come from?
  19. 19. We know some metadata… + What year is it? + Where did it come from? TEMPERATURE
  20. 20. 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
  21. 21. ----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. 2021, in prep] Surface Temperature Map Precipitation Map + TEMPERATURE
  22. 22. ----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. 2021, in prep] Surface Temperature Map Precipitation Map + TEMPERATURE
  23. 23. THE REAL WORLD (Observations) What is the annual mean temperature of Earth?
  24. 24. What is the annual mean temperature of Earth? THE REAL WORLD (Observations) Anomaly is relative to 1951-1980
  25. 25. 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) A CLIMATE MODEL (CESM1.1-LE)
  26. 26. 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) A CLIMATE MODEL (CESM1.1-LE)
  27. 27. What is the annual mean temperature of Earth? [CESM1 "Single Forcing" Large Ensemble Project]
  28. 28. 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
  29. 29. So what? Greenhouse gases = warming Aerosols = ?? (though mostly cooling) What are the relative responses between greenhouse gas and aerosol forcing?
  30. 30. Surface Temperature Map ARTIFICIAL NEURAL NETWORK (ANN)
  31. 31. INPUT LAYER Surface Temperature Map ARTIFICIAL NEURAL NETWORK (ANN)
  32. 32. INPUT LAYER HIDDEN LAYERS OUTPUT LAYER Surface Temperature Map “2000-2009” DECADE CLASS “2070-2079” “1920-1929” ARTIFICIAL NEURAL NETWORK (ANN)
  33. 33. 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)
  34. 34. 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]
  35. 35. 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]
  36. 36. Layer-wise Relevance Propagation BACK-PROPAGATE THROUGH NETWORK = EXPLAINABLE AI WHY? = LRP HEAT MAPS Machine Learning Black Box [Labe and Barnes 2021, JAMES]
  37. 37. 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]
  38. 38. 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
  39. 39. 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
  40. 40. 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
  41. 41. LAYER-WISE RELEVANCE PROPAGATION (LRP) Image Classification LRP https://heatmapping.org/ NOT PERFECT Crock Pot Neural Network WHY Backpropagation – LRP
  42. 42. [Adapted from Adebayo et al., 2020] EXPLAINABLE AI IS NOT PERFECT THERE ARE MANY METHODS
  43. 43. [Adapted from Adebayo et al., 2020] THERE ARE MANY METHODS EXPLAINABLE AI IS NOT PERFECT
  44. 44. Visualizing something we already know…
  45. 45. Neural Network [0] La Niña [1] El Niño [Toms et al. 2020, JAMES] Input a map of sea surface temperatures
  46. 46. 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
  47. 47. 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]
  48. 48. 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 Colder Warmer
  49. 49. 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 Colder Warmer
  50. 50. 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
  51. 51. 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 Colder Warmer
  52. 52. CLIMATE MODEL DATA PREDICT THE YEAR FROM MAPS OF TEMPERATURE AEROSOLS PREVAIL GREENHOUSE GASES PREVAIL STANDARD CLIMATE MODEL [Labe and Barnes 2021, JAMES]
  53. 53. OBSERVATIONS PREDICT THE YEAR FROM MAPS OF TEMPERATURE AEROSOLS PREVAIL GREENHOUSE GASES PREVAIL STANDARD CLIMATE MODEL [Labe and Barnes 2021, JAMES]
  54. 54. OBSERVATIONS SLOPES PREDICT THE YEAR FROM MAPS OF TEMPERATURE AEROSOLS PREVAIL GREENHOUSE GASES PREVAIL STANDARD CLIMATE MODEL [Labe and Barnes 2021, JAMES]
  55. 55. [Labe and Barnes 2021, JAMES] ARE THE RESULTS ROBUST? YES! COMBINATIONS OF TRAINING/TESTING DATA – 100 x
  56. 56. Low High HOW DID THE ANN MAKE ITS PREDICTIONS?
  57. 57. Low High HOW DID THE ANN MAKE ITS PREDICTIONS? WHY IS THERE GREATER SKILL FOR GHG+?
  58. 58. RESULTS FROM LRP [Labe and Barnes 2021, JAMES] Low High
  59. 59. Low High RESULTS FROM LRP [Labe and Barnes 2021, JAMES] WHAT IS SIGNIFICANT?
  60. 60. RESULTS FROM LRP [Labe and Barnes 2021, JAMES] Low High
  61. 61. Higher LRP values indicate greater relevance for the ANN’s prediction AVERAGED OVER 1960-2039 Aerosol-driven Greenhouse gas-driven All forcings Low High [Labe and Barnes 2021, JAMES]
  62. 62. Greenhouse gas-driven Aerosol-driven All forcings AVERAGED OVER 1960-2039 [Labe and Barnes 2021, JAMES]
  63. 63. DISTRIBUTIONS OF LRP AVERAGED OVER 1960-2039 [Labe and Barnes 2021, JAMES] Less relevant More relevant
  64. 64. DISTRIBUTIONS OF LRP AVERAGED OVER 1960-2039 [Labe and Barnes 2021, JAMES] Less relevant More relevant
  65. 65. ----ANN---- 2 Hidden Layers 10 Nodes each Ridge Regularization Early Stopping TEMPERATURE We know some metadata… + What year is it? + Where did it come from?
  66. 66. TEMPERATURE We know some metadata… + What year is it? + Where did it come from? Train on data from the Multi-Model Large Ensemble Archive
  67. 67. TEMPERATURE We know some metadata… + What year is it? + Where did it come from? NEURAL NETWORK CLASSIFICATION TASK HIDDEN LAYERS INPUT LAYER
  68. 68. STANDARD EVALUATION OF CLIMATE MODELS Pattern correlation RMSE EOFs Trends, anomalies, mean state Climate modes of variability
  69. 69. COMPARING CLIMATE MODELS LRP (Explainable AI) Raw data (Difference from multi-model mean) Low High Colder Warmer
  70. 70. What climate model does the network predict for each year of observations? Multi-Model Mean
  71. 71. APPLYING METHODOLOGY TO REGIONS PREDICTION FOR EACH YEAR IN OBSERVATIONS LRP (EXPLAINABLE AI) COMPARED TO RAW DATA
  72. 72. Higher Confidence Lower Confidence RANKING CLIMATE MODEL PREDICTIONS FOR EACH YEAR IN OBSERVATIONS
  73. 73. I. Explainable neural networks reveal patterns of climate change in large ensembles simulated with different combinations of external forcing II. Neural networks can be used to identify unique model differences and biases between large ensembles and observations III. But what about predictability in the climate system?
  74. 74. I. Explainable neural networks reveal patterns of climate change in large ensembles simulated with different combinations of external forcing II. Neural networks can be used to identify unique model differences and biases between large ensembles and observations III. But what about predictability in the climate system?
  75. 75. I. Explainable neural networks reveal patterns of climate change in large ensembles simulated with different combinations of external forcing II. Neural networks can be used to identify unique model differences and biases between large ensembles and observations III. But what about predictability in the climate system?
  76. 76. Global Warming Hiatus? …in research
  77. 77. Global Warming Hiatus? …in research
  78. 78. Global Warming Hiatus? …in research
  79. 79. Global Warming Hiatus? …in research
  80. 80. Global Warming Hiatus? …in research
  81. 81. Global Warming Hiatus? …in research
  82. 82. Global Warming Hiatus? …in the media, etc.
  83. 83. Global Warming Hiatus? >300 papers, to-date
  84. 84. Hiatus period? Temperature Anomaly (°C)
  85. 85. Are slowdowns (“hiatus”) in decadal warming predictable? • Statistical construct? • Lack of surface temperature observations in the Arctic? • Phase transition of the Interdecadal Pacific Oscillation (IPO)? • Influence of volcanoes and other aerosol forcing? • Weaker solar forcing? • Lower equilibrium climate sensitivity (ECS)? • Other combinations of internal variability?
  86. 86. Select one ensemble member and calculate the annual mean global mean surface temperature (GMST) START OF 10-YEAR TEMPERATURE TREND 2-m TEMPERATURE ANOMALY
  87. 87. Calculate 10-year moving (linear) trends 2-m TEMPERATURE ANOMALY
  88. 88. Plot the slope of the linear trends START OF 10-YEAR TEMPERATURE TREND 2-m TEMPERATURE ANOMALY
  89. 89. Calculate a threshold for defining a slowdown in decadal warming
  90. 90. Repeat this exercise for each ensemble member in CESM2-LE
  91. 91. Compare warming slowdowns with reanalysis (ERA5)
  92. 92. So how well does the neural network do?
  93. 93. Low High Colder Warmer
  94. 94. Low High Colder Warmer
  95. 95. Low High Colder Warmer
  96. 96. What about observations? Future (2012-) so-called “hiatus”
  97. 97. What about observations? Future (2012-) so-called “hiatus” Comparing observations to the IPO
  98. 98. What about observations? Future (2012-) so-called “hiatus” 2021 (preliminary) Looking ahead to the near- future… ?
  99. 99. What about observations? Low High [2003, 2004] [2016, 2017]
  100. 100. What about observations? Colder Warmer [2003, 2004] [2016, 2017]
  101. 101. INPUT [DATA] PREDICTION Machine Learning Explainable AI Learn new science!
  102. 102. MACHINE LEARNING IS JUST ANOTHER TOOL TO ADD TO OUR WORKFLOW. 1)
  103. 103. MACHINE LEARNING IS NO LONGER A BLACK BOX – WE CAN ADDRESS PHYSICAL MECHANISMS. 2)
  104. 104. 3) WE CAN LEVERAGE NOVEL DATA SCIENCE METHODS WITH NEW CLIMATE MODEL LARGE ENSEMBLES.
  105. 105. ASSESSING CLIMATE VARIABILITY & CHANGE WITH EXPLAINABLE NEURAL NETWORKS CLIMATE/EVENT ATTRIBUTION GFDL’s SPEAR-MED NATURAL, Hist_SSP245, Hist_SSP585 runs Experiments using GFDL’s SPEAR for decadal prediction DECADAL PREDICTION DETECTING EXTREME EVENTS E.g., Very Rapid Ice Loss Events (VRILEs) in the Arctic S2S FORECASTS OF OPPORTUNITY How will climate change affect teleconnections (FOO)? FUTURE DIRECTIONS
  106. 106. KEY POINTS Zachary Labe zmlabe@rams.colostate.edu @ZLabe 1. Machine learning is just another possible tool to add to our scientific workflow 2. Machine learning is no longer a black box – we can address physical mechanisms in the climate system. 3. We can leverage novel data science methods with new climate model large ensembles to investigate S2S/S2D predictability and attribution of extreme events.

×