Exploring the Frontiers in Earth System Modeling with Machine Learning and Big Data

This workshop brings a small but varied group of geoscientists and climate modelers together with machine learners, statisticians, and representatives of other fields where ML has already had a big impact. Discussions will center on how innovative and efficient ML methods will provide new, innovative and transformative ways of modeling and projecting the Earth system and extracting information from massive data volumes. Of particular importance will be an assessment of the challenges ahead in combining physical knowledge with data-driven methods. This is a challenging and promising research field where a concentrated effort will have a high impact both to advance science and to address topics of critical importance and high relevance for human society. Newly developed methods can open opportunities for innovative new hybrid ML-Earth system models, as well as analyses of models and observations. With enhanced analysis capabilities, we can look forward to revolutionary advancement of Earth system sciences to accelerate scientific understanding, modeling, and projecting climate change.

Detailed Description

Over the last decade, the volume of data from satellite sensors and Earth system models has increased by at least an order of magnitude. The exponential growth in data is threatening to overwhelm the relatively simple tools and methods currently used to analyze the data within the climate research community, and is providing new opportunities to improve Earth system and climate models with machine learning (ML).

ML is a promising and growing path of research that could contribute to substantially enhancing our understanding of the Earth system and reduce uncertainty in climate change projections. While ML is revolutionizing many fields, from molecular science, bioinformatics, and genomics to material sciences, in Earth sciences this revolution is still in its infancy with the potential of ML methods and ideas yet to be fully explored. One reason is that a large body of physical knowledge enabled to develop theories and Earth system models is derived from first principles. This purely physics-driven approach has its limits and there are many ways in which ML can be harnessed to improve our understanding, models, and to predict and project the Earth system from weather to climate scales.

ML excels at modeling highly complex functional relationships. One novel path forward involves ML modules for simulating small-scale processes in Earth system models that cannot be fully resolved and that are currently often heuristically parametrized. Another opportunity is for ML to deliver powerful tools for analyzing high-dimensional datasets which are especially prevalent in the Earth sciences. Additionally, ML, augmented by causal inference methods, allows more information to be extracted from observations and models on how processes interact causally.

This workshop brings a small but varied group of geoscientists and climate modelers together with machine learners, statisticians, and representatives of other fields where ML has already had a big impact. Discussions will center on how innovative and efficient ML methods will provide new, innovative and transformative ways of modeling and projecting the Earth system and extracting information from massive data volumes.

Of particular importance will be an assessment of the challenges ahead in combining physical knowledge with data-driven methods. This is a challenging and promising research field where a concentrated effort will have a high impact both to advance science and to address topics of critical importance and high relevance for human society. Newly developed methods can open opportunities for innovative new hybrid ML-Earth system models, as well as analyses of models and observations. With enhanced analysis capabilities, we can look forward to revolutionary advancement of Earth system sciences to accelerate scientific understanding, modeling, and projecting climate change.

Main Topics of the Workshop

(I) Hybrid modeling and subgrid-scale parametrizations

• ML tasks: Learning from observational data and high-resolution model output, Parameter estimation
• Approaches: Deep Learning (generative models, classification models), Unsupervised Learning, Transfer learning
• Unique Challenges: Incorporating physical constraints, Developing physical regularization techniques, Interpretability, Out-of-sample generalization, Uncertainty propagation

(II) Extreme event detection and forecasting

• ML tasks: Anomaly detection, Pattern Classification, Change detection, Dimension reduction, Uncertainty quantification
• Approaches: Deep Learning (spatial; space-time; sequence-to-sequence models); generative models
• Unique Challenges: Handling large-scale multivariate, spatio-temporal data; Developing parallel algorithms for distributed computing; Physical interpretability; Explainability; Robustness, generalization guarantees

(III) Extracting dynamical causal dependencies

• Analytics Tasks: Combining analyses of climate model experiments with observations; Variable extraction and dimension reduction from spatio-temporal data
• Approaches: Conditional independence-based algorithms; Structural causal modeling and invariant prediction techniques; Causal network comparison metrics
• Unique challenges: High-dimensional nonlinear interdependencies; Uncertainty propagation

(IV) Constraining climate change projections

• Analytics tasks: Dimension reduction; Uncertainty quantification
• Approaches: Causal transfer learning and transportability; Bayesian optimal weighting schemes; Data mining
• Unique challenges: Physical interpretation / plausibility

(V) Cross-cutting challenges for ML method development

• Physical interpretation/plausibility
• Explainability
• Incorporating physical constraints, physics-based regularization techniques
• Robustness, generalization guarantees