Characteristics of Climate Data

• Observational data obeying physical laws

is limited

• Climate models, e.g., NMME, SubX, etc.

• Vast amounts “imperfect” outputs

SSF Benchmark Dataset (1980 - recent)

• Prediction task - mean temperature in week 3-4

• Residualized w.r.t. climatology, 30-year average

on each location and each date

• Climate variables - atmosphere, land, and ocean

Exploration of ML Algorithms

• Investigate ML approaches to SSF forecasting over the contiguous U.S.

• ML models are doing better than climatology

• Coastal areas are overall easier to predict than inland

Interpretable Data-driven Climate Indices

• Transformation of climate variables via clustering

  • Spatially coherent

  • Predictive of key climate variables

• Improve predictive performance over raw data

Climate (Dynamical) Models

• Dynamics based on PDEs

(partial differential equations)

• Billions of variables governed by PDEs

• Several hundreds of parameters characterize

the PDEs

Improve Dynamical Models (PDEs)

• Develop ML tools to estimate PDE parameters

• Approximate Bayesian Inference

• Derivative-Free Optimization (DFO)

• Monte Carlo Tree Search (MCTS)

• Tune real-world climate model with billions of variables

Approximate Bayesian Inference

Gaussian prior based on the “observational” data

Initial Results CESM2

• Estimating clubbgamma and dcs

• 6 ensemble members

Observations

• Model parameters can be estimated fairly accurately after 6 months

Derivative-Free Optimization (DFO)

Initial Results

• 2-variable Lorenz-96 Model

Observations

• Integration over a long-time horizon

  • Chaotic, hard to optimize

• Integration over a short horizon

  • Akin to convex function, easy to optimiz