CLIMASIM — Climate Simulation with Scientific Machine Learning

Tirtha Tilak Pani, Prathamesh Dinesh Joshi, Raj Dandekar, Sreedath Panat, Rajat Dandekar

Track: Regular Track (Page limit: 6-8 pages)

Keywords: Scientific Machine Learning, Climate Modeling, Universal Differential Equations, Neural Ordinary Differential Equations, Energy Balance Models, Symbolic Regression, Carbon Cycle Dynamics, Climate-Carbon Feedback, Physics-Informed Machine Learning, Sparse Identification of Nonlinear Dynamics (SINDy), General Circulation Models (GCMs), Earth System Models (ESMs), Climate Forecasting, Mechanistic Interpretability, LASSO Regression, CO₂ Concentration Modeling, Temperature Anomaly Prediction, Biogeochemical Cycles, Computational Climate Science

TL;DR: Physics-guided neural networks enable accurate, fast, transparent climate modeling for rapid policy scenario testing.

Abstract: Climate change manifests as a complex, nonlinear dynamical system characterized by intricate interactions among atmospheric CO2 concentrations, surface and ocean temperatures, and anthropogenic forcing. While General Circulation Models (GCMs) and their successors, Earth System Models (ESMs), provide comprehensive simulations by incorporating detailed biogeochemical cycles, their computational demands remain prohibitive for rapid climate scenario exploration. Conversely, classical Energy Balance Models (EBMs) offer computational tractability at the expense of predictive accuracy. We present a novel framework that augments EBMs through scientific machine learning, enhancing accuracy while preserving physical interpretability. Specifically, we couple classical energy balance formulations with carbon cycle dynamics and evaluate performance under linearly increasing emission scenarios. Our methodology proceeds systematically: we first investigate Neural Ordinary Differential Equations (Neural ODEs) for climate forecasting, finding limited efficacy. Subsequently, we replace a critical carbon-climate feedback term with a neural network, constructing a Universal Differential Equation (UDE) that achieves error rates below 0.2\% across all climate variables for three distinct initializations. To ensure mechanistic transparency, we employ sparse symbolic regression via LASSO, successfully recovering learned dynamics across three initializations and six noise perturbation levels. Comparative benchmarking against statistical baselines (VAR and ARIMA) demonstrates superior forecasting performance in data-scarce regimes with known physical constraints. Our results establish that UDEs enable accurate climate state prediction while symbolic regression maintains interpretability, yielding a computationally efficient framework for rapid climate scenario exploration and mechanistically transparent climate modeling.

Submission Number: 39