Machine Learning for Gradient Estimation in Earth Science Simulations

Abstract:

To understand and predict the behavior of large, complex, and chaotic systems, Earth scientists construct sophisticated simulators from physical laws. These simulators generalize better to new scenarios, require fewer tunable parameters, and are more interpretable than non-physical deep learning, but procedures for differentiation through the simulator are typically unavailable. These missing gradients limit the application of many important tools for forecasting, model tuning, sensitivity analysis, or sub-grid-scale parametrization. We address this limitation by applying deep learning to simulation data, translating the simulator into a differentiable emulator to provide the missing gradients. Emulator training does not require analyzing simulator code or equations, so modified simulators do not require new gradient calculation routines. Using the L96 model system, we demonstrate that emulator-derived gradients enable accurate optimization-based sensitivity analysis, 4D-Var data assimilation, and closed-loop training of parametrizations without access to the simulator's source code or internal states. Recent results show these successes can be scaled up to larger and more realistic models, providing a potential basis for combining the inductive biases of physical models with the power and flexibility of machine learning. ​

Bio: