Physics Informed ML | Machine Learning & Atmospheric Processes Group

Image from NVIDIA showing their PhysicsNeMo Platform for modeling physical systems and deploying digital twins.

Physics informed machine learning blends traditional physical modeling with data driven learning. Rather than treating machine learning as a black box this approach introduces structure into the model itself or into the loss function based on known scientific principles such as conservation laws governing mass energy and momentum. In the geosciences where measurements are sparse and dynamics are complex this combination is essential for building trustworthy systems that generalize beyond the training data.

We are exploring physics informed ML methods across a range of applications from satellite retrievals to land surface modeling. For example many current precipitation and snow retrievals rely on either purely empirical models or physics only inversions that can break down under new climate regimes. By combining both approaches, we aim to build models that are more robust to noise while still obeying the constraints of the physical system.

Understanding the structure of a system in terms of learned physics can help enhance trust and transparency in the model output

One of our current focuses is on modeling northern hemisphere snow water equivalent. We are designing graph neural networks (GNNs) that represent terrain and atmospheric conditions as a network of interacting locations with physical constraints such as energy balance and mass conservation encoded directly into the model. These models learn from remote sensing data but also respect the governing physics of snow accumulation melt and transport.

By training these models to reflect the true water and energy budgets we can better capture snow dynamics under a range of conditions including melt events sublimation and refreezing. We also plan to integrate numerical snow model outputs into a hybrid learning setup where ML models can learn to correct known biases or missing processes in traditional approaches.

We are excited to continue expanding these methods to new areas of Earth system science. Physics informed learning offers a powerful way to reduce unrealistic model behavior improve scientific credibility and push machine learning into more operational and reliable domains. Whether for retrievals land surface fluxes or coupled modeling systems our goal is to improve the physical consistency of machine learning models across the geosciences.

Related Publications

2022

JAMES

Toward Efficient Calibration of Higher-Resolution Earth System Models

Christopher G. Fletcher, William McNally, John G. Virgin, and 1 more author

Journal of Advances in Modeling Earth Systems, 2022

Abs DOI

Projections of future climate change to support decision-making require Earth system models (ESMs) running at high spatial resolution, but at present this is computationally prohibitive. A major challenge is the calibration (parameter tuning) during the development of ESMs, which requires running large numbers of simulations to identify optimal values for parameters that are poorly constrained by observations. Here, we train a convolutional neural network (CNN) to emulate perturbed parameter ensembles from two lower-resolution (and thus much less expensive) versions of the same ESM, and a smaller number of higher-resolution simulations. Cross-validated results show that the CNN’s skill exceeds that of a climatological baseline for most variables with as few as 5–10 examples of the higher-resolution ESM, and for all variables (including precipitation) with at least 20 examples. This proof-of-concept study demonstrates a machine learning based approach that makes the process of constructing a higher-resolution emulator 20%–40% more computationally efficient, and thus offers the prospect of significantly more efficient calibration of ESMs.