Public Abstract

DE-SCL0000257: Physics-Informed Machine Learning of Cloud Microphysics for High-Resolution Earth System Modeling with Observationally Constrained Online Training

Award Status: Active

Institution: Argonne National Laboratory (ANL), Lemont, IL
UEI: JA8EP8EKX8Y9
PM: Nasiri, Shaima

Most Recent Award Date: 08/01/2026
Number of Support Periods: 1
PI: Feng, Yan

Current Performance Period: 08/01/2026 - 04/30/2027
Current Project Period: 08/01/2026 - 04/30/2027

Public Abstract

Physics-Informed Machine Learning of Cloud Microphysics for High-Resolution Earth System Modeling with Observationally Constrained Online Training Argonne National Laboratory: Yan Feng (Principal Investigator), Jiwen Fan, Rao Kotamarthi, Virendra Ghate, Hong Zhang, Sri H. Krishna Narayanan, Sid Gupta, and Yunpeng Shan NCAR: Sisi Chen and Lulin Xue Aeolus Labs Inc.: Gregory Wagner, Kai-yuan Cheng, and Paul Berg Accurate prediction of precipitation patterns and extremes is essential for energy sector resilience, yet Earth System Models (ESMs) remain limited by persistent biases arising from the parameterization bottleneck in cloud microphysics and turbulence. Microphysical processes, spanning from submicron to centimeter scales, are strongly modulated by small-scale turbulence that existing bulk schemes in ESMs cannot resolve. While high-fidelity simulations such as Direct Numerical Simulations (DNS) offer greater physical realism, their computational cost precludes direct use in ESMs. Existing Machine Learning (ML) approaches face a structural impasse: "offline" models suffer from numerical instability, while "online" models trained on coarse reanalysis data fail to capture underlying microphysical processes. To address these challenges, we propose PRECISION (Physics-Informed Machine Learning of Cloud Microphysics for High-Resolution Earth System Modeling with Observationally Constrained Online Training), a hybrid ML framework designed to bridge process-level realism and global-scale tractability. The framework comprises three stages: (1) DNS-based pre-training, where deep neural networks (NNs) learn residual corrections to bulk schemes and reduced-order representations of microphysics; (2) Neural Large Eddy Simulation (LES) development, integrating NN-based microphysics into a GPU-first, end-to-end differentiable atmospheric model; and (3) Observation-constrained online learning, using DOE ARM data to refine predictions across cloud regimes and reduce systematic biases. Full differentiability enables gradient-based nudging toward real-world observations, creating a "Neural LES” that learns unresolved microphysics–turbulence interactions directly from observations. In Phase I, we will demonstrate a Neural LES framework; in Phase II, we will enable online training and scale the framework to a neural mesoscale model suitable for kilometer-scale ESM simulations. Key AI innovations include process-resolving learning from DNS, adaptive and stable online training that mitigates distribution shift, and cross-scale generalization within a differentiable framework that transfers consistently from DNS/LES to mesoscale and ESMs. This project represents a paradigm shift, leveraging AI to unify a resolution-spanning hierarchy from DNS (millimeter-scale) to LES (meter-scale) and mesoscale systems, ultimately delivering reliable predictions of the water cycle and extreme weather for energy sector on the sub-seasonal to seasonal to decadal scales.

Physics-Informed Machine Learning of Cloud Microphysics for High-Resolution Earth System Modeling with Observationally Constrained Online Training

Argonne National Laboratory: Yan Feng (Principal Investigator), Jiwen Fan, Rao Kotamarthi, Virendra Ghate, Hong Zhang, Sri H. Krishna Narayanan, Sid Gupta, and Yunpeng Shan
NCAR: Sisi Chen and Lulin Xue
Aeolus Labs Inc.: Gregory Wagner, Kai-yuan Cheng, and Paul Berg

Accurate prediction of precipitation patterns and extremes is essential for energy sector resilience, yet Earth System Models (ESMs) remain limited by persistent biases arising from the parameterization bottleneck in cloud microphysics and turbulence. Microphysical processes, spanning from submicron to centimeter scales, are strongly modulated by small-scale turbulence that existing bulk schemes in ESMs cannot resolve. While high-fidelity simulations such as Direct Numerical Simulations (DNS) offer greater physical realism, their computational cost precludes direct use in ESMs. Existing Machine Learning (ML) approaches face a structural impasse: "offline" models suffer from numerical instability, while "online" models trained on coarse reanalysis data fail to capture underlying microphysical processes.

To address these challenges, we propose PRECISION (Physics-Informed Machine Learning of Cloud Microphysics for High-Resolution Earth System Modeling with Observationally Constrained Online Training), a hybrid ML framework designed to bridge process-level realism and global-scale tractability. The framework comprises three stages: (1) DNS-based pre-training, where deep neural networks (NNs) learn residual corrections to bulk schemes and reduced-order representations of microphysics; (2) Neural Large Eddy Simulation (LES) development, integrating NN-based microphysics into a GPU-first, end-to-end differentiable atmospheric model; and (3) Observation-constrained online learning, using DOE ARM data to refine predictions across cloud regimes and reduce systematic biases. Full differentiability enables gradient-based nudging toward real-world observations, creating a "Neural LES” that learns unresolved microphysics–turbulence interactions directly from observations.

In Phase I, we will demonstrate a Neural LES framework; in Phase II, we will enable online training and scale the framework to a neural mesoscale model suitable for kilometer-scale ESM simulations. Key AI innovations include process-resolving learning from DNS, adaptive and stable online training that mitigates distribution shift, and cross-scale generalization within a differentiable framework that transfers consistently from DNS/LES to mesoscale and ESMs. This project represents a paradigm shift, leveraging AI to unify a resolution-spanning hierarchy from DNS (millimeter-scale) to LES (meter-scale) and mesoscale systems, ultimately delivering reliable predictions of the water cycle and extreme weather for energy sector on the sub-seasonal to seasonal to decadal scales.