Lawrence Livermore National Laboratory: Yunyan Zhang (PI), Hassan Beydoun, Peter Bogenschutz, Emily de Jong, Lin Lin, Hsi-Yen Ma, Tianning Su, Xue Zheng

Argonne National Laboratory: Jiwen Fan, Yunpeng Shan

University of Oklahoma: Xuguang Wang, Yongming Wang

Pacific Northwest National Laboratory: Andrew Gettelman

NVIDIA

Over the past five years, a new generation of Earth system models has advanced rapidly including the development and application of global storm-resolving models (GSRMs), such as DOE’s Simple Cloud Resolving E3SM Atmosphere Model (SCREAM), for global simulations of deep convective processes and their multiscale interactions. However, GSRMs still face significant predictability challenges due to uncertainties in representing small-scale cloud microphysics and atmospheric turbulence. Biases in clouds and precipitation can lead to errors in surface water and energy budgets.

Artificial intelligence (AI) and machine learning (ML) offer strong potential to improve these processes and reduce model biases and uncertainties by:

• Learning from high-fidelity models, such as large eddy simulations (LES), to emulate complex fine-scale physical evolution
• Emulating model physics to accelerate ensemble prediction and model calibration
• Integrating observational data with model physics to support physics-informed model development and data assimilation

Through STITCH Phase I, we propose to:

• Build a testbed to implement and evaluate process emulators for turbulence and warm-rain microphysics against benchmark LES data and observations by DOE’s Atmospheric Radiation Measurement (ARM) facility
• Create an emulator of regionally refined SCREAM (RRM-SCREAM) for simulating clouds and precipitation around the ARM sites
• Develop a workflow to assess the impact of initial conditions on RRM-SCREAM simulations through data assimilation of ARM observations

In Phase I, we will generate LES training data to develop process emulators using a range of neural operator approaches. We will also generate RRM-SCREAM training data to build a cloud and precipitation emulator for regions around ARM sites. In parallel, we will generate initial conditions by assimilating ARM observations, including scanning radar reflectivity and sounding array measurements and evaluate the impact of those initial conditions on RRM-SCREAM simulations.

If successful, these Phase I efforts will establish the foundation for creating an ARM data cube by integrating RRM-SCREAM and ARM observations. This capability will support the development of an ARM digital twin for discovering new physics and further improving the representation of cloud microphysics, turbulence, and their emulators. Together, these capabilities will form a positive feedback loop that incorporates new observations, model advances, and machine learning methods to continually improve the digital twin and model predictability of water and energy at regional and global scales.