Building AI–Physics Digital Twins for Hurricanes With FourCastNet v2 and ERF

Mahesh Natarajan, Aaron M. Lattanzi, Ann Almgren, Jean Sexton, Weiqun Zhang, Travis Allen O'Brien, William D. Collins

Zenodo (CERN European Organization for Nuclear Research) · 12 October 2026 · Lawrence Berkeley National Laboratory / NERSC, University of California, Berkeley

Paper

In one sentence. Couples FourCastNet v2 boundary forcing with an 8 km ERF solver to simulate Hurricane Laura, reproducing rapid intensification to about 120 knots where AI-only models underestimate it.

Abstract

Abstract While purely physical numerical weather models are computationally expensive, pure artificial intelligence (AI) models often struggle to respect physical conservation laws and accurately capture localized extreme events, such as rapid hurricane intensification. To address this challenge, we present a research software workflow that enables a hybrid AI–physics digital twin architecture, combining the execution speed of deep learning with the physical rigor of high-resolution numerical solvers. Our workflow utilizes global weather reanalysis data (GDAS) as initial conditions to drive the ECMWF-FourCastNet v2 machine learning framework. FourCastNet v2 rapidly generates multi-day forecast trajectories, providing time-dependent boundary conditions to force the Energy Research and Forecasting (ERF) model–an open-source atmospheric solver built on the exascale AMReX framework. Using an 8 km horizontal resolution mesh with 42 million cells, ERF computes internal convective fluid dynamics, employing Smagorinsky turbulence, Kessler microphysics, and bulk surface parameterizations. We evaluate this framework on a 3-day tracking case study of Hurricane Laura (August 2020) executed across 16 NVIDIA A100 GPUs on NERSC’s Perlmutter supercomputer, completing in 1.5 hours. While both AI-only and hybrid approaches accurately constrain storm trajectory within the NOAA track forecast cone, AI-only predictions severely underestimate storm intensification. Conversely, the hybrid AI–physics framework reasonably reproduces the rapid intensification phase, reaching maximum wind speeds of ~120 knots, broadly consistent with IBTrACS observations. This work highlights the power of coupling AI boundary forcing with exascale physics solvers for high-fidelity extreme weather forecasting. Keywords Digital Twins, AI–Physics, Hurricanes, GDAS, IBTrACS, FourCastNet v2, ERF, AMReX

Labels

Labels and summary are generated by Claude and reviewed weekly. Wrong? Edit this entry.