AP3ESM: A Kilometer-Scale AI-Powered and Performance-Portable Earth System Model

Juepeng Zheng of Sun Yat-Sen University and the National Supercomputing Center in Shenzhen presents AP3-ESM version 2, a hybrid AI-physics coupled earth system model running at global one-kilometer resolution on China's exascale supercomputers, delivered on behalf of the developing group. Building on a decade of kilometer-scale modeling work including CESM on Sunway, the GRIST atmospheric model, and the LICOM ocean model, version 2 advances three directions: hybrid AI-physics fidelity, computational efficiency and portability, and coupled multiscale simulation capability. The atmosphere integrates a physics-constrained fully AI parameterization package that replaces the whole conventional physics suite to preserve consistency, with predicted tendencies constrained within a convex hull of learned basis tendencies to avoid extrapolation; the ocean adds an AI-based near-inertial wave parameterization using a residual MLP trained on 27 million samples from high-resolution simulations. Optimizations include OpenMP and Kokkos acceleration, an offline NetCDF/HDF5 partitioning tool that caches domain decomposition for reuse, and FP16 mixed-precision storage selected automatically while keeping arithmetic in FP32. On the Lanshan ARM-based system, the one-kilometer atmosphere reaches about 1.7 simulated years per day and the ocean about 3.1, with the coupled system projected around one SYPD, roughly three times faster than version 1; results are also shown on a GPGPU system and the new Sunway. Simulations capture typhoon structure, Himalayan topographic winds, and post-typhoon ocean mixed-layer deepening.