AI-for-Science in Earth Science by Integrations of Simulations/Data/Learning on Heterogeneous Supercomputers
Please login to view abstract download link
Since 2015, we have advanced the BDEC initiative to integrate Simulation, Data, and Learning (S+D+L). To fully leverage heterogeneous supercomputers such as Wisteria/BDEC‑0 and Miyabi, both equipped with CPUs and GPUs, we have developed the advanced software infrastructure h3‑Open‑BDEC. This platform enables AI-for-Science workflows and has produced internationally recognized achievements across fields including earth science. In this presentation, we will present the following three topics. (1) Integration of 3D Seismic Wave Propagation and Real-Time Data Observation/Assimilation: Seism3D-DAF assimilates 3D seismic wave propagation simulations with earthquake observation data provided by JDXnet (Japan Data eXchange network) in real time. We present a preliminary forecasting system developed on Wisteria/BDEC-01 using h3-Open-BDEC. Our current activities for optimization of the codes on GPU cluster (Miyabi-G) will be also introduced in the presentation. (2) Simulation-AI Coupling for Global Atmospheric Model: We coupled the CPU based NICAM atmospheric model with a GPU based machine learning surrogate in real time, conducting experiments on the Wisteria/BDEC-01 and Miyabi supercomputers using h3-Open-BDEC. The AI model was trained to reproduce microphysics tendencies from eight input variables at each vertical layer, accurately capturing NICAM’s spatial structures. Computational performance was also evaluated on both platforms. (3) Ensemble Coupling for Weather/Climate Simulation: We proposed an innovative computational method ensemble coupling, that combines low-resolution ensemble calculations with a high-resolution single one in global weather/climate simulations. We will demonstrate examples in which applying ensemble coupling achieved more than 100x improvement in computational efficiency compared with conventional ensemble-based data assimilation methods.
