We develop and use AI-assisted simulation tools to model magnetized plasmas, pervasive in the universe from galaxies to stars and planets, and core to engineering applications like fusion reactors. Present-day tools employ direct numerical simulation (DNS), aiming to capture all relevant scales but facing prohibitive computational costs and vast data output. To tackle this, we apply artificial intelligence to DNS data to learn accurate, generalizable sub-grid scale (SGS) models of small-scale processes, reducing computational costs and data storage needs. Our tools are scalable to Exascale computing, enabling faster, more sustainable progress in science and engineering. A proof-of-concept has confirmed their feasibility. This project aims to apply our framework to study solar convection, that drives Sun's magnetic field generation, unresolved over 70 years. Our simulations, using LUMI-G supercomputing, hope to shed new light on this problem by using unprecedented resolution.

2026

2028