Accelerating 3D Solar atmosphere modelling with Graph Neural Networks

Simulating the sun's atmosphere requires solving complex non-local, non-linear systems of coupled equations. A massive computational bottleneck that typically restricts researchers to 1.5D approximations or requires millions of CPU hours for full 3D snapshots. We present a solution that bypasses these limitations using Graph Neural Networks (GNNs) to predict atomic population states directly from 3D physical parameters. We will demonstrate how we modeled the simulation domain as a learnable graph, using message passing to replicate the non-local physics of photon scattering. This approach yielded a robust surrogate model that is 10^4 times faster than standard numerical solvers.