JointDiff: one multimodal diffusion model maps fusion experiments to physics both ways

bravo_abad · x · 2026-08-20

Michael Jones and coauthors introduce JointDiff, a multimodal diffusion model for inertial-confinement fusion that learns a joint distribution over simulation parameters, scalar diagnostics and neutron images — moving both ways between a fusion experiment and the physics behind it.

The ML idea is unusually transferable: instead of separate surrogates for forward prediction, inverse design and missing-data reconstruction, JointDiff learns all three simultaneously. Given simulation inputs it predicts observables; given partial observations it infers possible physical inputs; hide a diagnostic and it reconstructs a distribution for the missing measurement. That distributional output matters because inverse scientific problems are rarely one-to-one. Trained on more than 443,000 3D simulations.

Original post →

More from Research

Research channel →