Work

Hybrid World Models for Fusion Plasma Simulation

Year: 2026

Type: article

Abstract: We present FUSIONWM, a physics-informed World Model for fusion plasma based on the Joint Embedding Predictive Architecture (JEPA)It integrates 11 differentiable physics constraints into the training loss and predicts plasma state evolution across four tokamak configurations at 1.52 times the speed of a Rust PDE solver.A systematic ablation over seven predictor architectures reveals that architecture choice --- not parameter count --- determines multi-step rollout stability, with JEPA providing a consistent 42% dream-error reduction.We identify GRU, CfC, and MLP-Mixer as structurally compatible with photonic hardware, targeting 30-50 times energy-efficiency gains via Q.ANT's photonic processor.

Source: Zenodo (CERN European Organization for Nuclear Research)

DOI: 10.5281/zenodo.19914272

Authors: Neubürger, Felix, Kaufmann, Rafael, Kopinski, Thomas

Institution: South Westphalia University of Applied Sciences

Language: English

Topic: Model Reduction and Neural Networks

Subfield: Statistical and Nonlinear Physics

Field: Physics and Astronomy

Domain: Physical Sciences

Open Access status: green

OpenAlex ID: W7159342394

FWCI:

Cites:

Cited by:

Related to: