Joint Diffusion as a Multi-Modal Surrogate for Inertial Confinement Fusion Simulations

Inertial Confinement Fusion (ICF) research increasingly relies on machine

learning methodologies to bridge the gap between high-fidelity

hydrodynamics simulations and experimental shot data. Capsule

simulations generate complex, multi-modal datasets—including images

and scalar observables—that predict corresponding experimental

diagnostics. Due to the computational cost of physics-based simulations,

machine learning-based surrogate models have been developed to rapidly

predict simulation outputs, accelerating the workflow by several orders of

magnitude.

In this work, we present a generative framework for solving the ICF inverse

problem: predicting distributions of capsule input parameters that best

reproduce experimental observations and associated uncertainties. While

Markov Chain Monte Carlo (MCMC) methods have been the standard

approach, we demonstrate that generative diffusion models offer

improvements in both convergence and computational efficiency. Our

diffusion-based approach is inherently multi-modal and invertible, enabling

joint inference and prediction of both scalar and image-based

observables. This framework is adaptable across a range of ICF platforms

and experimental configurations.

Furthermore, we discuss strategies for enhancing model performance

through fine-tuning on limited experimental datasets, highlighting the

potential for data science to accelerate discovery and optimize

experimental design in ICF research.