Generative surrogate modeling of hydrodynamic instabilities using stochastic interpolants
Please login to view abstract download link
The advent of scientific deep learning has brought powerful and versatile new tools for building surrogate models of unprecedented quality. In particular, generative models based on diffusion [Song et al., 2020] or stochastic interpolants [Albergo, Boffi and Vanden-Eijnden, 2023] provide a way to generate samples from complex and very high dimensional distributions. While learning these representations generally requires a lot of data, downstream applications may require even more and prove completely unfeasible without these intermediate surrogates. For example, exploring a large parameter space for optimization, sensitivity analysis or sampling problems typically requires thousands to millions of evaluations, which are out of reach of direct simulations. These enhanced capabilities pave the way for applications to systems of practical interest such as inertial confinement fusion (ICF). Although numerous experimental configurations have been explored over the years, leading to significant and repeated energy gains since 2022, exploration of the design space remains limited, and a large fraction of experiments still fails to reach high gains. There are indeed many sources of perturbations that can lead to inhomogeneous heating and compression of the target. One of them is the development of hydrodynamic instabilities at the target's interfaces [Zhou, Sadler and Hurricane, 2025], such as Rayleigh-Taylor's and Richtmyer-Meshkov's, which may evolve to turbulent mixing zones that increase the energy losses. The use of advanced surrogates, possibly in conjunction with artificial intelligence agents [Shachar et al., 2025], has the potential to reveal interesting new designs that could increase yield or reduce production costs and complexity. This work focuses on the development of high-fidelity surrogate models for hydrodynamic instabilities using the stochastic interpolant framework, which encompasses diffusion models. This framework can be adapted to perform various tasks such as full-field emulation, stochastic prediction, reconstruction of partially observed fields and statistical inference. Ongoing and future directions include using these models to build surrogate optimization loops and applications to configurations more representative of ICF.
