Laser- and shock-induced droplet dynamics: A machine learning benchmark for complex multiphase flows

ICLR 2026 Conference Submission20815 Authors

Keywords: Scientific Machine Learning, Compressible multiphase flow, Data-driven surrogates, Droplet dynamics, Autoregressive rollout, Neural operators, Transformers

TL;DR: We introduce two complex compressible multiphase datasets and train baseline models autoregressively, which provide insight into the underlying challenging physics.

Abstract: Compressible multiphase flow is central to numerous engineering applications, characterized by complex wave dynamics and challenging shock-interface interactions. Despite their importance, they remain significantly missing from existing benchmarks in the Scientific Machine Learning (SciML) community, limiting progress on generalization to impactful real-world scenarios. To address this issue, we introduce two exemplary datasets from this class, Laser-Induced Droplet Explosion (LIDE) and Shock-Induced Droplet Aero-breakup (SIDA), providing researchers with valuable references to establish reliable baselines and push boundaries of SciML. Due to the high computational cost of simulating these processes with full fidelity, we explore data-driven surrogate models designed to efficiently approximate the underlying physics at reduced cost. We benchmark these datasets on diverse architectures--UNet, Fourier Neural Operator (FNO), Vision Transformer (ViT), Scalable Operator Transformer (ScOT), and Residual Network (ResNet)--trained autoregressively and compared across varying parameter counts. A comprehensive set of ablations is carried out to analyze the performance of the models. We identify key scenarios, such as incorporating temporal sequence information and conditioning, that enable the models to accurately capture the rich and nonlinear physics embedded in the datasets. Code and datasets will be made available upon request.

Primary Area: datasets and benchmarks

Submission Number: 20815