Neptuna: Large-Scale ML Benchmark for Compressible Multiphase Flows

Researchers have introduced Neptuna, the first large-scale benchmark for machine learning surrogates in shock-driven compressible multiphase flows. The benchmark comprises 2.4 TB of high-fidelity 2D and 3D datasets featuring shock-induced bubble collapse and droplet breakup. These flows involve strong nonlinear interactions, compressibility, sharp discontinuities, and multiphase effects, making surrogate modeling challenging. The dataset is hosted on Hugging Face under the repository FluidVerse, with sample videos, metadata, and inference rollout plots provided in supplementary material. The study evaluates diverse surrogate model families on the benchmark, aiming to advance reliable ML surrogates for applications such as bubble collapse and droplet breakup.