Neptuna: A Comprehensive Machine Learning Framework for Benchmarking Complex Multiphase Flows

Researchers introduced Neptuna, a new benchmark for shock-driven compressible multiphase flows.

The benchmark includes 2.4 TB of 2D and 3D data on bubble collapse and droplet breakup.

It tests multiple surrogate model families, loss functions, and adaptive weighting methods such as SoftAdapt and GradNorm.

Neptuna is the first large-scale benchmark for this difficult fluid regime, enabling more rigorous model comparison and better simulation surrogates.