Research map / Probabilistic and scientific modeling / Differential equations and PDEs
Neptuna: A Comprehensive Machine Learning Framework for Benchmarking Complex Multiphase Flows
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Introduces Neptuna, a 2.4 TB benchmark of shock-driven compressible multiphase flows (bubble collapse and droplet breakup in 2D/3D) and benchmarks CNO, FFNO, ConvNeXt, ScOT, Poseidon and DPOT under MSE vs composite losses with SoftAdapt/GradNorm weighting.
Problem
Machine learning surrogates for compressible multiphase flows with shocks and material interfaces lack large benchmark datasets and evaluation protocols that jointly cover compressibility, discontinuities, and multiphase effects.
Approach
A multiphase shock-driven dataset suite generated with the ALPACA/RDEMIC solver, plus a benchmarking framework that pairs composite losses (MSE with H1, SSIM, IRMSE) with adaptive loss balancing (SoftAdapt, GradNorm) and a broad set of 12 evaluation metrics including physics-informed ones.
Key results
- No single surrogate is best across all datasets and metrics; fine-tuned DPOT-M performs best among 2D baselines.
- Composite losses improve IRMSE by roughly 40-50% on the shock-droplet dataset versus about 5-7% on bubble-collapse datasets.
- SoftAdapt gives the most consistent gains over MSE-only training with almost no overhead, more so than GradNorm.
Surprising findings
- Fine-tuned Poseidon-B performs substantially worse than other baselines despite having the largest parameter count among 2D models.
- Incorporating a metric into the composite training loss does not guarantee improved inference performance on that metric; SSIM gains were empirical.
Limitations
- No evaluated surrogate accurately recovers small secondary droplets in the SIE regime.
- 3D baselines trained under a 24-hour limit, yielding far fewer epochs, so 3D conclusions are tentative.
- Composite loss terms were selected after broad experimentation, risking selection bias on evaluation metrics.
Experimental setup
Datasets: 2D-SABW 3D-SABW 2D-SRBA 2D-SDBA 3D-SDBA
Benchmarks: Neptuna (2D/3D shock-driven multiphase datasets)
Metrics: VRMSE MAE MLW wfRMSE IRMSE SSIM PCC iqRMSE-enstrophy H2-seminorm shock-focused RMSE
Baselines: ConvNeXt CNO FFNO ScOT Poseidon-B DPOT-M
Builds on: PDEBench APEBench ALPACA RDEMIC
Closest papers by meaning
- Bubbleformer: Forecasting Boiling with Transformers (NeurIPS 2025 · similarity 0.85)
- PINNacle: A Comprehensive Benchmark of Physics-Informed Neural Networks for Solving PDEs (NeurIPS 2024 · similarity 0.85)
- MultiPDENet: PDE-embedded Learning with Multi-time-stepping for Accelerated Flow Simulation (ICML 2025 · similarity 0.84)
- PARCv2: Physics-aware Recurrent Convolutional Neural Networks for Spatiotemporal Dynamics Modeling (ICML 2024 · similarity 0.84)
- Poseidon: Efficient Foundation Models for PDEs (NeurIPS 2024 · similarity 0.84)
- Geometric and Physical Constraints Synergistically Enhance Neural PDE Surrogates (ICML 2025 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.