Benchmarking neural surrogates on realistic spatiotemporal multiphysics flows
Runze Mao, Rui Zhang, Xuan Bai, Tianhao Wu, Teng Zhang, Zhenyi Chen, Minqi Lin, Bocheng Zeng, Yangchen Xu, Yingxuan Xiang, Haoze Zhang, Shubham Goswami, Pierre A. Dawe, Yifan Xu, Zhenhua An, Mengtao Yan, Xiaoyi Lu, Yi Wang, Rongbo Bai, Haobu Gao, Xiaohang Fang, Han Li, Hao Sun, Zhi X. Chen · 2025
Read Benchmarking neural surrogates on realistic spatiotemporal multiphysics flows by Runze Mao, Rui Zhang, Xuan Bai, Tianhao Wu, Teng Zhang, Zhenyi Chen, Minqi Lin, Bocheng Zeng, Yangchen Xu, Yingxuan Xiang, Haoze Zhang, Shubham Goswami, Pierre A. Dawe, Yifan Xu, Zhenhua An, Mengtao Yan, Xiaoyi Lu, Yi Wang, Rongbo Bai, Haobu Gao, Xiaohang Fang, Han Li, Hao Sun, Zhi X. Chen (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2512-18595.txt.
Introduction
Runze Mao and a large team argue that neural surrogates for fluid physics look better than they are, because tests use simplified toy problems. Their 2025 benchmark, REALM, grades such models on realistic coupled flows instead. It bundles 11 high fidelity datasets, ranging from canonical problems to propulsion and fire safety scenarios. A standard protocol covers training and evaluation, with physics aware preprocessing and a robust rollout strategy.
The team benchmarks more than a dozen model families, including spectral operators, convolutional models, Transformers, pointwise operators, and graph networks. Three robust trends emerge. A scaling barrier, set by dimensionality, stiffness, and mesh irregularity, makes rollout errors grow fast. Performance is governed mainly by built in architectural biases, not by size. Most sobering, high scores can hide unphysical behavior, as models miss key transient structures and totals. REALM gives the field a strict testbed for building models engineers could actually trust.
Catalog record
| CREATOR | Runze Mao, Rui Zhang, Xuan Bai, Tianhao Wu, Teng Zhang, Zhenyi Chen, Minqi Lin, Bocheng Zeng, Yangchen Xu, Yingxuan Xiang, Haoze Zhang, Shubham Goswami, Pierre A. Dawe, Yifan Xu, Zhenhua An, Mengtao Yan, Xiaoyi Lu, Yi Wang, Rongbo Bai, Haobu Gao, Xiaohang Fang, Han Li, Hao Sun, Zhi X. Chen |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 129 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2512-18595