Latent Generative Solvers for Generalizable Long-Term Physics Simulation
Zituo Chen, Haixu Wu, Sili Deng · 2026
Read Latent Generative Solvers for Generalizable Long-Term Physics Simulation by Zituo Chen, Haixu Wu, Sili Deng (2026) 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-2602-11229.txt.
Description
Neural PDE solvers usually trade generality for balance over long rollouts. Zituo Chen and Sili Deng close that gap with the Latent Generative Solver, a three-part system. A physics variational autoencoder compresses twelve PDE families into one shared latent space.
A pyramidal flow-forcing model then generates the next latent by flow matching, guided by context updated from its own forecasts. Training noise, backed by a contraction bound, explains long-horizon balance. Pretrained on millions of paths, the solver wins on nearly all systems at medium rollouts and cuts long-rollout error sharply, using far less recurrent compute. It also adapts fast to a held-out Kolmogorov flow.
Catalog record
| CREATOR | Zituo Chen, Haixu Wu, Sili Deng |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 48 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2602-11229