![[2602.20399] GeoPT: Scaling Physics Simulation via Lifted Geometric Pre-Training](https://arxiv.org/static/browse/0.3.4/images/arxiv-logo-fb.png){kind=logo}
[2602.20399] GeoPT: Scaling Physics Simulation via Lifted Geometric Pre-Training
Summary
GeoPT introduces a novel approach to scaling physics simulations by utilizing lifted geometric pre-training, enhancing model efficiency and reducing data requirements.
Why It Matters
This research addresses a significant challenge in physics simulation by proposing a method that leverages synthetic dynamics to improve model training efficiency. By reducing the need for labeled data and accelerating convergence, GeoPT has the potential to transform how neural simulators are developed and applied in various industries, making high-fidelity simulations more accessible.
Key Takeaways
- GeoPT enhances physics simulations by integrating synthetic dynamics with geometry.
- The model reduces labeled data requirements by 20-60%, improving training efficiency.
- Pre-training on over one million samples leads to significant performance gains in industrial benchmarks.
- The approach bridges the gap between geometry and physics, unlocking scalable neural simulation.
- GeoPT's methodology could influence future developments in various fields requiring physics simulations.
Computer Science > Machine Learning arXiv:2602.20399 (cs) [Submitted on 23 Feb 2026] Title:GeoPT: Scaling Physics Simulation via Lifted Geometric Pre-Training Authors:Haixu Wu, Minghao Guo, Zongyi Li, Zhiyang Dou, Mingsheng Long, Kaiming He, Wojciech Matusik View a PDF of the paper titled GeoPT: Scaling Physics Simulation via Lifted Geometric Pre-Training, by Haixu Wu and 6 other authors View PDF HTML (experimental) Abstract:Neural simulators promise efficient surrogates for physics simulation, but scaling them is bottlenecked by the prohibitive cost of generating high-fidelity training data. Pre-training on abundant off-the-shelf geometries offers a natural alternative, yet faces a fundamental gap: supervision on static geometry alone ignores dynamics and can lead to negative transfer on physics tasks. We present GeoPT, a unified pre-trained model for general physics simulation based on lifted geometric pre-training. The core idea is to augment geometry with synthetic dynamics, enabling dynamics-aware self-supervision without physics labels. Pre-trained on over one million samples, GeoPT consistently improves industrial-fidelity benchmarks spanning fluid mechanics for cars, aircraft, and ships, and solid mechanics in crash simulation, reducing labeled data requirements by 20-60% and accelerating convergence by 2$\times$. These results show that lifting with synthetic dynamics bridges the geometry-physics gap, unlocking a scalable path for neural simulation and potentially...