GeoPT: Scaling Physics Simulation via Lifted Geometric Pre-Training
arXiv:2602.20399v1 Announce Type: new 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
arXiv:2602.20399v1 Announce Type: new 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 beyond. Code is available at https://github.com/Physics-Scaling/GeoPT.
Executive Summary
This article presents GeoPT, a unified pre-trained model for general physics simulation that bridges the geometry-physics gap through lifted geometric pre-training. By augmenting geometry with synthetic dynamics, GeoPT enables dynamics-aware self-supervision without physics labels, leading to improved industrial-fidelity benchmarks in fluid and solid mechanics. The results demonstrate a 20-60% reduction in labeled data requirements and a 2$ imes$ acceleration in convergence. This scalable path for neural simulation has significant implications for various industries and opens up possibilities for applications beyond physics simulation. The code is available on GitHub, allowing for further research and development.
Key Points
- ▸ GeoPT is a unified pre-trained model for general physics simulation
- ▸ Lifted geometric pre-training bridges the geometry-physics gap
- ▸ GeoPT achieves improved industrial-fidelity benchmarks in fluid and solid mechanics
Merits
Strength in Scalability
GeoPT's pre-training approach allows for a significant reduction in labeled data requirements, making it a more scalable solution for neural simulation.
Improved Performance
The results demonstrate a 20-60% improvement in industrial-fidelity benchmarks, indicating that GeoPT can produce high-quality simulations with reduced data requirements.
Synthetic Dynamics
The introduction of synthetic dynamics in pre-training enables dynamics-aware self-supervision without physics labels, a significant innovation in the field.
Demerits
Limited Generalizability
The results are specific to industrial-fidelity benchmarks in fluid and solid mechanics, and it is unclear whether GeoPT will generalize well to other physics tasks or domains.
Dependency on Synthetic Dynamics
The effectiveness of GeoPT relies on the quality of the synthetic dynamics used in pre-training, which may introduce additional complexity and variability.
Expert Commentary
GeoPT represents a significant advancement in the field of neural simulation, as it addresses a fundamental limitation in the existing approaches. The introduction of synthetic dynamics in pre-training is a creative solution to the geometry-physics gap, and the results demonstrate its effectiveness. However, further research is needed to fully understand the generalizability and robustness of GeoPT. Additionally, the dependence on synthetic dynamics is a potential limitation that requires careful consideration. Nevertheless, GeoPT has the potential to revolutionize the field of neural simulation and open up new possibilities for applications in various domains.
Recommendations
- ✓ Further research is needed to explore the generalizability of GeoPT to other physics tasks and domains.
- ✓ The development of more sophisticated synthetic dynamics methods could further improve the performance and robustness of GeoPT.
