MIT's GeoPT Halves Physics Simulation Data Needs, Doubles Convergence via Geometric Pre-Training

GeoPT: Scaling Physics Simulation via Lifted Geometric Pre-Training

A paper accepted at ICML 2026 from MIT CSAIL and Tsinghua University presents GeoPT , a framework that pre-trains neural simulators on large-scale 3D geometry data augmented with synthetic dynamics, dramatically improving data efficiency and convergence speed for downstream physics simulation tasks.

Core Idea: Geometry + Synthetic Dynamics

Instead of relying on expensive industrial CFD or FEM simulations for pre-training, GeoPT leverages freely available 3D geometric datasets (e.g., ShapeNet). The pipeline:

3D geometry → random velocity field assigned to points → particles move along velocities → particles stop on contact with surface → time-varying geometric features extracted

This process converts static shapes into dynamic geometric–physical interactions (wind resistance, water impact, collision deformation, light propagation) at a fraction of the cost. The authors report that generating one geometric-dynamics sample tracking 36,864 points takes 0.2 seconds on 80 CPUs , which is millions of times faster than industrial CFD.

Pre-training Scale

Source: ShapeNet (cars, airplanes, ships, robots)

10,000+ unique geometries

Multiple random dynamic conditions per geometry

Total pre-training samples: 1,346,300

Downstream Evaluation

GeoPT is tested on five industrial physics simulation tasks with deliberately small training sets (~100 samples per task, 20–50 test samples):

Automotive aerodynamics (surface pressure, flow field)

Aerospace aerodynamics

Ship hydrodynamics (hull resistance, waves)

Vehicle collision (max Von Mises stress)

Light transport on complex geometry

Key Results

Physical training data reduced by 20–60% while matching full-data performance.

Convergence accelerated up to 2× compared to training from scratch.

For ship hull simulation (air + wave forces), labeled data reduced by 60% and peak accuracy reached 4× faster than the strongest baseline.

GeoPT outperforms state-of-the-art neural simulators across all benchmarks.

Ablation: Geometry Diversity > Trajectory Diversity

The authors investigate whether pre-training should prioritize more geometries or more dynamic trajectories per geometry. Results show that geometric diversity is far more critical . A model exposed to many distinct car/airplane/ship shapes learns abstract geometry–physics relationships, whereas seeing many trajectories of a single shape yields limited generalization. Scaling to a larger, more diverse geometry repository (GeoPT-Huge) further improves performance on unseen tasks like car collision deformation and light transport—even though the model was never explicitly trained on 3D collision or ray-tracing physics.

Scaling Behavior & Future Work

Unlike neural simulators trained from scratch that overfit on small data, GeoPT continues to benefit from increased model and data scale, demonstrating a viable scaling path for neural physics engines. Future directions include simulating weather patterns, testing material properties, and generating realistic video.

Paper: GeoPT: Scaling Physics Simulation via Lifted Geometric Pre-Training (ICML 2026) OpenReview PDF: https://openreview.net/pdf/3df9a842bf6bf180192840eca36de1f7eda780d3.pdf MIT News: https://news.mit.edu/2026/ai-models-simulate-wider-range-of-real-world-scenarios-0810

Code example

来源：ScienceAI
本文
约2000字
，建议阅读
5
分钟
既然真正的物理模拟数据这么昂贵，那能不能先让 AI 只用大量免费的 3D 几何数据，让它了解日常机械互动，更准确地模拟现实世界。