🤖 machine learning

GeoPT: Scaling Physics Simulation via Lifted Geometric Pre-Training

GeoPT introduces a unified pre-trained model for general physics simulation that bridges the gap between static geometry and dynamics by augmenting geometric pre-training with synthetic dynamics, thereby significantly reducing labeled data requirements and accelerating convergence across diverse industrial benchmarks.

Original authors: Haixu Wu, Minghao Guo, Zongyi Li, Zhiyang Dou, Mingsheng Long, Kaiming He, Wojciech Matusik

Published 2026-05-21

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Haixu Wu, Minghao Guo, Zongyi Li, Zhiyang Dou, Mingsheng Long, Kaiming He, Wojciech Matusik

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you want to teach a computer to predict how wind flows around a new car design or how a metal beam crumples during a crash. Traditionally, to teach this, you have to run massive, expensive computer simulations for every single training example. It's like trying to teach a student to drive by making them drive a real car on a real highway for every lesson—it's slow, dangerous, and costs a fortune in fuel.

The paper introduces GeoPT, a new way to train these "neural simulators" that is much cheaper and faster. Here is how it works, using simple analogies:

The Problem: The "Static Map" Trap

Usually, to train these AI models, researchers try to pre-train them on a huge library of 3D shapes (like a massive digital museum of cars, planes, and boats). They teach the AI to recognize the shape of these objects.

The problem is that shape alone doesn't tell the whole story.

The Analogy: Imagine showing a student a static map of a city. They can learn where the streets are. But if you ask them, "How will traffic flow if it rains and everyone is rushing to work?", the static map doesn't help. The map has no concept of movement or force.
The Result: When the researchers tried to use this "static map" training for physics, the AI actually got worse at predicting physics. It learned to group parts of a car together based on their shape, but it failed to understand how air would actually move over them.

The Solution: "Lifting" the Geometry

The authors realized they needed to bridge the gap between "static shape" and "dynamic physics" without running expensive simulations. They invented a method called Lifted Geometric Pre-Training.

The Core Idea:
Instead of just showing the AI a static 3D shape, they "lift" the shape into a dynamic world by adding fake wind and movement.

The Setup: They take a 3D shape (like a toy car).
The "Lift": They randomly shoot imaginary particles (like tiny wind gusts) at the car from all different angles and speeds.
The Game: They ask the AI: "If a wind gust hits this car from the left, how does the 'distance' to the car's surface change as the wind travels?"
- The AI doesn't need to know the real physics of air. It just needs to learn how the geometry interacts with the movement of these imaginary particles.
- It's like teaching a student to drive by having them navigate a virtual obstacle course with random wind gusts, rather than driving a real car on a real road.

Why This Works

By training on millions of these "what-if" scenarios (random wind hitting random shapes), the AI learns a universal intuition about how objects interact with forces.

The "Prior" Knowledge: The AI learns a "muscle memory" for physics. It understands that if something moves fast, it pushes harder; if it hits a wall, it stops. It learns the rules of the game without needing to know the specific score of the game yet.
The "Fine-Tuning": Once the AI has this general intuition, you only need a tiny bit of real data (20% to 60% less than before) to teach it the specific rules of a real-world task, like predicting the exact wind pressure on a specific new airplane wing.

The Results

The paper tested this on real industrial problems:

Aerodynamics: Predicting wind around cars, planes, and ships.
Crash Simulation: Predicting how metal deforms when a car hits a wall.
Light Transport: Even predicting how light bounces around a room (radiosity).

The Gains:

Less Data Needed: They reduced the need for expensive, real-world simulation data by 20% to 60%.
Faster Training: The models learned 2 times faster.
Better Accuracy: The predictions were more accurate, even with less data.

The Bottom Line

GeoPT is like giving a physics student a massive library of "what-if" scenarios to play with before they ever see a real textbook. By learning how shapes behave under random, made-up forces, the AI builds a strong foundation that makes it incredibly easy to learn the specific, real-world physics later. It turns a bottleneck (the high cost of generating real physics data) into a scalable path for the future of engineering design.

Technical Summary: GeoPT: Scaling Physics Simulation via Lifted Geometric Pre-Training

1. Problem Statement

Neural simulators have emerged as efficient surrogates for classical numerical solvers, capable of accelerating physics simulation in scientific discovery and engineering design. However, scaling these models to industrial fidelity is bottlenecked by the prohibitive cost of generating high-fidelity training data. Unlike vision or language, where web-scale data and self-supervised learning enable massive scaling, neural simulators remain dominated by supervised learning on data generated by expensive numerical solvers. For instance, generating a single industrial-fidelity aerodynamics sample can cost tens of thousands of CPU-hours.

While abundant off-the-shelf 3D geometries exist in public repositories, a fundamental gap prevents their direct use for pre-training: supervision based solely on static geometry ignores the dynamics (e.g., flow velocity, impact forces) that drive physical systems. Previous attempts to pre-train on geometry alone (e.g., predicting Signed Distance Fields or vector distances) fail to capture the coupled interaction between geometry and dynamics, often leading to negative transfer or degraded performance on downstream physics tasks.

2. Methodology: Lifted Geometric Pre-Training

The authors propose GeoPT, a unified pre-trained model that bridges the geometry-physics gap through dynamics-lifted geometric pre-training. The core innovation is augmenting static geometry with synthetic dynamics to enable self-supervision without requiring physics labels.

2.1. The Geometry-Physics Gap

The paper demonstrates that native pre-training (predicting static geometric features like vector distance fields from geometry alone) learns representations based on static shape cues. In contrast, physics simulations require representations that encode how spatial points correlate under specific dynamic conditions (e.g., aerodynamic flow). Native pre-training fails because it lacks any notion of dynamics, leading to aggregation patterns that do not align with physical flow structures.

2.2. Lifting Geometry to Physics Space

To bridge this gap, GeoPT introduces a "lifting" operation that expands the pre-training space from pure geometry to a joint geometry-dynamics space:

Synthetic Dynamics: Instead of relying on expensive physics simulations to determine velocity fields, the method constructs synthetic dynamics by randomly sampling per-particle velocity vectors ( $v$ ) from a bounded ball.
Trajectory Tracking: For a given geometry $G$ and a query point $x$ , the system simulates a particle trajectory over a time horizon $\tau$ using the equation $dx_t/dt = v \cdot \mathbb{1}_G(x_t)$ , where $\mathbb{1}_G$ is an indicator function that halts the particle at the geometry boundary.
Dynamics-Aware Supervision: The self-supervision target is the trajectory of geometric features (specifically vector distance fields) along these synthetic paths. The model is trained to predict the evolution of these features under the sampled velocity fields.

This process effectively teaches the model to obey conservation laws (specifically, the collisionless transport equation with sticking boundaries) under arbitrary dynamics, creating a "dynamics-aware prior."

2.3. Fast Fine-Tuning

After pre-training on over one million samples derived from diverse 3D shapes (cars, aircraft, watercraft), GeoPT is fine-tuned for specific physics tasks. The fine-tuning process replaces the randomly sampled synthetic velocities with task-specific velocity fields ( $V_S$ ) that encode the actual simulation settings (e.g., angle of attack, impact direction, freestream speed). This unified interface (Geometry + Velocity) allows a single pre-trained model to adapt to diverse physics domains.

3. Key Contributions

Lifted Pre-Training Paradigm: The paper introduces a novel self-supervised framework that augments static geometry with synthetic dynamics, enabling the learning of geometry-dynamics coupled representations without physics labels.
GeoPT Model: A pre-trained neural simulator trained on over 1.3 million solver-free samples, demonstrating consistent improvements across fluid and solid mechanics benchmarks.
Scalability and Efficiency: The method significantly reduces the dependency on expensive physics simulations, offering a scalable path to pre-training neural simulators using abundant geometric data.
Generalizability: The approach is shown to generalize across different physics domains (aerodynamics, hydrodynamics, crash simulation, and even radiosity) and different backbone architectures (e.g., Transolver).

4. Experimental Results

The authors evaluated GeoPT on five industrial-scale benchmarks:

Aerodynamics: DrivAerML (cars), NASA-CRM, and AirCraft.
Hydrodynamics: DTCHull (ships).
Solid Mechanics: Car-Crash simulation.
Radiosity: Light transport simulation (Cornell box).

Key Findings:

Data Efficiency: GeoPT reduces the labeled physics data requirements by 20% to 60% while achieving performance comparable to models trained from scratch with full data.
Convergence Speed: The method accelerates convergence by up to 2× compared to training from scratch.
Generalization: GeoPT shows larger improvements in tasks with high geometric diversity (e.g., DTCHull with varying hull curvatures) compared to tasks with limited geometric variation.
Cross-Domain Transfer: The model successfully transferred to radiosity simulation, a domain with fundamentally different governing physics, demonstrating its potential as a general-purpose prior.
Scaling: Experiments show that GeoPT benefits from increased model size (up to 32 layers) and diverse pre-training data, avoiding the overfitting issues often seen in limited-data industrial simulations.

5. Significance and Claims

The paper claims that GeoPT provides a scalable pathway for neural simulators by leveraging the asymmetry between the scarcity of physics labels and the abundance of 3D geometry data. By "lifting" the pre-training space to include synthetic dynamics, the authors successfully bridge the gap between static geometry and dynamic physics.

The significance lies in:

Alleviating the Data Bottleneck: It offers a practical solution to the high cost of generating training data for industrial design, potentially accelerating AI-aided software workflows.
New Self-Supervised Paradigm: It challenges the notion that pre-training must be confined to the native space of the data, demonstrating that introducing synthetic external factors (dynamics) can unlock transferable representations for higher-dimensional downstream tasks.
Foundation Model Potential: The results suggest that GeoPT can serve as a step toward general-purpose physics foundation models, capable of adapting to diverse physical systems through simple reconfiguration of the dynamics input.

The authors maintain a modest stance, noting that while the lifted paradigm bridges the geometry-physics gap, the current parameterization of dynamics (velocity fields) may not capture all nuances of complex material properties or boundary conditions, leaving room for future framework extensions.