Abstract
Neural physics simulators are often highly sensitive to mesh topology—the specific triangulation of a 3D shape—even when the underlying geometry remains identical. This sensitivity poses a significant challenge for high-fidelity simulations, such as radar sensing or aerodynamics, where consistent predictions are required across varied mesh representations. This paper demonstrates that unsupervised autoencoder pretraining on large-scale mesh datasets significantly reduces this sensitivity. By forcing models to learn the underlying "language" of geometry before exposure to specific physical tasks, these models produce more consistent predictions across diverse topological representations.
Introduction
In computational physics, objects are typically represented as meshes. However, there is no unique mesh for any given shape; a flat surface might be represented by two large triangles or thousands of smaller ones. While the underlying physics remains constant, neural networks often overfit to these specific mesh structures.
Standard deep learning techniques like data augmentation (e.g., vertex jittering) are often insufficient in physics contexts. Because simulators are hyper-sensitive to minute surface changes—where a tiny bump can drastically alter a radar signature—one cannot simply "jitter" data without re-running expensive simulations to obtain new ground-truth labels. This creates a bottleneck: how can we train models to be robust to mesh variations without an impractical amount of simulated training data?
Methodology
The authors investigate whether pretraining can bridge the gap between topological sensitivity and physical accuracy. They focus on a radar cross-section (RCS) simulation task, where the objective is to predict a radar response from a 3D mesh input.
Model Architecture
The proposed architecture consists of three primary components:
- Face Embedding: This layer converts mesh faces into vectors. The researchers compared three approaches: Direct (geometric features), Graph-based (using convolutions to capture neighborhood information), and Tokenization-based (using a learned codebook).
- Aggregator: A Transformer-based module that pools individual face embeddings into a global object representation.
- Decoder: A Multi-Layer Perceptron (MLP) that maps the global representation to the final radar response.
Training Strategies
The core of the study compares three distinct training paradigms:
- Training from Scratch: The model is trained directly on the RCS task with no prior geometric knowledge.
- Supervised Pretraining: The encoder is pretrained to classify objects (e.g., identifying a "chair" or "table") using the ShapeNet dataset.
- Unsupervised Autoencoder Pretraining: The encoder is trained to reconstruct the mesh itself. This approach forces the model to capture granular surface manifold properties.
Experimental Setup
The study utilized two primary datasets:
- Basic Shapes Benchmark: A new dataset introduced by the authors to isolate the effects of topology. It contains 3,000 objects (Cubes, Cylinders, Spheres). Each object features one "simple" mesh and 99 "complex" variations that are topologically distinct but geometrically identical.
- ShapeNet: A standard dataset of approximately 20,000 3D models used for the pretraining phase.
Key Results and Insights
The primary metric for success was Variation MSE, which measures the change in a model's prediction when the mesh topology changes for the same geometric object.
The Power of Pretraining
The results indicate that Graph-based embeddings combined with Autoencoder pretraining achieved the best performance, yielding a Variation MSE of 125.6. In contrast, training the same architecture from scratch resulted in a Variation MSE of 216.6—nearly doubling the sensitivity to topological noise.
Reconstruction vs. Classification
One of the most significant findings was that autoencoder (reconstruction) pretraining outperformed supervised (classification) pretraining. While classification requires identifying a general category, reconstruction requires a deep understanding of the mesh's surface properties. This granular understanding translates more effectively to downstream physics simulations.
Local Features and Topology as Noise
The researchers found that graph-based embeddings significantly outperformed direct embeddings. Because physics—such as radar waves bouncing off a surface—is inherently local, the relationship between adjacent faces is more critical than the absolute coordinates of individual vertices. Pretraining effectively teaches the model to treat topological variations as "noise" to be ignored, focusing instead on the invariant geometric features.
Limitations and Future Work
Despite the improvements, a performance gap remains. An "Ideal" model trained on every possible mesh variation achieved a Variation MSE of 1.2, suggesting that while pretraining reduces sensitivity, it does not yet eliminate it.
Future research directions include:
- Scaling: Utilizing much larger 3D datasets (millions of objects) for pretraining, analogous to the scaling seen in Large Language Models.
- Loss Function Engineering: Developing pretraining losses that explicitly enforce embedding consistency across different meshes of the same object.
- Equivariance: Integrating rotation-invariant structures into mesh embeddings to further harden models against irrelevant spatial changes.
Summary updated