Attention to Detail: Fine-Scale Feature Preservation-Oriented Geometric Pre-training for AI-Driven Surrogate Modeling (2504.20110v2)

Abstract: AI-driven surrogate modeling has become an increasingly effective alternative to physics-based simulations for 3D design, analysis, and manufacturing. These models leverage data-driven methods to predict physical quantities traditionally requiring computationally expensive simulations. However, the scarcity of labeled CAD-to-simulation datasets has driven recent advancements in self-supervised and foundation models, where geometric representation learning is performed offline and later fine-tuned for specific downstream tasks. While these approaches have shown promise, their effectiveness is limited in applications requiring fine-scale geometric detail preservation. This work introduces a self-supervised geometric representation learning method designed to capture fine-scale geometric features from non-parametric 3D models. Unlike traditional end-to-end surrogate models, this approach decouples geometric feature extraction from downstream physics tasks, learning a latent space embedding guided by geometric reconstruction losses. Key elements include the essential use of near-zero level sampling and the innovative batch-adaptive attention-weighted loss function, which enhance the encoding of intricate design features. The proposed method is validated through case studies in structural mechanics, demonstrating strong performance in capturing design features and enabling accurate few-shot physics predictions. Comparisons with traditional parametric surrogate modeling highlight its potential to bridge the gap between geometric and physics-based representations, providing an effective solution for surrogate modeling in data-scarce scenarios.

Fine-Scale Feature Preservation in AI-Driven Surrogate Modeling

The paper "Attention to Detail: Fine-Scale Feature Preservation-Oriented Geometric Pre-training for AI-Driven Surrogate Modeling" presents a method designed for the sophisticated task of preserving fine-scale geometric features in AI-driven surrogate modeling. With the increasing reliance on surrogate models as alternatives to computationally intensive physics-based simulations, accurately predicting physics is contingent on the detailed geometric representation supplied by such models.

Overview

AI-driven surrogate modeling offers a promising alternative to resource-consuming simulations for 3D design and manufacturing. However, preserving fine-scale geometric features remains an unresolved challenge, especially for mechanical simulations sensitive to intricate design details. This paper introduces a self-supervised geometric representation learning method aimed at capturing these fine-scale geometric features in non-parametric 3D models. The proposed method separates geometric feature extraction from downstream tasks, leveraging geometric reconstruction to guide the learning of latent embeddings. Key innovations—such as near-zero level sampling and a batch-adaptive attention-weighted loss function—facilitate enhanced encoding of design features.

Methodology

The authors delineate a two-stage training strategy – pretraining followed by downstream application. During pretraining, a parametric graph neural network processes Boundary Representation (B-Rep) data to learn structured latent spaces focused on geometric reconstruction loss. This involves:

• Near-zero level sampling: Optimizes SDF value sampling near geometry surfaces to capture thin-shell features, reducing typical sampling inefficiencies.
• Batch-adaptive attention-weighted loss: Dynamically adjusts the loss function to focus on areas with significant geometric variation, indicating fine-scale changes.

For verification, case studies were conducted on crash box and bottle designs, chosen for their structural significance and complexity. The pretraining validated the methodology's ability to learn latent vectors that accurately predict design parameters, with R² scores consistently exceeding 0.99.

Results

The effectiveness of this approach is pronounced in few-shot learning scenarios. When applied to reaction force and nodal deformation fields, pretrained models outperform traditional parametric surrogate models in data-scarce environments. This holds particularly true when utilizing latent vectors directly without extensive retraining, showcasing the potential of geometric pretraining to bridge gaps in non-parametric data scenarios.

Implications and Future Directions

This paper addresses a critical gap—preservation of fine-scale geometric details in surrogate modeling—to enhance structural performance predictions. Practically, this approach yields refined predictions in mechanical stress analysis and deformation modeling without the need for simulation-based parameter information. Theoretically, it offers a robust foundation for further exploration into self-supervised learning for CAD data, especially as CAD-related workflows grow increasingly complex and data-rich.

Future research should focus on expanding the scalability of this method to diverse CAD repositories and exploring multi-modality integrations to enhance representation quality. Moreover, refining automated architecture selection processes could improve model robustness across varying dataset complexities.

In summary, this research provides a significant contribution to the field of AI-driven surrogate modeling by introducing techniques that preserve intricate geometric details vital for accurate physical predictions, particularly in data-scarce scenarios.