Physics-Constrained Diffusion for Lightweight Composite Material Design

Hangwei Qian, Yang He, Bingjin Chen, Mohit Sharma, Ivor Tsang

Keywords: Physics-Constrained Diffusion, Generative Models, Diffusion, AI for Material Science, Composite Material, Inverse Design

TL;DR: We propose a physics-constrained diffusion model tailored for composite material design.

Abstract: Composite materials play a pivotal role in diverse engineering applications, particularly in the development of lightweight yet high-performance structures. However, their generative design has received far less attention than that of other material classes. A major challenge is that existing generative models often push component weights toward extreme minima, sometimes even yielding negative values, which are physically impossible for material compositions. In this work, we propose a diffusion-based generative framework tailored for lightweight composites. Specifically, we introduce a physics-constrained diffusion model (PCDiff) that integrates domain-specific constraints into the denoising process, ensuring generated candidates are both high-fidelity and physically plausible. In particular, we enforce two key constraints, i.e., non-negativity and sum-to-one conditions on composite compositions, through regularization within the diffusion process. Experimental evaluations demonstrate that our approach consistently outperforms existing generative models in terms of validity, density, and coverage with respect to target physical properties. This study underscores the potential of physics-guided generative modeling for accelerating the discovery of lightweight composites.

Submission Track: Findings, Tools & Open Challenges

Submission Category: AI-Guided Design

Institution Location: Singapore, Singapore

AI4Mat Journal Track: Yes

AI4Mat RLSF: Yes

Submission Number: 88