Generative–predictive AI framework for joint design and mechanical prediction of bicontinuous composite microstructures

Abstract Designing high-performance composites requires exploring vast microstructural design spaces, which is computationally expensive for bicontinuous architecture using traditional simulations. We present an end-to-end artificial intelligence framework combining a denoising diffusion probabilistic model (DDPM) and a multimodal surrogate predictor to jointly generate and evaluate composite microstructures. Trained on 2000 phase-field-generated binary composites (0.55:0.45 stiff-to-soft ratio) with simulated stress fields, the DDPM co-generates realistic configurations and von Mises responses. From each stress map, 28 features and image embeddings feed a multimodal neural network that predicts Young’s modulus and ultimate tensile strength ( R 2 = 0.95 and 0.75). This generative–predictive pipeline accelerates composite design and inverse discovery. Graphical abstract