Answers are grounded in this paper and cite their sources.
In this work, we present an eXplainable artificial intelligence (XAI)‐driven framework to automate the inspiration‐driven design of periodic metamaterial units based on stochastic microstructures (e.g., stochastic natural microstructures). The proposed framework leverages XAI as a critical geometric feature learning agent to drive a closed‐loop evolutionary design generation and optimization process. In each iteration, XAI methods identify property‐critical geometric features from exceptional samples in the current population, and these geometric features are employed to generate the next generation of microstructure designs using a feature‐steered diffusion (FSD) generative model. The FSD model integrates a variational autoencoder (VAE) latent space with a property‐conditioned diffusion backbone and employs a parameter‐efficient low‐rank geometry adaptation strategy to inject XAI‐identified geometric features (spatial priors) into the generative process without full retraining. This critical geometric feature identification–microstructure generation process is iterated until the convergence criteria are satisfied. Two XAI approaches, gradient‐weighted regression activation mapping (Grad‐RAM) and activation maximization texture synthesis (AMTS), are employed and compared in the framework. A bootstrapping‐based initial sample generation approach is also introduced to enhance the performance of the Grad‐RAM–driven iterative design process. To validate the effectiveness of this framework, three case studies covering maximization of Poisson's ratio, minimization of Poisson's ratio, and minimization of the coefficient of thermal expansion are presented. The case studies demonstrate that the proposed framework consistently steers evolution toward higher‐performance designs.
No takes yet. Share an insight, caveat, or question.
Feng et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: