A new framework for generative design, real-time prediction, and inverse design optimization: application to microstructure

Abstract Optimizing composite materials for specific applications requires the ability to design microstructures that meet targeted mechanical performance, as well as to predict their behavior efficiently. This study introduces a new framework based on Rank Reduction Autoencoders (RRAEs) to enable fast and accurate predictions in both forward (mechanical response prediction) and inverse (microstructure design) directions. The framework employs two separate RRAEs models: one to encode and generate microstructure geometries and another to encode and predict their mechanical responses. These models are connected in latent space using regression mapping, enabling fast transitions between geometries and solutions. This approach supports real-time prediction of composite behavior as the microstructure evolves and enables inverse design by generating microstructures that match specific target properties. Using the RRAEs model, the framework creates a regularized and continuous latent space that allows accurate interpolation and reliable sampling. This facilitates efficient exploration of the design space and the discovery of new microstructure configurations with potentially improved mechanical performance.