Process effect prediction method based on visual generative models

This study aims to develop a conditional visual generation model that maps process-structure relationships, generating scanning electron microscopy (SEM) images of metallographic microstructures under specified conditions based on input process parameters for process improvement reference. The Omni-loss is used to optimize the conditional discriminator/generator. A MAE-CGAN model was proposed to demonstrate the fidelity of generated microstructures and their correlation with applied processing methods. The model integrating visual transformer and Conditional Generative Adversarial Networks (CGAN) offers several advantages: the CGAN framework theoretically supports modeling any material system; trained generators can produce specific microstructure images under designated process conditions. And the prediction datasets are evaluated using quantitative indicators proposed in previous work. The results showed that the pre-trained Masked Auto-Encoder (MAE) model demonstrated robust critical structural feature extraction capabilities with the relatively small image dataset. The quantitative analysis of existing samples yielded results consistent with the expert evaluations, and the generated samples hold great potential for guiding future experimental settings.