EXTENDED ABSTRACT: Advanced metallic structural materials, such as superalloys, have important applications in ffelds like aerospace. The microstructure, as a bridge connecting composition, process, and performance, plays a crucial role in understanding physical phenomena like strengthening mechanisms and optimizing performance. Currently, microstructural characterization of metallic materials largely relies on expert experience, which inevitably introduces subjectivity, and is inefficient, making it difffcult to handle the increasing scale of data. Due to its superior image processing capabilities, deep learning has been applied to microstructure characterization in recent years. However, characterizing the microstructure of metallic materials is challenging due to the small available data scale, and directly training deep learning models often leads to overfftting. Pre-trained models or transfer learning may offer solutions, but the source and target data often belong to different domains (e.g., pre-trained models based on ImageNet), resulting in poor transfer performance. More importantly, deep learning often overlooks the vast amount of physical laws or constraints embedded in the microstructure, leading to algorithm-driven designs that do not align with physical realities.To address these issues, Inconel 625 superalloy was selected as the research object. Starting from 25 sets of microstructure image-yield strength data, methods were developed to accurately predict mechanical properties based on a small number of microstructure images and to perform high-ffdelity inverse design of the microstructure. (1) First, without considering performance labels, image augmentation algorithms were used to signiffcantly expand the 25 sets of images. Then, a variational autoencoder (VAE) was used to learn from unlabeled data, resulting in a latent space representation of the microstructure that is performance-neutral, more robust, and has better generalization capabilities. Applying the pre-trained model to data with performance labels improved the microstructure characterization ability and revealed that the latent variables followed the Hall-Petch relationship with yield strength, signiffcantly improving the prediction accuracy of yield strength. In contrast, convolutional neural networks (CNNs) were found to fail in learning structured microstructure representations, resulting in severe overfftting. (2) Based on this, attempts were made to achieve inverse design of the microstructure using the trained VAE model. However, the image quality generated by the VAE alone was poor, with signiffcant distortion or loss of physical information such as grain boundaries. By introducing U-Net, a diffusion probabilistic model (DDPM), and designing a loss function, the generation quality of microstructure images was effectively improved. Comparing experimental images with generated images revealed good consistency in physical characteristics such as grain area and grain orientation. A multi-objective performance-oriented microstructure active learning design framework was further developed, enabling high-ffdelity microstructure designs based on speciffc performance requirements.

Keywords：Microstructure characterization; deep learning; strengthening mechanisms; inverse design.