Ruihao Yuan , Jinshan Li', Torah Lookman
State Key Laboratory of Solidification Processing, Northwestern Polytechnical University, Xi'an 710072, China;

AiMaterials Research LLC, Santa Fe, NM 87501, USA 

EXTENDED ABSTRACT: Advanced metal structural materials, crucial in aerospace and other industries, achieve tailored performance through variables like composition, processing, and microstructure. However, their complex interplay makes establishing effective composition-processing-structure-performance models challenging. A兀ificial intelligence techniques, including machine learning, have recently gained prominence in metal structural materials research, reducing development cycle and costs for new alloys. Obtaining microstructure and performance data for metal structural materials, particularly in extreme conditions, is difficult, hindering predictive model accuracy.
To address this, the report suggests pre-training self-supervised models using abundant unlabeled data. These models enhance mechanical performance prediction using limited labeled alloy data, and the report aims to understand their improved performance from an interpretability perspective, guiding alloy design. The approach comprises two main aspects: (i) Using natural language processing, latent knowledge about titanium alloy design is extracted from extensive unlabeled data, generating feature embeddings for alloys. These embeddings enable simultaneous predictions of various alloy properties while exploring the alloy composition space for optimal performance. A pre-trained model, highly adaptable and interpretable through attention mechanisms, is trained using the Transformer architecture on a large dataset and fine-tuned for downstream tasks, offering tailored guidance for titanium alloy design. (ii) In a labeled dataset featuring 25 sets of superalloy electron backscatter diffraction (EBSD) images and performance data, establishing a structure-performance relationship is challenging. To address this, a pre-trained model is obtained by combining a Variational Autoencoder (VAE) and image enhancement techniques using unlabeled data. This pre-trained model excels at extracting comprehensive and robust microstructure information, improving mechanical performance prediction. By comparing microstructural differences between high and low­performance alloys through anomaly detection algorithms, critical areas influencing alloy mechanical performance, such as small-grain regions, are identified.
Keywords: Pre-training models; Feature extraction; Small data; Alloy design;