Harmonizing physical and deep learning modeling: A computationally efficient and interpretable approach for property prediction

Abstract

Physical modeling and deep learning are known for their respective advantages in interpretability and computational efficiency. Nonetheless, efficiently predicting properties of metallic materials while maintaining interpretability presents a formidable challenge. This study proposes a novel solution by introducing dual-output deep learning model that simultaneously predicts stress-strain partitioning and mechanical properties through a two-component architecture. The initial component uses U-Net model trained on stress and strain partitioning generated from crystal plasticity (CP) simulations, thereby enhancing interpretability. Subsequently, this information is used to predict the properties in the second component. The prediction results demonstrate the validity of this approach, accurately predicting high stress at the martensite-martensite interface, high strain at the ferrite-martensite interface, and properties. In addition, the minimal computational cost significantly improves efficiency compared to conventional CP method. This innovative methodology represents a significant advancement, achieving harmonious balance between interpretability, computational accuracy, and efficiency in properties prediction of metallic materials.