Design of superalloys with multiple properties via multi-task learning

Abstract

The development of new materials requires the collaborative design of multiple properties, requiring an analysis of the interactions amongst material composition, processing methods, and individual properties. Traditional data-driven materials design approaches typically rely on single-task models that operate independently, often neglecting the shared insights across related tasks. To overcome this limitation, we propose a collaborative design framework that employs multi-task learning for the development of novel Co-based superalloys. In this framework, six thermodynamic and microstructural property tasks share a common encoder, which effectively captures the underlying influence of alloy compositions across different properties. Each task then utilizes its own dedicated decoder to ensure precise predictions. As a result, the average normalized error for the predictions of the six properties is reduced by 37.5 % compared to conventional single-task learning methods. Furthermore, latent high-dimensional variables are extracted from the common encoder, and utilized to identify promising exploration directions for optimal properties, as indicated by the projection of these variables, which aids in screening new alloys. We successfully designed new alloys that satisfy the targeted criteria: low density (<9 g cm^-3), suitable freezing ranges (<60 °C) and processing windows (>80 °C), optimal γ′ sizes (<550 nm after aging at 1100 °C for 168 h and <200 nm after aging at 1000 °C for 24 h), high γ′ solvus temperature (>1200 °C) without detrimental phases, and strong oxidation resistance. This framework represents a promising approach for collaborative materials design, leveraging shared information to enhance the development of multiple properties.