Thermal prediction of additive friction stir deposition through Bayesian learning- enabled explainable artificial intelligence

Abstract

Additive friction stir deposition is an emerging solid-state metal additive manufacturing technology that can conveniently and economically produce fully-dense, high-end components, rendering it promising for high-value manufacturing industries, such as defense, aerospace, and space. Like in other metal additive processes, the thermal characteristics in additive friction stir deposition govern the quality and properties of the printed products but is challenging to predict and control in this complex manufacturing system. Recent advances in artificial intelligence provide a powerful data-driven prediction paradigm, which nevertheless necessitates extensive training datasets and struggles to be applied across diverse material systems that exhibit significant variations in thermal and material flow characteristics during printing. Relying solely on physics-based modeling also lacks success due to potential model discrepancies, unknown material parameters, and high computational cost. Here, we propose to address these challenges by developing an explainable artificial intelligence scheme via Bayesian learning, in which the physics-based surrogate model is calibrated and updated via machine learning of in-situ monitoring data, resulting in a physics-informed, data-driven model for temperature distribution during additive friction stir deposition. The efficacy of this approach is demonstrated in printing of an Al-Mg-Si alloy, in which fast, accurate temperature prediction is achieved using a moderate number of physics simulation trials and a small number of in-situ measurements. This approach also offers physical insights into previously unknown parameters, such as the interfacial slip/stick coefficient and back heat transfer constant.