Physics-informed machine learning across manufacturing processes: Recent advances, challenges, and directions

초록

Data-driven methods have shown remarkable promise across manufacturing disciplines. However, their application to manufacturing processes, where physical behaviors dominate, faces several limitations including the need for large datasets, the risk of physically implausible results, and limited interpretability. Physics-informed machine learning (PIML), which incorporates governing physical laws into deep neural networks, has emerged as a promising direction for overcoming these limitations by enabling physically consistent and data-efficient modeling. In this article, we systematically analyze PIML applications from a manufacturing-process perspective. We organize the manufacturing landscape into five representative categories based on their dominant physical mechanisms and the availability of relevant literature: mechanical processes, chemical processes, thermal-driven processes, additive manufacturing, and semiconductor fabrication. We comprehensively review and discuss domain-dependent characteristics, including governing physics, data regimes, and modeling objectives across manufacturing processes. We further discuss fundamental challenges of PIML approaches, along with practical considerations for real-world deployment. This review aims to provide a foundation for accelerating physics-informed learning and its advancement within manufacturing research and practice.

키워드