Physics-Informed Machine Learning in Design and Manufacturing: Status and Challenges

Abstract Machine learning (ML) technique is a critical tool to promote optimal design and ensure reliable and efficient products and processes in the manufacturing industry, since it can discover hidden knowledge and build complex relationships by learning patterns from data. However, the inherent ‘black-box’ nature of ML presents a major challenge in interpreting the mechanism and outcomes of the models. Moreover, reliable ML predictions are highly dependent on the amount and quality of training data. To address these issues, physics-informed machine learning (PIML), also known as scientific machine learning, has emerged as a new research field. PIML incorporates physical and domain knowledge into ML models to guide the ML training process, which enables more interpretable and reliable models. To fully leverage the advantages of PIML and promote the advancement of design and manufacturing, it is essential for researchers to understand the available PIML methodologies and the technical challenges of PIML methods. This article provides a systematic review of the state-of-the-art in PIML, focusing on the methodologies of integrating physics into ML. The PIML techniques can be grouped into three categories, including hybrid models, physical loss-based models, and physics-embedded architectures. Each of these categories is further stratified according to different integration approaches and ML models. The methods and applications of each technique are summarized. In addition, the technical challenges and potential opportunities of PIML are critically analyzed and discussed, providing a roadmap to narrow the research gaps in PIML.