Physics-informed machine learning for manufacturing applications

Physics-Informed Machine Learning (PIML) has emerged as an innovative approach that integrates previous knowledge from physics into machine learning models, improving their ability to solve engineering problems. Within the PIML framework, Physics-Informed Neural Networks (PINNs) have gained importance due to their ability to enforce physical laws as constraints directly during the training process, reducing the need for large and labeled datasets. This thesis begins with a theoretical exploration of PIML, distinguishing PINNs as a specific class of physics-informed machine learning models. We then analyze the core principles and key challenges of PINNs, including simple applications of the framework to illustrate its foundational concepts. Building on this theoretical background, we applied the PINN framework to engineering and manufacturing scenarios, starting with a 2D linear elasticity model, where a PINN is used to approximate stress and displacement fields subject to physical constraints. The study is then extended to a 2D transient heat diffusion problem and after reviewing the theoretical background of Resistance Spot Welding (RSW), we developed a parametric equation to estimate the nugget growth and implemented it with an informed machine learning approach. A comparative analysis with traditional numerical methods, specifically the Finite Element Method (FEM), is done to validate the predictions of the PINN models for the simulated scenarios.