Metal additive manufacturing, such as Laser-Based Powder Bed Fusion (L-PBF) and Directed Energy Deposition (DED), presents an enabling opportunity for creating complex metal parts with design freedom. The unique thermal cycle of rapid heating, fast solidification, and melt-back during metal AM may cause very complex melt pool dynamics. The complex kinetic process and thermal history may lead to various quality issues of the printed parts. Therefore, the understanding and prognosis of metal pool dynamics remains the central intractable problem for printing high-quality metal parts or new alloys. Compared with physics-based simulation models, machine learning has the potential to handle high-dimensional and massive process data for efficient surrogate modeling and decision-making. However, pure data-driven machine learning models are black-box, inherently computation-intensive, and storage-intensive. A deep knowledge gap exists between machine learning and physics-based modeling in predicting melt pool dynamics.To address these limitations, this dissertation has developed a Physics-Informed Machine Learning (PIML) framework that integrates governing physical laws and small datasets for the forward prediction of key thermomechanical responses, including temperature evolution, melt pool flow behavior, and thermal stress distribution, as well as the inverse learning of unknown constants in the governing equations. The proposed framework further leverages the complementary strengths of physics-based and data-driven methods – combining the physical interpretability of finite element analysis (FEA) with the efficiency of machine learning.The research progresses from standalone Physics-Informed Neural Network (PINN) models to hybrid FEA-regulated PINN (FEA-PINN) architectures. These hybrid models address common challenges in conventional PINNs, such as error accumulation and poor generalization in time-dependent problems, by incorporating periodic corrective feedback from FEA simulations. The resulting framework demonstrates accuracy comparable to traditional FEA models while significantly reducing computational cost and improving scalability for parametric studies.
Rahul Sharma (Thu,) studied this question.