Physics-Informed Thermal Field Reconstruction for Melt-Pool Monitoring in Laser Powder Bed Fusion

Abstract: Laser Powder Bed Fusion (LPBF) is an Additive Manufacturing (AM) process used to manufacture complex, high-performance components in aerospace and automotive fields. Despite its advantages, LPBF is prone to process anomalies that may lead to defects, reduced part quality, or build failure. Melt-pool monitoring is used for capturing and analyzing the dynamics of the manufacturing process. In-situ, layer-wise image data of the melt pool can be treated as a time series
and modeled for online anomaly detection. However, acquiring fully healthy experimental data over an entire layer, or even over simple scan segments, remains challenging because the process is naturally sparse, highly dynamic, and contains unavoidable anomalies. 

To address this limitation, simulation results enhanced with experimental process data can be used to formulate a dynamic data-driven monitoring and diagnostic framework for LPBF. The main challenge is the high computational cost, since accurate thermal-field prediction requires solving multi-physics partial differential equations (PDEs), which can be expensive even for relatively small prints. For this reason, machine-learning-based surrogate models can be used to accelerate full-field temperature prediction. In this work, a Physics-Informed Neural Network (PINN) is developed and assessed for temperature-field reconstruction. The model is trained using simulation data generated under varying process parameters, including laser power, scan speed, and geometry, while a simplified heat-transfer PDE is incorporated as the physics-based constraint. Once trained, the PINN predicts the temperature field with significantly
reduced computational time, allowing in-situ monitoring and anomaly diagnosis while maintaining low error relative to simulation results. This framework supports dynamic data-driven melt-pool monitoring by enabling fast thermal-field prediction for healthy baseline generation and future in-situ anomaly detection in LPBF. 

Keywords: Dynamic data-driven systems · Laser powder bed fusion · Additive manufacturing · Melt-pool monitoring · Physics-informed neural networks · Thermal-field prediction · Anomaly detection