Physics-Informed Neural Network for Real-Time Thermal Analysis and Melt Pool Behavior in Multi-Material Laser Metal Deposition

The Laser Metal Deposition (LMD) operation is characterized by rapid thermal cycles and complex thermal-material interaction. This work focuses on predicting peak melt pool temperature for distinct substrate-powder combinations using a pyrometer sensor and an infra-red camera integrated with the LMD experimental set-up for real-time data extraction and characterization of the bead features. In addition, this study suggest application of dual-model architecture that integrates governing physical laws with data-driven learning approach using PINNs. For the development of PINN model, an extensive experimental investigation was conducted for three different substrates (Al3002, Ti and Mild steel) and five distinct powders (Ti6Al4V, AA2024, SS316L, Cu, and Nitinol). For ensuring physical consistent predictions, governing heat conduction equation along with appropriate boundary conditions are incorporated into the loss functions. Thereafter, the PINN model’s performance was benchmarked against machine learning algorithms namely, Random Forest (RF), Polynomial Regression (PR), Decision Trees (DT), Artificial Neural Networks (ANN). The developed PINN temperature estimated module outperformed all benchmarks with an R ² of 0.97 and a Mean Squared Error (MSE) of 417.67 K². While, the bead geometry prediction module yielded an R ² value above 0.96. Based on SHAP analysis, it was identified that scanning speed and beam power were the two most dominant input parameters in LMD process that influences the predictive model performance. This research work was an attempt to emphasize the effectiveness of hybrid approaches that integrates physical principles with machine learning especially in industrial regimes with limited data sets.