Abstract

       Over the last few years, Physics-informed (PI) machine learning methods have established themselves as ideal candidates to solve quality variability in laser powder bed fusion additive manufacturing due to complex interactions, defect formations, and building parameters. The research develops a novel physics-informed learning framework that fuses physics knowledge with further state-of-the-art ensemble or neural network-based models to predict build outcomes and dynamically optimize processing parameters in quasi-real-time settings. Key physics-informed descriptors including volumetric energy density (VED) and linear energy density (LED) were extracted and fed into multiple ensemble algorithms including XGBoost, Random Forest, LightGBM, CatBoost, and gradient boosting. Simultaneously, the physics-informed neural networks were trained with thermodynamic constraints for physical consistency of the predictions. The Harvard LPBF database provided the experimental dataset which contains 1579 samples from six commercially important alloys. The findings revealed that the physics-informed neural network yielded greater performance, with an overall coefficient of determination of 0.827, with values exceeding 0.78 for four of the six alloys. Inclusion of descriptors based on physics considerably enhanced the predictive capability of the neural network, reducing the mean absolute error by up to 4.7 percent. Bayesian optimization was used to find the best parameters for mitigating defects. Moreover, the proposed framework achieved closed-loop control using layer-wise simulation with an average optimization time of 3.8 ms per step. The results obtained show that interfacing of physical knowledge with machine learning enhances reliability, prediction capability, and real-time process control in LPBF manufacturing.