PHYSICS-INFORMED LOSS FUNCTIONS WITH EXPLAINABLE AI TO PREDICT EMISSION IN POWDER BED FUSION

Despite the growing prevalence and use of custom metal-based parts produced by powder bed fusion (PBF) processes, overheating remains an issue that prevents stability and zero-defect parts from being printed. Data-driven methods utilizing existing data of prior prints are often employed to help avoid overheating. Despite the impressive predictive power associated with these methods, there is no guarantee that the model's behavior will align with the physical understanding of the printing process. That is, the model may capture a trend within the data that is impossible in the real-world setting. One solution to this problem is using custom physics-based loss functions designed with the known physics knowledge of the PBF printing process to enforce model behavior that obeys physical understanding. This paper proposes two custom loss functions to enforce a physics-obeying model. These loss functions are then implemented into a developed sequential neural network, and the model's predictive performance is evaluated using root mean square error. To increase transparency and interpretation of the model, an explainable AI technique, Accumulated Local Effects (ALE), measures the ability of the loss functions to enforce physical behavior in the model. It is observed that the custom loss functions successfully enforce physical behavior in the model with a slight compromise in predictive power.