Machine learning framework for accurate melt pool dimensions prediction in selective laser melting of Inconel 718 under high linear energy density: Surpassing Rosenthal equation limitations

Abstract

Melt pools generated in the course of the selective laser melting procedure dictate every major defect, including warping, lack of fusion, cracking, and porosity. Therefore, optimizing melt pool sizes for manufacturing quality components is essential. Various research groups developed analytical models based on physics to anticipate melt pool geometries quickly. Nevertheless, the predictive capability of analytical models is constrained by certain assumptions incorporated during their development. The predictive capacities of the Rosenthal equation were investigated in the current study. The Rosenthal equation showed quite good accuracy in predicting melt pool dimensions under lower linear energy density printing conditions. However, the prediction capabilities are reduced at high linear energy density process parameter settings. To overcome the limitation of the Rosenthal equation, machine learning methods were applied in the current research. Predictions generated by machine learning models are not inherently tied to the underlying physical phenomena governing a process; rather, they are contingent upon the quality of the data used during the training process In this study, a dataset was generated using the Rosenthal equation, encompassing various combinations of laser power and scanning speed, spanning from low to high linear energy density. Data points corresponding to lower linear energy densities, where the Rosenthal equation performed well, were used to train machine learning models. The research evaluated three machine learning models: neural network regression, support vector regression, and, K-nearest neighbors regression. Notably, all three machine learning models outperformed the Rosenthal equation, showcasing higher accuracy and reliability, particularly under high linear energy density printing conditions. The best-performing neural network model yielded root mean square error/mean absolute error values of 28.93/22.57 for melt pool width and 26.24/24.38 for melt pool depth prediction. While the support vector regression model performed poorly, registering the highest errors of 39.88/33.18 for melt pool width and 36.11/34.26 for melt pool depth.