Machine learning analysis for melt pool geometry prediction of direct energy deposited SS316L single tracks

Among the metal additive manufacturing techniques, directed energy deposition (DED) is least investigated, particularly in the context of machine learning (ML)-based process-structure correlation. To address this aspect, we performed the planned experiments for continuous deposition of single tracks of austenitic stainless steel (SS316L) by varying the process parameters. Based on extensive analysis of the melt pool quality in terms of defect morphology, the process map for DED of SS316L was created. This can help in decision-making regarding process parameter selection. Within the limitation of a small dataset, a number of statistical learning algorithms with tuned hyperparameters were trained to predict the geometrical parameters of single tracks (width, depth, height, track area, melt pool area). Based on an extensive evaluation of the performance metrics and residual error analysis, the Gaussian Process Regression (GPR) model was found to consistently predict all of the geometrical parameters better than other ML algorithms, with a statistically acceptable coefficient of determination (R2) and root mean square error (RMSE). An attempt has been made to rationalise the superior performance of GPR in low data regime, over linear regression or gradient boosting machine (GBM) in reference to the underlying statistical framework.