Prediction of density and surface roughness in LPBF-printed parts from recycled SS316L powder using random forest regression model

This paper uses machine learning algorithm to predict the density and surface roughness of 316L stainless steel parts manufactured using recycled powder based on the process parameter and powder characteristics. The advancement in Laser Powder Bed Fusion (LPBF) has enabled the production of complex and high-performance metallic components. However, the high cost of virgin powders and the substantial material waste generated during the AM process present economic and environmental challenges. The developed models with dedicated system interface were built on the understanding of the effect of powder characteristics on the part properties, which included layer thickness, hatch spacing, laser power, and scanning speed. Feature relationships were analysed using a correlation heatmap, highlighting strong interdependencies such as the inverse correlation between density and surface roughness (R = 0.98), and the alignment between laser power and scanning speed (R = 0.74). The RFR model was trained on datasets of varying sizes, and its performance was evaluated using standard error metrics (MAE, MSE, RMSE) and the coefficient of determination (R²). The model achieved high predictive accuracy, with an R² of 0.821 for density and 0.795 for surface roughness from the initial 16 dataset. Error metrics were significantly lower than previous studies: MAE of 0.218 for density and 0.256 for surface roughness. Performance improved with larger datasets, reaching an R² of 0.973 for density and 0.942 for roughness at 250 samples, though a slight drop in accuracy was observed beyond this point due to potential data noise. The Random Forest model demonstrated strong capability in predicting quality outcomes in LPBF processes, outperforming earlier works in both accuracy and consistency. The developed system provides a model tool to inform AM optimization effectively, especially when supported by carefully selected features and appropriate dataset sizes.