Data-driven complementation and prediction of process-relative density data for SLM IN718 alloy

Optimization of multiple process parameters remains a challenge in additive manufacturing, and building data-driven property prediction models serve as an effective approach to address this challenge. Accumulating literature data has laid a fundamental data foundation for data-driven modeling. In this work, process parameters (laser power, scanning speed, hatch spacing, and layer thickness) and corresponding relative density data extracted from published studies on IN718 alloy fabricated by selective laser melting (SLM) are adopted as dataset samples. The expectation-maximization (EM) algorithm is employed to impute missing parameter values from collected literature. Three prediction models for relative density are established based on the sparrow search algorithm (SSA)-optimized generalized regression neural network (GRNN), the Kepler optimization algorithm (KOA)-optimized random forest (RF), and extreme gradient boosting (XGBoost). Statistical evaluation indicators including coefficient of determination (R²), root mean square error (RMSE), mean absolute error (MAE), and mean relative error (MRE) verify that all three models achieve favorable prediction accuracy. Among them, the KOA-XGBoost model delivers the optimal predictive performance with R², RMSE, MAE, and MRE of 0.967, 0.806, 0.437, and 0.47%, respectively; the KOA-RF model ranks second, with corresponding values of 0.938, 1.102, 0.466, and 0.52%; the SSA-GRNN model exhibits relatively inferior accuracy, with R², RMSE, MAE, and MRE equal to 0.899, 1.399, 0.588, and 0.63%, respectively. Satisfactory prediction outcomes are obtained when three models are validated against independent experimental datasets, and the accuracy ranking follows the order: KOA-XGBoost>KOA-RF>SSA-GRNN, demonstrating robust stability and strong generalization capability of the developed models.