Data-driven prediction and surrogate-based parameter selection for density and surface roughness in SLM of 316L SS under data scarcity

Selective laser melting (SLM) of 316L stainless steel involves strongly nonlinear interactions between processing parameters and resulting part properties, which renders conventional trial-and-error optimization strategies inefficient. This study proposes a data-centric framework that combines machine learning with evolutionary optimization to accurately predict and computationally select candidate processing parameters for surface roughness Ra and relative density. Support vector regression (SVR), random forest, XGBoost, CatBoost, TabNet, artificial neural networks (ANN), and a stacked ensemble, were evaluated under different data configurations. Results clearly demonstrated that data availability is the primary factor influencing predictive accuracy. Under data-scarce conditions, CatBoost was the most effective at predicting surface roughness, achieving a coefficient of determination (R²) of 0.83. Data augmentation substantially improved all models, enabling the ANN and the SVR-based stacking model to achieve R² values of 0.95. For the simultaneous surface roughness and density prediction, the multi-output ANN outperformed all other approaches, recording R² values above 0.98. This best-performing ANN surrogate model was subsequently coupled with the non-dominated sorting genetic algorithm II and differential evolution to select key SLM processing parameters. The computationally identified candidate parameter sets exhibited extremely low surrogate-predicted deviations from the specified target values, below 10⁻³ % for density and 10⁻⁴ µm for Ra. Although no additional SLM experiments were performed to validate these combinations, their physical plausibility was assessed by mapping them onto known melting regimes. Overall, the proposed framework provides a robust computational approach for the data-driven exploration and screening of promising processing windows in SLM, prior to experimental validation.