Data-driven modeling of AlSi10Mg mechanical properties: predictive approach to process parameters selection in laser powder bed fusion

• Data-driven and physics-informed modeling optimizes LPBF AlSi10Mg with high accuracy. • Bidirectional LSTM outperforms RNN/LSTM with < 10 % error and physics-aligned interpretations. • Symbolic regression balances parameter influence for transparent process optimization. • Robust generalization reduces costs and improves part performance in industrial applications. • Machine learning framework accelerates AlSi10Mg development for tailored LPBF properties. Data-driven methods for accelerated design and optimization of additively manufactured AlSi 10 Mg components based on the prediction of their mechanical properties are presented. The bidirectional long short-term memory demonstrates superior performance compared to symbolic regression, achieving maximum relative errors below 7.83 % using 5-fold cross-validation for all predicted mechanical properties. Despite the possibility of achieving accurate predictions, symbolic regression and random forest methods yield feature importances that conflict with established LPBF knowledge, suggesting sequence learning methods may be better suited for capturing the process’s complex relationships. The models, validated against experimental datasets, exhibit accuracy within 10 % of typical physical measurement tolerances, demonstrating their industrial applicability for rapid process parameter screening. This approach enables efficient exploration of the additive manufacturing process space and facilitates the design of components with tailored mechanical performance.