Machine learning enabling prediction in mechanical performance of Ti6Al4V fabricated by large-scale laser powder bed fusion via a stacking model

Determining appropriate process parameters in large-scale laser powder bed fusion(LPBF)additive manufacturing pose formidable challenges that necessitate advanced approaches to minimize trial-and-error during experimentation.This work proposed a data-driven approach based on stacking ensemble learning to predict the mechanical properties of Ti6Al4V alloy fabricated by large-scale LPBF for the first time.This method can adapt to the complexity of large-scale LPBF data distribution and exhibits a more generalized predictive capability compared to base models.Specifically,the stacking model utilized artificial neural network(ANN),gradient boosting regressor,kernel ridge regression,and elastic net as base models,with the Lasso model serving as the meta-model.Bayesian optimization and cross-validation were utilized for model optimization and training based on a limited data set,resulting in higher predictive accuracy compared to traditional artificial neural network model.The statistical analysis of the ANN and stacking models indicates that the stacking model exhibits superior performance on the test set,with a coefficient of determination value of 0.944,mean absolute percentage error of 2.51%,and root mean squared error of 27.64,surpassing that of the ANN model.All statistical metrics demonstrate superiority over those obtained from the ANN model.These results confirm that by integrating the base models,the stacking model exhibits superior predictive stability compared to individual base models alone,thereby providing a reliable assessment approach for predicting the mechanical properties of metal parts fabricated by the LPBF process.