Machine learning-based prediction of roughness in additively manufactured Ti-6AL-4V parts

Abstract

Traditional trial-and-error methods for optimizing surface quality in additive manufacturing (AM) are inefficient and costly due to complex, non-linear process dynamics. This study investigates the prediction of surface roughness in AM Ti-6Al 4V parts using machine learning (ML) models. A dataset was constructed by varying key process parameters (laser power, scanning speed, and scanning interval) during sample fabrication. Following empirical surface roughness measurements, predictive models including BPNN, RBF, CNN, RF, SVM and LSTM were evaluated. The three best-performing baseline models (BPNN, CNN, and SVM) were subsequently coupled with PSO and GWO to further enhance predictive accuracy. Results indicate that among the baseline models, CNN performed best. Optimization significantly improved model stability and accuracy. Specifically, on the training and test sets respectively, the optimal hybrid model (GWO-CNN) achieved R² values of 0.9716 and 0.9302, MAE of 0.5077 and 0.7604, and RMSE of 0.6183 and 0.9429. This demonstrates that integrating ML with optimization provides a precise, efficient approach to controlling AM surface roughness.