Machine learning-based prediction of microgroove geometry & surface roughness in femtosecond laser machining of SiCp/Al

Abstract

Femtosecond laser machining is a non-contact method for fabricating microstructures on difficult-to-machine materials like silicon carbide particle reinforced aluminium (SiCp/Al). However, the non-linear and complex relationship between laser processing parameters, microgroove geometry, and surface quality makes the parameter selection very challenging and often dependent on traditional trial-and-error experimentation. In this paper, different machine learning regression models are developed to predict the groove depth, groove width, and surface roughness of microgrooves using laser power, scanning speed, and defocus distance as input features. A dataset of 60 experiments is used to train and evaluate the Linear Regression, Random Forest, Support Vector Regression, XGBoost, and K-Nearest Neighbors models. The results show that nonlinear models outperform the linear regression, which indicates the presence of complex parameter-response relationships. SVR achieves the best overall performance with higher prediction accuracy and more stable generalization across all the output variables. The analysis indicates that the surface roughness is more predictable than the geometric features, while groove depth shows higher variability due to material heterogeneity and laser-material interactions. SHAP analysis identifies that the laser power is a dominant factor. However, a limited dataset constrains the generalization, and future research should focus on expanding these datasets and integrating the physics-informed machine learning models.