Data-Driven Analysis and Prediction for Ultrafast Laser Processing by Machine Learning: A Comparison Study

In recent years, laser surface processing has become an essential technique for steel materials processing. The precision and quality of laser processing outcomes are heavily influenced by various processing parameters, such as pulse mode, pulse frequency, defocusing amount, marking frequency, and pulse energy. However, conventional approaches including trial-and-error and experimental design, are inefficient and limited in capturing intricate correlations between laser parameters and resultant outcomes. Therefore, this study presents a data-driven framework for predicting and analyzing laser processing outcomes. We conducted a series of laser surface texturing experiments to collect data on laser parameters and their corresponding outcomes. The dataset includes random combinations of pulse energy, defocusing amount, pulse frequency, pulse mode, and marking frequency. These experiments served as the foundation for model training and correlation analysis. We employ Spearman’s rank correlation coefficient and maximal information coefficient to establish correlations between laser parameters and processing outcomes. By comparing multiple machine-learning models, we evaluate their performance in predicting processing outcomes. The results demonstrate that the data-driven framework effectively identifies key correlations between laser parameters and processing outcomes, and the comparison of machine learning models reveals that random forest performs best in predicting depth, support vector regression excels in predicting surface roughness, and multi-layer perception provides the most accurate predictions for diameter, offering valuable insights for selecting optimal laser parameters in industrial applications.