In-line spectroscopic scatterometry-based monitoring with data-driven analysis for high-quality laser surface texturing

Real-time monitoring of laser surface texturing is essential for process robustness by detecting deviations in surface topography. This study presents a compact coaxial optical monitoring system based on spectroscopic scatterometry to identify such deviations caused by processing disturbances. The system records zero-order reflected spectra from laser-fabricated textures, correlating spectral features with variations in processing conditions. Two texturing methods are investigated: Laser-Induced Periodic Surface Structures (LIPSS), producing sub-micron ripples, and Direct Laser Interference Patterning (DLIP), generating hierarchical micro/nano patterns. Reference spectra acquired under controlled variations in Focal offset Distance (FoD) are used to train machine learning (ML) classifiers to detect process deviations. Decision tree models were chosen for their low latency and interpretability, with SHapley Additive exPlanations (SHAP) used to identify key wavelengths for model compression, enabling faster inference without compromising accuracy. The pruned models achieve classification accuracies of over 99% on both LIPSS and DLIP textures and enable near an order of magnitude improvements in inference speed. Beyond the classification capabilities, the same approach was extended to predict the texture height of micropillar structures using regression models. High prediction accuracies (R² > 0.97) were achieved, showing the potential of spectral data not only for detecting process deviations but also for quantitative geometry estimations. This approach enables rapid, in-axis detection of processing disturbances together with their effects on multiscale surface topographies without requiring complex inverse modelling or surface reconstructions. Further, it offers a scalable and real-time solution for intelligent control in laser texturing of freeform surfaces. • A coaxial spectroscopic scatterometry approach is demonstrated for in-line detection of processing disturbances in LIPSS and DLIP. • Machine learning-driven models detect focal offset deviations with >99% accuracy using a reduced set of wavelengths. • SHAP-guided feature reduction achieves sub-millisecond inference suitable for embedded deployment. • Spectral signatures enable non-contact prediction of texture height without explicit inverse models.