Toward real-time anomaly detection in direct laser interference patterning using autoencoder analysis of plasma emission signals

Direct laser interference patterning (DLIP) is a high-precision surface engineering technique with strong potential for scalable industrial surface functionalization. However, ensuring consistent pattern quality requires monitoring strategies that are sensitive to subtle process deviations. This study introduces a lightweight autoencoder (AE)-based framework for morphology-referenced anomaly detection in DLIP using time-resolved plasma emission signals acquired with an off-axis photodiode setup. Surface morphologies are characterized by confocal microscopy and scanning electron microscopy, and a Gini coefficient-based homogeneity index is used to generate binary reference labels—“healthy” and “non-healthy”—for signal classification. The signals are segmented into local analysis windows, transformed into frequency-domain spectra, linked to these morphology labels, and used to train the AE only on spectra from “healthy” processing conditions. Three reconstruction-error metrics, namely Euclidean distance, Manhattan distance, and Mahalanobis distance (MahaD), are systematically compared. Among them, MahaD achieves the best performance, with 91% classification accuracy and 95% sensitivity for detecting nonhomogeneous surface formation, reflecting the benefit of covariance-aware scoring for correlated spectral features. The proposed framework demonstrates that compact photodiode-based plasma emission monitoring, Gini-based surface labeling, and lightweight AE modeling can be combined to assess DLIP surface quality from process signals. As a proof-of-concept study, it establishes a methodological foundation for future inline plasma-based process monitoring and for closed-loop quality control strategies after broader validation in industrial applications.