Direct Laser Interference Patterning (DLIP) is an advanced, high-precision technique for large-scale surface functionalization [1]. In addition, effective integration of DLIP into manufacturing environments demands reliable real-time monitoring of process stability. Plasma emission signals, containing rich information regarding laser–material interaction, offer a promising basis for in-situ process evaluation. Autoencoders (AEs), a class of neural network-based models frequently used in signal processing, enable anomaly detection through efficient feature extraction and signal reconstruction [2]. This preliminary study presents an AE-based approach to classify plasma signals collected during the picosecond DLIP on stainless steel. Texturing was conducted using the ELIPSYS® system (SurFunction GmbH) with a picosecond laser (12 ps, 1064 nm). Plasma emission signals were captured using a photodiode-based sensor (Figure 1a). The signals were segmented, windowed and transformed via Fast Fourier Transform (FFT) before serving as input for AE model (Figure 1b). The model was trained on signals labeled as “healthy” via prior statistical clustering. Classification into “healthy” or “non-healthy” was based on reconstruction error and validated using surface topography measurements. The Receiver Operating Characteristic (ROC) curve showed a strong classification capability of the AE model, with an AUC of 0.80 indicating reliable discrimination between classes (Figure 1c). Finally, different challenges related to model generalization and interpretability in the context of data-driven process control for laser-based manufacturing are discussed.