Learning Operating Regimes from Experimental Data in Optical Feedback Systems

Interferometers based on self mixing (SM) in semiconductor lasers [1] can provide compact and sensitive readouts for future sensing applications, but their signals are sensitive to alignment, feedback strength, and speckle decorrelation. As a result, conventional machine learning (ML) pipelines must be used more strategically when the system is not perfectly calibrated, as the experiments are naturally open to external perturbations, alignment shifts, and other uncontrolled effects. In this work, we show that the operating regimes of an optical feedback (OF) setup can be discovered directly from experimental SM signals, and this regime identification is more critical than classifier complexity. We convert each SM waveform into a low dimensional feature vector using t-distributed stochastic neighbor embedding (t-SNE). Density-Based Spatial Clustering of Applications with Noise (DBSCAN) clustering is then used to find stable, high density regions while discarding low density, unstable measurements. Classical classifiers and a compact convolutional neural network (CNN) were trained on datasets cleaned with DBSCAN. In a discrimination task involving five letters (U, B, A, R, I), the optimal configuration achieved macro F1 scores of 0.87 while removing only 14 percent of the data. This suggests that the primary issue is unstable operating regimes, not the model's capability, pointing to a path for robust optical diagnostics that focus on distinguishing a few tissue or cell classes. Overall, the results indicate that, for the investigated OF dataset, identifying and filtering unstable operating regimes has a larger impact on classification reliability than increasing model complexity alone. The contribution of this work is therefore not a new learning algorithm, but an application oriented workflow showing how unsupervised manifold analysis can be used as a data driven regime identification step before supervised classification. This strategy is relevant for future OF based sensing scenarios, including biomedical diagnostics [2], where the task may involve separating a limited number of tissue or cell states under imperfect calibration.