PAQC: A hybrid-intelligence framework for automated quality control of Raman spectroscopy via physics-aware and deep feature fusion

Raman spectroscopy is a critical analytical technique across numerous engineering disciplines; however, the reliability of its measurements is often compromised by artifacts stemming from complex instrumental and sample-specific variations. Automating quality assessment poses a significant challenge: conventional physics-based thresholds, though interpretable, lack the flexibility to accommodate diverse anomalies, while data-driven deep learning approaches typically overlook valuable domain knowledge. To address this gap, we propose PAQC—a hybrid-intelligence method that synergistically integrates explicit physical priors with implicit representation learning. Our approach processes unlabeled spectral data through dual parallel streams: an explicit knowledge stream that computes physically-informed quality metrics to produce high-confidence pseudo-labels, and an implicit knowledge stream that uses a peak-attention autoencoder, guided by known peak locations, to extract discriminative deep features. These two streams are cohesively fused through a progressive contrastive learning network, yielding a highly separable feature space tailored for anomaly detection. For practical deployment, a Mahalanobis distance-based classifier enables real-time quality diagnosis of individual spectra. Evaluated on a real-world dataset using a rigorous 5-fold cross-validation protocol, PAQC achieves state-of-the-art performance with an F1-score of 98.98% ± 0.67%, while critically maintaining a perfect 100.00% recall, underscoring its effectiveness as a robust and scalable solution for automated quality control in knowledge-sensitive engineering applications.