Intelligent baseline correction for Raman spectroscopy via deep learning-based parameter optimization

초록

Raman spectroscopy is a widely used technique for the nondestructive analysis and identification of chemical substances. However, Raman spectral data acquired in real-world environments often exhibit baseline drift, necessitating correction for accurate interpretation. The conventional and representative signal processing method, Adaptive Iteratively Reweighted Penalized Least Squares (airPLS), achieves strong correction performance but requires manual tuning of multiple parameters, which limits its practicality. To address this issue, we propose airPLS-NN, a novel baseline correction system trained on synthetically generated datasets that integrates synthetic data generation and labeling, deep learning-based automatic parameter tuning, and robust correction strategies tailored to the characteristics of Raman spectral data. Experimental results on synthetic datasets demonstrate that our system significantly outperforms conventional approaches, reducing the Fréchet Inception Distance (FID) by 64.19% and the Euclidean Distance (ED) by 45.47%, while increasing the Pearson Correlation Coefficient (PCC) by 1.73% and the Cosine Similarity (CS) by 1.38%. Furthermore, inference experiments on real experimental Raman spectra confirm the system's robust generalization capability, delivering stable performance even in cases involving complex and unseen baseline shapes. Finally, the proposed system achieves superior computational efficiency compared to other neural network-based methods, making it suitable for deployment in resource-constrained environments. © 2026 Elsevier B.V.

키워드