AirNet: A Deep Learning-Driven Auto Baseline Correction Algorithm Balancing Global Smoothness and Local Fidelity

Baseline correction is a critical preprocessing step to eliminate non-Raman scattering backgrounds in the Raman/SERS spectrum, ensuring accurate peak positions and intensities for qualitative and quantitative analysis. Recently, various machine learning-based approaches have been proposed for automatic baseline correction. Nevertheless, their generalizability is constrained, since neither background physicochemical origins nor critical fitting parameters are fully understood. Therefore, we developed AirNet, an automated baseline correction algorithm that integrates deep learning with chemometrics to balance global smoothness and local fidelity. AirNet includes three steps: (1) Raman peaks and baselines are identified with initialized weights by a ResUnet-based model; (2) an optimal smoothing parameter is adaptively selected by a multi-indicator evaluation strategy; (3) a reliable baseline is achieved with refined weights by a dynamic and robust adaptive iteratively reweighted penalized least squares (Dr-airPLS) under optimized smoothness. On both simulated and experimental Raman spectra, AirNet outperforms algorithms like airPLS, OP-airPLS, and DIRAS in both accuracy and generalizability, with a speed of 0.3 s per spectrum. Furthermore, the Homologous Model of AirNet ensures consistent baseline correction across homologous spectra. The segmented fitting strategy applies a region-specific smoothing parameter, facing a Raman spectrum with drastically changed background gradients. AirNet extracts high-fidelity Raman spectral information, providing a solid basis for reliable spectrum-structure correlation.