Abstract
Baseline correction is a critical preprocessing step in Raman and surface-enhanced Raman spectroscopy analysis. The adaptive iterative reweighted penalized least-squares (airPLS) method is widely used due to its simplicity and efficiency, but its effectiveness is often hindered by challenges such as baseline smoothness, parameter sensitivity, and inconsistent performance under complex spectral conditions. To address these limitations, we developed an optimized airPLS algorithm (OP-airPLS) that systematically fine-tunes key parameters by using an adaptive grid search method. We further implemented a machine learning model to predict these parameters through spectral shape recognition. A data set of 6000 simulated spectra representing 12 spectral shapes (comprising three peak types and four baseline variations) was used for evaluation. On average, OP-airPLS achieved a percentage improvement (PI) of 96 ± 2%, with the maximum improvement reducing the mean absolute error (MAE) from 0.103 to 5.55 × 10(-4) (PI = 99.46 ± 0.06%) and the minimum improvement lowering the MAE from 0.061 to 5.68 × 10(-3) (PI = 91 ± 7%). The optimal parameters for each spectral shape were found to reside within a well-defined linear region in the parameter space. While OP-airPLS significantly improved enhanced baseline correction accuracy, it required substantial computational resources and relied on known true baselines. To overcome these constraints, a machine learning approach combining principal component analysis and random forest (PCA-RF) was developed to directly predict optimal parameters from input spectra. The PCA-RF model demonstrated robust performance and achieved an overall PI of 90 ± 10% while requiring only 0.038 s to process each spectrum. When this method is applied to real spectra, its baseline estimation performance is constrained by both the signal-to-noise ratio and the similarity of the spectral shape to the training data.