Adaptive Physics-Aware Raman Baseline Correction with Machine Learning Predicted Parameters

Accurate baseline correction is critical for reliable Raman spectral interpretation. Traditional algorithmic methods often require manual tuning of regularization parameters, while recent machine learning and neural network approaches automate correction but lack generalizability and user control. We have developed a new approach to baseline correction which adaptively resolves baseline distortions without manual intervention - DIRAS (Dynamic Iterative Reweighted Autoregressive Spectral baseline correction). DIRAS uses a fixed regularization parameter (λ), which performs robust batch correction by iteratively reweighting residuals. We further used Structural Similarity Index Measure (SSIM) as an objective for λ optimization and trained a deep learning model to learn the nonlinear mapping between raw spectral features and optimal regularization. The resulting framework (DIRAS+) was capable of real-time spectrum-specific λ prediction. Applied to two SERS data sets, DIRAS+ outperformed ALS and SEALS in preserving peak fidelity, reducing intraclass variability and minimizing baseline distortion. Importantly, in downstream chemometric workflows, DIRAS improved calibration and model performance, yielding lower errors and enhancing analytical sensitivity. DIRAS and DIRAS+ together provide robust, scalable, and user-adaptable solutions for high-throughput Raman spectroscopy applications.