Deep learning-assisted Raman spectroscopic quantification of total terpenes in Curcuma kwangsiensis

A rapid and efficient strategy was developed to quantify the total terpene content of Curcuma kwangsiensis by combining Raman spectroscopy with a Residual-Squeeze-and-Excitation one-dimensional convolutional neural network (Res-SE 1D-CNN). A dataset of 1400 spectra obtained from 26 batches across 6 production areas in Guangxi was constructed. Major terpenoids were determined by HPLC, and UV-based total terpene content served as the reference. Density functional theory calculations were performed to verify the theoretical Raman peaks of Germacrone and support spectral band assignments. Compared with Partial Least Squares Regression, Support Vector Machine, and Backpropagation Neural Network models, the proposed Res-SE 1D-CNN achieved superior predictive accuracy (R² = 0.998, RMSE = 0.360). This study establishes a robust quantitative tool for evaluating complex traditional Chinese medicine systems and provides practical support for geographical traceability and rapid on-site quality control, with potential for extension to multi-marker natural product analysis.