DFT-assisted Raman spectroscopic characterization and machine learning classification of three rice types from the...

A DFT-assisted Raman framework integrating difference-derivative-PCA highly distinguishable Raman bands screening and difference visualization (DD-PCA-HSV), characteristic peak correlation analysis, and machine learning classification was developed for three milled rice types (SXJ-100, FYXZ, and BJR) from the Taihu Lake Basin. Five independent lots were analyzed for each type. Lots L1-L4 formed the development dataset, and L5 was reserved for independent testing. DFT calculations on simplified amylose and amylopectin models, together with literature, supported the assignment of 22 characteristic Raman peaks. Grouped four-fold cross-validation of DD-PCA-HSV on L1-L4 reproducibly retained six bands at 439, 576, 939, 1126, 1262, and 1460 cm^-1. Z scores of these bands characterized spectral differences among the three rice types in the development dataset, and the same relative trends were reproduced in L5. Differences between BJR and SXJ-100/FYXZ were mainly associated with pyranose-ring and glycosidic-linkage vibrations, whereas differences between SXJ-100 and FYXZ were mainly associated with CO and CCO stretching vibrations. Linear correlation analysis combined with a dual-threshold rule identified 1080/1106 and 524/615 cm^-1 as discriminative peak pairs. Subsampling analysis indicated that 30 spectra were sufficient for repeatable R² estimation. The corresponding linear correlation features were reproduced in L5. Among RF, DT, and SVM models, SVM using normalized spectra achieved the best independent test performance, with accuracy, macro-precision, macro-recall, and macro-F1 all reaching 0.98. Overall, the framework distinguished the investigated rice types while linking discriminative Raman features to molecular vibrations, thereby providing a chemically interpretable basis for Raman-based rice characterization.