Quantitative Analysis of Multicomponent Iron-Based Compounds by Combining Raman Spectroscopy with Deep Learning

Quantitative identification of iron-based corrosion products is crucial for elucidating phase evolution and reaction pathways in complex oxidation environments. However, Raman-based compositional analysis remains challenging due to severe spectral overlap and nonlinear interactions among components in multicomponent systems. In this work, a data-driven framework is developed for the quantitative analysis of iron oxide mixtures using Raman spectroscopy. Representative corrosion-related phases, including α-FeOOH, α-Fe2O3, and Fe3O4, were systematically characterized to establish their vibrational fingerprints. A series of 27 mixtures with controlled compositions was prepared, and the data set was expanded to 2700 spectra through data augmentation to improve model robustness. A deep learning model based on ResNet18 was implemented to capture complex spectral correlations and nonlinear mixing behavior, and its performance was benchmarked against convolutional neural network (CNN) and partial least-squares (PLS) approaches. The ResNet18 model achieves superior predictive accuracy, with a coefficient of determination (R2) of 0.9627 and a mean absolute error (MAE) of 3.18%. The incorporation of residual connections mitigates gradient degradation, enabling stable training and improved generalization. The absence of systematic bias in residuals further confirms the reliability of the model. These results demonstrate an effective strategy for resolving overlapping vibrational features and provide a generalizable approach for quantitative analysis of complex multicomponent Raman systems.