Physics-Informed Machine Learning Overcomes Particle-Size Nonlinearities That Confound the Raman Spectrometry of Granular Carbonaceous Materials

Abstract Raman spectroscopy supports multivariate regression models that can offer a rapid and nondestructive approach to quantifying the composition of granular materials. Soils constitute a class of such materials for which chemical analysis provides vital information for agriculture and environmental stewardship. However, soil samples typically present highly heterogeneous mixtures of varying particle size and diverse optical properties owing to light-absorbing constituents. Although particle-size effects on Raman spectrometry in scattering media have long been considered in the context of pharmaceutical analytical technology, the quantitative influence of particle size in opaque, absorbing materials remains poorly understood. Here, we investigate the effect of particle size and light absorption in controlled mixture systems that mimic soils. These mixtures, containing particles of varying sizes and optical properties, isolate the ways in which fluorescence, granularity and light absorption affect the accuracy of machine learning regression models. Using a dual-laser Raman system, we collected both conventional Raman and Shifted-Excitation Raman Difference Spectroscopy (SERDS) spectra, showing that light-absorbing compounds attenuate Raman intensities nonlinearly to an extent that varies with particle size. Multivariate regression models link this behavior to optical extinction governed by the surface area of the absorbing grains. We account for this behavior in terms of a simple physical model that incorporates the physics of light scattering and absorption in heterogeneous systems of varying particle size. This hybrid, physics-informed machine learning framework outperforms purely data-driven approaches in predicting sample composition from Raman spectra, highlighting the importance of physical characteristics in the analysis of light-absorbing granular materials.