Physics-regularized simulation-based inference for quantitative near-infrared spectroscopy

Quantitative near-infrared spectroscopy often fails outside narrow calibration regimes because spectral features and nuisance factors co-vary across measurement conditions. We introduce a physics-informed simulation-based inference framework that decomposes NIR spectra into three components: an analyte-specific signature extracted via constrained multivariate curve resolution, instrument drift characterized from reference measurements, and empirically estimated correlated noise. These components drive a generative model that synthesizes training data, which is used to train neural networks to predict analyte concentrations without direct training on labeled experimental samples. On lactate quantification in phosphate-buffered saline and fetal bovine serum across multiple sessions with uncontrolled temperature and pH variation, the method achieves a mean absolute error of 0.37 mM—comparable to commercial point-of-care analyzers. The framework enables sample efficient quantitative spectroscopy by leveraging physics-informed spectral decomposition and realistic noise modeling rather than exhaustive experimental calibration.