Data-driven local linearization for nonlinear computational spectroscopy

Accurate spectral reconstruction from nonlinear photodetectors remains a fundamental challenge for miniaturized computational spectrometers because global inverse models cannot adequately represent strongly nonlinear device responses. Here, we introduce a piecewise-linear machine-learning framework for nonlinear optical spectral sensing using photocurrent data acquired from a GeSe/InSe heterojunction device. The method partitions a photocurrent response space into locally linear regimes and reconstructs spectra using linear regression within each regime. We utilize two grouping strategies: supervised logistic-regression classification when light-source labels are available and an unsupervised ensemble of hierarchical clustering models when labels are unavailable. Applied to experimentally measured photocurrent vectors, both approaches enable accurate recovery of the incident power spectra despite the globally nonlinear device behavior. The clustering-based strategy provides the most consistent performance, with 72% of cases achieving log average normalized mean-squared error less than −1 (NMSE < 0.37), compared with 44% for the classification-based approach (median Avg NMSE of 0.12 vs 1.26, non-overlapping 95% CIs). These results show that nonlinear detector responses can be effectively linearized in a data-driven manner, establishing local grouping combined with linear inversion as a practical and explainable route for computational spectroscopy with compact tunable optoelectronic devices.