Variable selection plays a central role in spectroscopic calibration. However, most existing methods treat it as a purely data-driven optimization task, without explicitly incorporating the physicochemical mechanisms of spectral responses. Herein, we propose a physics-informed spectral-structure synergy optimization (SSSO) method that integrates characteristic spectral lineshapes (CSLs) with structured synergy effects to enhance consistency and interpretability. First, a physics-informed sparse Bayesian dictionary learning strategy is proposed to explicitly model CSLs using a sparse Gaussian dictionary, while structure-aware priors are employed to characterize the intrinsic properties of distinct spectral components. Variational Bayesian inference (VBI) is then applied to obtain approximate posterior distributions. Based on the solutions, the full spectrum is decomposed into chemically meaningful peak structures, thereby achieving adaptive nonuniform spectral segmentation. To further exploit synergistic effects among these structures, a structure-based bootstrap sampling strategy is introduced. This strategy generates diverse structural combinations and iteratively compresses the number of retained structures based on predictive performance, ultimately selecting the optimal synergistic structural combination. Experimental results demonstrate that SSSO achieves superior predictive performance while ensuring physicochemical interpretability, with the selected variables consistently aligning with the chemical bonds of the target analytes.
Wu et al. (Fri,) studied this question.