Diffusion Model-Driven Spectral Noise Modeling and Generative Reconstruction Strategy

Abstract

Modern spectroscopic techniques struggle with quantitative accuracy in complex environments due to unavoidable noise and matrix interference. While machine learning (ML) can improve performance, its reliance on massive data sets contradicts the few-shot reality of industrial applications, where limited data exacerbates noise-induced overfitting. To address this challenge, this paper proposes the universal spectral noise modeling and generative reconstruction (SNMGR) strategy. This strategy builds upon diffusion probabilistic models to effectively suppress strong interference effects in few-shot scenarios through the deep integration of noise interference modeling and generative reconstruction. To verify the effectiveness of SNMGR, this study conducts experiments using the quantitative analysis of lithium ore via femtosecond LIBS as a typical case. The research constructs the generative model based solely on 7 calibration samples totaling 210 spectra and systematically compares the quantitative performance of five mainstream ML models with distinct architectures on 3 unknown test samples. The results indicate that the integration with SNMGR enables the quantitative prediction accuracy of all five models to achieve significant and consistent improvements over their baseline counterparts. Taking the CNN-SNMGR model as an example, its predictive mean absolute error drops substantially from 1.801 ppk to 0.611 ppk representing a decrease of 66.06%, while the root-mean-square error decreases from 2.226 ppk to 0.778 ppk representing a decrease of 65.05%. This study demonstrates that SNMGR provides a viable solution for spectral interference correction under resource-constrained conditions. Furthermore, its universality in coupling with various ML models opens a new pathway for the large-scale implementation of spectral quantitative technologies in complex industrial environments.