Laser spectral enhancement and analysis based on blind-spot networks and Kolmogorov-Arnold networks

Laser-induced breakdown spectroscopy (LIBS) has garnered significant attention in fields such as materials science and environmental monitoring, owing to its rapid, non-destructive, and multi-element detection capabilities. However, the widespread adoption of LIBS is hindered by two major limitations: spectral noise interference and limited accuracy in quantitative analysis. To address these challenges, we propose a self-supervised deep learning noise reduction framework (Blind-Spot Spectral Denoising Network-BSSDN) based on a Blind-Spot Networks, along with a quantitative analysis model leveraging the Kolmogorov-Arnold Networks (KANs) architecture, to significantly enhance the analytical performance of LIBS. BSSDN eliminates the need for clean reference data by employing a central-blind-spot convolutional mechanism combined with a dynamic masking strategy. It reduces the mean spectral relative standard deviation (RSD) from 73.8 % to 23.9 % across six certified stainless-steel reference materials (JZG201-JZG206B, 36 × 36 mm solid blocks, supplied by the Shanghai Institute of Metrology and Testing), while preserving 90 % of characteristic peaks and maintaining excellent intensity fidelity. For quantitative analysis, our model addresses the limitations of traditional Multi-Layer Perceptron (MLPs) through adaptive spline basis functions and localized weighting. It achieves a mean coefficient of determination (R²) of 0.978 for nine elements and reduces the root mean square error of prediction (RMSEP) by over 45 %. Systematic comparisons highlight the framework's superiority over conventional methods in noise suppression, signal fidelity, and prediction accuracy. This work establishes a robust framework for high-precision LIBS applications in industrial inspection and trace analysis, while pioneering a approach to intelligent spectral data processing using self-supervised learning.