Overcoming Correlation Degradation in Photoacoustic Spectroscopy: High-Precision Gas Concentration Prediction With a GA-Optimized Dual Stacking Architecture

Systematic residuals induced by correlation degradation represent a critical bottleneck in gas concentration prediction using photoacoustic spectroscopy (PAS). Conventional models, particularly those with single-layer architectures, lack an explicit residual correction mechanism, making it challenging to mitigate these systematic biases. To address this challenge, a multilayer perceptron model, termed GA-DSA-MLP, is proposed by integrating a Dual Stacking Architecture (DSA) with Genetic Algorithm (GA) optimization. Through a two-stage mapping mechanism, the model achieves explicit modeling and rectification of systematic residuals, effectively alleviating the limitations of traditional single-layer networks in feature extraction and error self-correction. Complementing this architecture, the global search capability of the GA provides robust parameter initialization to enhance feature extraction performance under severe correlation degradation. Experimental results demonstrate that the GA-DSA-MLP surpasses benchmarks like OLS and PSO-EAP-CNN by over 70% in key error metrics. The model requires approximately 17 ms for single inference, with 89.3% of samples exhibiting a relative error below 0.5%. Overall, this work presents a novel lightweight end-to-end model that, to our knowledge, is the first to incorporate an explicit residual correction mechanism specifically for addressing PAS correlation degradation. It provides a reproducible and effective framework for high-precision gas concentration prediction.