Neural Network-Enhanced FMCW Gas Spectroscopy

Abstract

Detecting multicomponent gases over extensive concentration ranges with laser spectroscopy faces challenges of complex configurations, intricate spectral analysis, and reduced accuracy. Neural networks offer transformative potential for advancing laser spectroscopy by facilitating real-time optimization and automation of experimental processes. Here, we report a frequency-modulated continuous-wave (FMCW) spectroscopic system enhanced by a feedforward neural network (FNN) algorithm. FMCW encodes gas absorption spectra into optical signals, enabling both spatial localization and spectroscopic analysis. By training the FNN for the target gas components, our approach can analyze broadband superposed spectra to determine gas concentrations in the mixture. In the proof-of-concept demonstration for the measurement of C2H2and CO2mixtures, the FNN achieves high accuracy in demodulating mixed gases, with residuals less than ± 2 ppm (100–900 ppm of C2H2) and ± 0.3% (80% to 96% CO2), respectively. The FNN outperforms traditional concentration inversion methods regarding linear dynamic response, attaining high-precision quantification across 5 orders of magnitude (R2> 0.99999). Our approach exhibits advantages of simplified signal processing and enhanced measurement accuracy, positioning it as a promising candidate for quasi-distributed sensing applications in environmental, medical, and industrial contexts.