Machine Learning-Assisted Design Optimization of T-type Photoacoustic Cell for Enhanced Acetylene Detection in Photoacoustic Spectroscopy

A truncated cone-double petal T-type photoacoustic cell (TCDPT-PAC) was proposed for acetylene (C₂H₂) detection in this paper. The Gradient Boosting Decision Tree (GBDT) model with Non-dominated Sorting Genetic Algorithm (NSGA-II) was used to optimize the dimension of TCDPT-PAC. The frequency responses and acoustic field distribution of the TCDPT-PAC were simulated with the finite element analysis (FEA) method. Acoustic pressure and quality factor (Q-factor) of the TCDPT-PAC were significantly enhanced after its dimensional optimization. The mean relative errors (MREs) between the evaluation predicted by the GDBT-NSGA-II model and those obtained from finite element simulations for the corresponding size parameters were as low as 0.27% for acoustic pressure and 0.70% for the Q-factor. The sensitivity of the TCDPT-PAC was 20.48 pm/ppm, which was 2.95 times that of the conventional H-PAC (6.94 pm/ppm) under the same optical path. For Allan deviation analysis, with an integration time of 100 seconds, the minimum detection limit (MDL) for the TCDPT-PAC was 2.93 ppb. The machine learning-assisted method rapidly optimized the novel TCDPT-PAC and balanced the limitations of the conflict between acoustic pressure and the Q-factor, providing a new approach for the development of high-performance PACs.