Abstract

Laser absorption spectroscopy enables non-intrusive gas detection, but its performance degrades in multi-species environments with spectral overlap, noise, and unknown interferents. This thesis develops machine learning–enhanced diagnostics to address these challenges across combustion and environmental sensing. Deep denoising autoencoders were applied to shock tube data to recover low-SNR absorbance signals during pyrolysis of hydrocarbon mixtures. An unsupervised spectral decomposition framework, HT-SIMNet, was introduced to isolate species features without requiring labeled data, while a blind source separation method, UnblindMix, reconstructed both species concentrations and spectral profiles directly from complex mixtures. Feature transformations based on spectral derivatives improved detectability of weak absorbers, and a randomized smoothing–based classifier, VOC-certifire, provided robust and certifiable identification of VOCs under perturbations. Together, these tools enable real-time, reference-free, and interference-resilient sensing across diverse gas-phase systems.  

Biography