Neural Network Architectures for Absorption Spectroscopy

Abstract

Abstract State-of-the-art neural network architectures in image classification and natural language processing were applied to absorption spectroscopy applications by interpreting the data structure accordingly. A model was designed for temporal interpolation of background spectra and a different model was created for gas concentration fitting. The networks were trained on experimental data provided by a wavelength modulation spectroscopy instrument and the best performing architectures were analyzed further to evaluate generalization performance, robustness and transferability. A BERT-styled fitter achieved the best performance on the validation set and reduced the mean squared error of fitted amplitude by 99.5 %. A U-Net styled convolutional neural network reduced the mean squared error of the interpolation by 93.2 %. Evaluation on a test set provided evidence that the combination of model interpolation and linear fitting was robust and the detection limit was improved by 52.4 %. Transferring the trained models to a different spectrometer setup was tested and showed no chaotic out-of-distribution effects. Additional fine-tuning further helped increasing the performance of the transferred model. Overall the proposed model architectures can be applied to spectroscopy tasks if the data structure is interpreted the right way and the pre-trained networks are robust and can be transferred to other spectrometer setups.