Application and integration of deep learning in FAIMS for identifying acetone concentration

In practical applications, analytical instruments are used for both qualitative and quantitative analysis. However, for high-field asymmetric-waveform ion mobility spectrometry (FAIMS), most studies to date have been focused on the qualitative analysis of substances, with limited research on quantitative analysis. Explored here is the feasibility of using deep learning in FAIMS for quantitative analysis, aided by redesigning the FAIMS upper computer. Integrating spectrum creation and deep learning analysis into the FAIMS upper computer boosts the processing and analysis of FAIMS data, laying a foundation for applying FAIMS practically. For analysis using image processing, multiple FAIMS spectral lines obtained under different conditions are converted into a three-dimensional thermodynamic map known as a FAIMS spectrum, and multiple FAIMS spectrum are preprocessed to obtain the data set of this experiment. The principles of partial-least-squares regression and the XGBoost and ResNeXt models are introduced in detail, and the data are analyzed using these models, while exploring the effects of different model parameters and determining their optimal values. The experimental results show that the pre-trained ResNeXt deep learning model performs the best on the test set, with a root mean square error of 0.86 mg/mL, indicating the potential of deep learning in realizing quantitative analysis of substances in FAIMS.