Spatially-resolved fluorescence (SRF) and deep-learning algorithms for the measurement of agents’ concentrations

Abstract Fluorescence is a non-destructive, rapid, and cost-effective technique widely employed for detecting chemical and biological agents. However, its performance in complex mixtures is often hindered by spectral overlap and limited specificity. A novel approach of spatially-resolved fluorescence (SRF) is introduced to overcome these issues. By acquiring fluorescence spectra as a function of spatial position under multi-wavelength excitation, SRF simultaneously captures both emission and absorption characteristics, enhancing the informational content of the measurement. The technique is first developed through a theoretical model, which outlines the conditions under which spatially-resolved data improve the accuracy of concentration estimation. This is followed by experimental validation using aqueous mixtures of tyrosine and tryptophan, from which spatial fluorescence maps were obtained for testing purposes. A convolutional neural network (CNN) is applied to solve the inverse problem of determining agent concentrations from SRF data, even under non-linear conditions (such as the concentrations’ non-linearity phenomenon). The model is trained entirely on synthetic datasets generated from a calibrated numerical simulation based on reference spectra. Experimental maps are used solely to test model performance. Despite the lack of experimental data in training, the CNN achieves high accuracy, with coefficients of determination ( R 2 ) exceeding 9x% in most cases. The SRF-CNN framework demonstrates strong potential for accurate, label-free quantification in complex samples. Future enhancements may include time-resolved measurements or broader excitation wavelength ranges to increase analytical power.