| This work primarily investigates use of the neural network(NN) method to analyze spectral data collected in single molecule detection(SMD) and identification (SMI) experiments. The 2-layer neural networks, with sigmoid as the activation function, are constructed and trained on a set of simulated data using back-propagation and the delta-learning rule. The trained networks are then used for identification of photon bursts in subsequent simulations. Results show that the NN method yields better identification of individual photon bursts than the traditional maximum likelihood estimation (MLE), particularly in cases where the fluorophores have disparate fluorescence quantum efficiencies, absorption cross-sections, or photodegradation efficiencies.;In addition, this work reports several improvements over the prior version of the Monte Carlo simulation program. The improved version considers the fluorescence probability as the convolution of the pure exponential decay function characterized by the fluorescence lifetime and the instrument impulse response function in the experiment. The setting of the time window is then implemented by monitoring the variation of signal and noise. A number of problems have been investigated by using the improved version. In particular, the effects of the number and widths of the bins within the time window on the precision of identification of molecules are studied. The results from the improved version of the simulation show that only a small number of bins (4--8) are required to achieve approximately 90% correct predictions with the NN method. Bin widths chosen in accordance with the intuitive algorithm, or equal bin widths, generally give better predictions.;Experimental improvements are also reported in this work. In particular, the transit time of BODIBY-TR(D-6116) dye molecules in an SMD experiment was improved to less than 200 mus, and a circuit is implemented to accomplish fast and continuous data collection to be used in future single molecule identification experiments. |
