Analysis of Dynamic EPR Spectra of pH-Sensitive Nitroxides Using Machine Learning

This research presents the application and comparison of the most popular machine learning algorithms (multilayer perceptron, ensemble of networks, convolutional neural network) to predict the shape of dynamic EPR signals of pH-sensitive nitroxide radicals (NR) with varying degrees of rotational correlation from EPR spectra obtained during the analysis of solid-phase systems. A training sample example consisting of theoretically modelled EPR spectra of pH-sensitive nitroxide radicals was prepared for the network learning. The result of applying and comparing machine learning algorithms with results obtained from simulations by the specialized software ODFR4 demonstrated that the multilayer perceptron and ensemble of networks have good accuracy in prediction signal shapes, however, within rather narrow scope of applicability. The convolutional neural network showed the worst results in predicting signal shapes. The results of this work can be useful for the development of software design to quickly estimate the parameters of analyzed systems by processing the EPR spectra of pH-sensitive nitroxide radicals.