Problems of indirect measurements arising in experimental science belong to the class of inverse problems (IP), which are often ill-posed, non-linear, and have high dimensionality. Machine learning methods are used to solve IP due to their ability to resist these unfavorable properties. In this study, we solve an IP of determination of concentrations of components in multi-component solutions by their Raman spectra. The results demonstrated by various machine learning methods are compared by the solution error and by their resilience to various types of noise encountered in experimental spectroscopy. The best results were demonstrated by multi-layer perceptrons with two hidden layers and by convolutional neural networks with one convolutional layer. For multiplicative noise with high noise level, the most noise resilient algorithms were random forest and gradient boosting.
|