Аннотация:The inverse problem of magnetotelluric sounding is a highly non-linear ill-posed inverse problem with high dimension both at the input and at the output. One way to reduce the incorrectness is to narrow the scope of the problem. In our case, this can be implemented in the form of a complex algorithm, which first makes the choice of one of the narrower classes of geological sections and then performs the solution of the regression inverse problem within the selected class. In the present study, we investigate the effectiveness of the implementation of the first phase of this algorithm. The neural network solution of the problem of classification of magnetotelluric sounding data was considered. We estimate the maximum accuracy of classification, perform search for optimal parameters, and test the results for resilience to noise in the data.