Аннотация:Standard methods for training MLP neural networks solve small-dimensional geophysical inverse problems. For nonlinear problems with N~100 or more, convolutional neural networks trained with deep learning are required. A custom CNN with specialized transformations (e.g., data compression) is proposed, enabling solutions for N~10^3 without initial approximation. The inversion process takes seconds and is independent of problem dimensionality (2D/3D). Numerical tests on synthetic and field data confirm the efficiency of the approach.