Inverting magnetotelluric responses in a three-dimensional earth using fast forward approximations based on artificial neural networks

The most computationally intensive step in 3D magnetotelluric (MT) inversion is the calculation of the forward response. This fact makes any modelling which requires many function evaluations, including genetic algorithms and Markov chain Monte Carlo inversion, extremely time consuming. Using Artificial Neural Networks (ANNs) it is possible to approximate these expensive forward functions with rapidly evaluated alternatives. Using a limited subset of resistivity models created in a simple parameterisation, this work is the first to apply ANNs to approximate the 3D MT forward function. Training data are generated with a compute time of two weeks, and after training the ANN is able to reproduce forward responses at arbitrary site locations with accuracy similar to the level of typical data errors. To evaluate the accuracy of the models, we show that these forward responses may be used to successfully invert MT data in an evolutionary framework. Examples are shown in both synthetic and real-world scenarios, and results are compared with those from traditional inversion algorithms. We conclude that the trained ANN inversion has a fraction of the run-time of a traditional inversion and is successful at modelling the space of its limited parameterisation.