Inversion using adaptive physics-based neural network: Application to magnetotelluric inversion

The neural networks applied to geophysical inversion has had limited success due to the need for large training data sets and the lack of generalizability to out-of-sample scenarios. The deterministic regularized inversion often requires a good starting model to avoid possible local minima in highly nonlinear problems. We have developed an inversion-based neural network (IBNN) procedure that combines the advantage of deterministic and neural network inversions in a coupled inversion scheme. The new inversion algorithm is formulated as a constrained problem solved by minimizing an objective function composed of data misfit, neural network misfit, and a coupling model objective function that links the two inversion schemes through a reference model. We investigate two strategies to update the reference model in the coupling model objective function using either a fully-trained network or a progressively trained network that keeps evolving and learning. We extend the analysis of the progressively-trained network inversion without preparing a training data set, which is suitable for most exploration scenarios. We demonstrate the convergence of our procedure in recovering high-resolution resistivity models when applied to synthetic MT data.