Integrating neural networks in least-squares inversion of airborne time-domain electromagnetic data br

Airborne time-domain electromagnetic surveys produce extremely large data sets with thousands of line kilometers of data and millions of possible models to explain the data. Inversion of such data sets to obtain the resistivity structures of the subsurface is computationally intensive and involves calculation of a significant number of forward and derivative responses for solving the least-squares inverse problem. The flight altitude of the airbornesystem needs to be included in the modeling, which adds further complexity. We propose to integrate neural networks in a dampediterative least-squares inversion framework to expedite the inver-sion process. We train two separate neural networks to predict theforward responses and partial derivatives independently for abroad range of resistivity structures and flight altitudes. Data in-version is not only used for producing the final subsurface modelsbut also used during data processing, or to produce intermediateresults during a survey. With these purposes in mind, we providethree inversion schemes with a tunable balance between computa-tional time and modeling accuracy: (1) numerical forwardresponses used initially in combination with neural networkderivatives, and the derivatives switched to a numerical solutionin final iterations, (2) numerical forward responses in combina-tion with neural network derivatives used throughout theinversion, and (3) only neural network forward responses andderivatives used in inversion. Experiments on field data find thatwe improve inversion speed without any loss in modeling accu-racy with our first approach, whereas the second scheme gives asignificant speedup at the cost of minor and often acceptable de-viations in the inversion results from the conventional nonlinearinversion. The last approach is the fastest and captures the overallresistivity structures quite well. Therefore, depending on the modeling accuracy, inversion speedup factors of up to 50 are realized by using the proposed schemes