Geoid Inversion For Mantle Viscosity With Convolutional Neural Networks



Journal Title

Journal ISSN

Volume Title



The Earth’s radial viscosity profile affects the extent to which density contrasts in the mantle manifest themselves in the long-wavelength shape of the non-hydrostatic geoid. Consequently, geodynamicists have used the observed geoid along with an estimate of the Earth’s density structure to invert for the viscous structure of the mantle. The inversion procedure, however, is one that has no direct solution, relies upon density estimates that have a high degree of uncertainty, and is computationally costly. In this study, we provide a first attempt at using machine learning as a low-cost solution to the geoid inverse problem. We carry out two separate solutions with convolutional neural networks (CNNs) and compare the results. The first solution uses two CNNs, where the first predicts viscous-response kernels from data characterizing the geoid and density structure of the Earth. The second network then predicts a radial viscosity profile from the viscous-response kernels. The second solution, instead, predicts a radial viscosity profile directly from the geoid and density data. We find that in the two-network solution, both CNNs make accurate predictions on the test data, but when predictions from the first network are fed into the second as input, the results are poor. We find that the single network solution allows us to obtain a smooth, long-wavelength estimate of the Earth’s radial viscosity profile.



Geophysics, Geodynamics, Machine learning, Convolutional neural networks, Inversion, Geoid, Viscosity, Mantle