Modular Neural Networks for Seismic Tomography

We propose in this paper a modular approach for the problem of traveltime inversion or seismic tomography. This problem consists in the inference of the velocity of wave propagation in the subsurface after an explosion has been produced at the surface, relying on such waves’ traveltimes. These traveltimes are recorded by several receivers on the surface. In the present work, we consider data synthetically generated, thanks to the use of a particular "Earth-Model". An Earth-model is a multilayered media in which each layer is homogeneous, that is, the seismic wave’s propagation velocity in each layer is constant, and each layer's thickness is different. We compare, on these synthetic data, a Multilayer Perceptron (MLP) to a modular neural architecture. We show that the modular approach is better suited for the inversion problem stated, and study the experimental conditions in which the potential of this approach is optimally exploited.