Practical deep learning inversion using neural architecture search and a flexible training dataset generator

Summary Deep learning has the potential to estimate velocity models directly from shot gathers, which would reduce the turn-around time of seismic inversion. Our study addresses two challenges in implementing deep learning techniques for seismic inversion: the practical generation of a large amount of training data and the search for the best neural network architecture. First, we propose a flexible system which parametrically generates velocity models to create a large-scale, complex and fully synthetic training dataset, without using a target subsurface model. Using this system, we created 300,000 synthetic velocity models for our experiments. Second, we employ neural architecture search techniques to design a suitable neural network using Optuna, an automatic hyperparameter optimisation framework. We incorporated the residual network into an encoder–decoder model and optimised its architecture. Thus, we obtained an optimal neural network model consisting of more than 100 hidden layers. We evaluated our model using the Marmousi2 model and the 1994 Amoco statics test dataset. The model demonstrated comprehensible estimations of the benchmark velocity models.