Full waveform inversion with random shot selection using adaptive gradient descent

Full waveform inversion (FWI) is a powerful yet computationally expensive technique that can yield subsurface models at high resolution. Randomly selected shots ("mini-batches") can be used to approximate the misfit and the gradient of FWI, thereby reducing its computational cost. Here, we present a methodology to perform mini-batch FWI using the Adam algorithm, an adaptive optimization scheme based on stochastic gradient descent. It provides for stable model updates by smoothing the gradient across iterations and can also account for the curvature of the optimization landscape. We describe empirical criteria to choose the hyperparameters of the Adam algorithm and the optimal mini-batch size. The performance of the outlined scheme is illustrated on synthetic data from the Marmousi model.