A hybrid stochastic-deterministic optimization method for waveform inversion

Present-day high quality 3D acquisition can give us lower frequencies and longer offsets with which to invert. However, the computational costs involved in handling this data explosion are tremendous. Therefore, recent developments in full-waveform inversion have been geared towards reducing the computational costs involved. A key aspect of several approaches that have been proposed is a dramatic reduction in the number of sources used in each iteration. A reduction in the number of sources directly translates to less PDE-solves and hence a lower computational cost. Recent attention has been drawn towards reducing the sources by randomly combining the sources in to a few supershots, but other strategies are also possible. In all cases, the full data misfit, which involves all the sequential sources, is replaced by a reduced misfit that is much cheaper to evaluate because it involves only a small number of sources (batchsize). The batchsize controls the accuracy with which the reduced misfit approximates the full misfit. The optimization of such an inaccurate, or noisy, misfit is the topic of stochastic optimization. In this paper, we propose an optimization strategy that borrows ideas from the field of stochastic optimization. The main idea is that in the early stage of the optimization, far from the true model, we do not need a very accurate misfit. The strategy consists of gradually increasing the batchsize as the iterations proceed. We test the proposed strategy on a synthetic dataset. We achieve a very reasonable inversion result at the cost of roughly 13 evaluations of the full misfit. We observe a speed-up of roughly a factor 20.