Massive computation of seismic traveltimes is widely used in seismic processing, for example, for the Kirchhoff migration of seismic and microseismic data. Implementation of the Kirchhoff migration operators uses large precomputed traveltime tables (for all sources, receivers, and densely sampled imaging points). We have tested the idea of using artificial neural networks for approximating these traveltime tables. The neural network has to be trained for each velocity model, but then the whole traveltime table can be compressed by several orders of magnitude (up to six orders) to the size of less than 1 MB. This makes it convenient to store, share, and use such approximations for processing large data volumes. We evaluate some aspects of choosing neural-network architecture, training procedure, and optimal hyperparameters. On synthetic tests, we find a reasonably accurate approximation of traveltimes by neural networks for various velocity models. A final synthetic test shows that using the neural-network traveltime approximation results in good accuracy of microseismic event localization (within the grid step) in the 3D case.

中文翻译：

---

使用人工神经网络的行程表压缩，用于微地震数据的基尔霍夫迁移处理

地震旅行时间的大规模计算广泛用于地震处理中，例如，用于地震和微地震数据的基尔霍夫偏移。基尔霍夫（Kirchhoff）迁移算子的实现使用大型的预先计算的旅行时间表（适用于所有源，接收器和密集采样的成像点）。我们已经测试了使用人工神经网络来近似这些行程表的想法。必须为每个速度模型训练神经网络，但是随后可以将整个行程表压缩几个数量级（最多六个数量级），以使其大小小于1 MB。这样可以方便地存储，共享和使用这些近似值来处理大数据量。我们评估选择神经网络架构，训练过程和最佳超参数的某些方面。在综合测试中，我们通过神经网络为各种速度模型找到了行进时间的合理准确的近似值。最终的综合测试表明，在3D情况下，使用神经网络的行进时间逼近可以很好地实现微地震事件定位的精确度（在网格步骤内）。