Sesimic Traveltime Tomography Using Deep Learning

Seismic inversion methods performed by a deep neural network trained in a supervised learning manner have shown successful inversion performance in synthetic data examples that target small areas. These deep-learning-based seismic inversions use time-domain wavefields as input data and subsurface velocity models as output data. Since the time-domain wavefields include both traveltimes and amplitudes of seismograms, the size of the input data is considerably large. Therefore, studies that apply deep-learning-based seismic inversions trained on large amounts of field-scale data have not yet been conducted. In this study, to apply the deep-learning-based seismic inversion technique to field-scale data, the velocity models are predicted using only traveltimes of seismic waves as the input data instead of the full time-domain wavefields. If the traveltime information is used as input data, the resolution of the inversion result is diminished, but the data size is significantly decreased, which can reduce GPU memory usage and speed up network training. We call this approach deep-learning traveltime tomography. The results obtained from this method can also be used as initial velocity models for full-waveform inversion. For network training, a large number of field-scale synthetic velocity models and corresponding first-arrival traveltimes with towed-streamer acquisition are created, and then the network is trained with the synthetic dataset. As a result of performing deep-learning traveltime tomography on an example of synthetic velocity models simulating the seafloor strata, inversion results similar to the labels were obtained. Therefore, it was confirmed that the deep-learning traveltime tomography method can immediately predict a field-scale velocity model, unlike the existing deep-learning-based seismic inversion.