Using Full Waveform Inversion and Deep Learning with Crosswell Seismic Data to Estimate Subsurface Anomalies

Conventional full waveform inversion (FWI) of seismic data has been quite successful, but still faces many challenges especially with high frequencies and incomplete survey coverage. Deep learning (DL) has recently garnered considerable interest in its potential to assist or enhance FWI. However, the inner workings of deep neural networks (DNN) might not be well understood, and interpreting how DNNs perform their estimates may be limited. We present a hybrid FWI-DL procedure that first maps crosswell seismic shot gathers to preliminary velocity images, and then a DNN is trained to rectify errors in the resultant interwell images. Numerical examples show that this procedure can effectively remove some artifacts from conventional FWI. Compared to a pure DL inversion, the hybrid FWI- DL approach not only requires fewer samples for training, but also facilitates the interpretation of the trained DNN. We examine feature maps at various depths in a DNN for the FWI-DL inversion. It is observed that the DNN decomposes preliminary velocity images into zones with good correspondence to the wavepath coverage. We also demonstrate that when applied to data contaminated with strong white noise, the hybrid FWI-DL approach trained with noise-free data can still make high-quality predictions without retraining for the noisy data. An application of the DL inversion technique on a time-lapse crosswell seismic field dataset from West Texas depicts an anomaly caused by hydraulic stimulation that is indiscernible in the anisotropic tomography results. The FWI-DL workflow promises significant improvements in crosswell analysis as well as other geometries and data sets.