Deep learning-based VSP velocity model building using first-arrival traveltime

The development of DAS technology has brought new opportunities and challenges to the field of oil and gas resource exploration. The first arrival traveltimes, due to the receivers being located in the well in DAS-VSP/VSP type data, carry more information from the deep underground strata, which offers the possibility for building depth domain velocity models using first arrivals. This work aims to explore the capabilities of artificial neural networks to establish underground velocity models based on the VSP first arrival traveltimes. We create the training dataset by utilizing the zero-offset velocity from VSP data. According to the observation system identical to the target data, the forward modeling is performed to acquire the corresponding first arrivals. In training datasets, the input data are the modeling first arrivals, while the corresponding velocity models are the labels. We designed a network and trained it on these datasets, then applied the trained network to the target first arrivals to acquire the predicted velocity model. Both synthetic data and field data demonstrate the validity and effectiveness of our proposed approach.