Well logs driven transfer learning for seismic velocity model building

This is the authors' abstract. We don't add key points for this paper.

: Recent advancements in machines learning have developed data-driven, end-to-end approaches for velocity model building, but these methods still struggle with the challenge of limited labeled data for practical applications. In this study, we present WellSeisNet, a novel well-seismic fusion network designed to overcome this limitation. WellSeisNet leverages well logs to fine-tune a pre-trained model, enabling it to predict improved velocity models from initial velocity models and the corresponding migration images. This approach significantly reduces the dependency on extensive labeled datasets, making it more feasible for real-world applications. WellSeisNet integrates two key components: a classic conditional generative adversarial network (cGAN) and a novel well-log adapter module. The cGAN is responsible for generating realistic velocity models, while the well-log adapter module enhances the network's performance by incorporating velocity distributions learned from well logs. Specifically, the adapter module uses adaptive instance normalization to regularize the feature maps with a velocity distribution learned from well logs, ensuring that the model generalizes well to new data without overfitting to the training set. Moreover, a transfer learning method is proposed with finetuning with well logs. By initially training the network on synthetic data, we can then efficiently fine-tune it using only well logs instead of legacy velocity models in field applications. The resulting model produces detailed and geologically plausible velocity models. Numerical experiments, conducted on the Marmousi II model, BP 2004 model and field data, demonstrate that WellSeisNet significantly improves accuracy and model generalization, even when fine-tuned with just one or two wells.