Velocity Model Estimations Using Deep Learning

This technical report presents the application of a deep neural network based on Fourier Neural Operators (FNOs) for velocity model building (VMB) in geophysical contexts. The study outlines how the FNO network can replace traditional tomographic inversion engines, enabling velocity predictions on field datasets without the need for residual move-out picking or masking. The report details the training process of the network, which includes mapping velocity errors directly from data to the image domain and utilizing a large dataset for improved accuracy. It discusses the advantages of this approach, such as faster convergence and the ability to handle complex geological features. The methodology section describes the architecture of the network, which incorporates integral operator blocks and convolutional neural networks to facilitate long-range dependencies in data. Field data examples from the Norwegian Sea and the Agung area illustrate the effectiveness of the FNO network in enhancing velocity models and improving seismic imaging, demonstrating its potential to accelerate the VMB sequence significantly.