Velocity model building with a modified fully convolutional network

We introduce a machine learning based method to estimate the P-wave velocity models directly from the prestack seismic traces using a modified fully convolutional network. The network is tuned to map multi-shot seismic traces to velocity models. We train the network with pairs of synthetic velocity models and their corresponding multi-shot seismic traces which are simulated from the velocity models with acoustic wave equations. Multiple shots are used as channels in the network to increase data redundancy. The training process is expensive, but it only occurs once up front. The trained network is then used to predict velocity models from testing seismic traces, and it shows satisfactory prediction results. The cost for predicting velocity models is negligible once the training is complete.