Building PSDM velocity models in complex structure land environments is difficult. A machine-learning method using a convolutional neural network (CNN) incorporates both human and artificial intelligence to overcome these difficulties. The supervised learning process used a large representative dataset to train the CNN, learning the convolutional weights that best map the input seismic shot records to the target velocity model. With careful consideration of both training data and network architecture, the CNN can accurately predict velocity models on both synthetic and field data.

• Built convolutional neural network (CNN) to estimate TTI PSDM velocity models from shot gathers
• Traditional automated methods for PSDM velocity estimation in complex-structure land areas are unstable, so we rely on the human understanding of the geology to build geologic models
• The goal is to use machine learning to supplement human learning
• Field-data example shows imaging improvements on certain shallow reflectors, which shows the potential of the method

Greg presents poster in Kuala Lumpur