Machine Learning for the Imaging of Sparse Seismic Data in Blocky Subsurface Models

Abstract:

Convolutional neural networks (CNNs) with a U-net architecture are trained for the imaging of sparse and noisy seismic data. Synthetic subsurface interface models with blocky layers are first constructed and Gaussian beam modeling is applied for the fast generation of synthetic seismic reflection data. A CNN with a U-net is an encoder-decoder neural network architecture consisting mainly of two paths, the contracting path (encoder) and expanding path (decoder). Each path consists of repeated applications of convolution, concatenation, activation functions, max pooling and dropout operations which play the roles of capturing and reconstructing important features from the input images. The CNN model is first trained and validated using labeled subsurface interface models and corresponding seismic reflection data. A mean-squared-error (MSE) loss function is utilized to measure the difference between the imaged subsurface models with trial subsurface models, and an Adam optimizer is used for the optimization of the neural network. The trained CNN models are then utilized to image both regularly and irregularly spaced sparse seismic reflection data with noise. We also investigate methods to reduce the size of the training dataset which will also decrease the time for data synthesis and CNN training. A trained CNN has the potential to image sparse seismic data with multiple subsurface layers.