Deep learning for automatic geophysical interpretation with uncertainty quantification

Geophysical interpretation such as picking faults and geobodies, analyzing well logs, and picking arrivals is a tedious, manual, and time-consuming process. Deep learning is a data-driven technique that has been getting more attention recently in different fields, such as medical imaging and computer vision. With large volumes of available data of different types and advancements in computing technology, geophysics is a promising field for applying deep learning. Applying deep learning to geophysical interpretation can make the process faster and the workflow less subjective. Decision-making based on interpretation is uncertain. Therefore, uncertainties in geophysical interpretation are very important. To utilize the deep learning models effectively, uncertainties from data and models’ parameters need to be quantified. In this dissertation, I address the problem by including uncertainties in several deep learning-based interpretation algorithms, and show the feasibility of applying them to various geophysical interpretation problems on different types of data. First, I develop a generative adversarial network to produce data that have style from a particular region in the field data. Different styles allow to generate different data to train various convolutional neural networks for automatic fault picking in 2D seismic images. I use a bootstrapping method to generate prediction scenarios and quantify the uncertainties from training data. Second, I introduce an end-to-end network for picking channel geobodies in 3D seismic volumes, which includes uncertainties from data and the model’s parameters. This workflow is fast and easy to quantify uncertainties, not only from data, but also from the parameters of a neural network. I then apply a similar workflow to quantify the uncertainties from the model’s parameters in picking channel facies and faults simultaneously in 3D seismic volumes. I also analyze the relationship between quantified uncertainties and geologic features in the seismic volumes. Apart from applying the workflow to the segmentation problem, I design a recurrent style network for predicting missing sonic logs from gamma-ray, density, and neutron porosity logs. This is a regression problem with two outputs of compressional and shear sonic logs. The workflow generates mean prediction and quantile values for upper and lower bounds. In the last chapter, I apply a transformer-based network for picking arrivals of earthquake data. I change from discrete labels of 0 and 1, where ones are picks, to continuous distributions with peaks at picks. This helps to quantify the uncertainties of the picking algorithm along time. Finally, I discuss some limitations and suggest some possible future research topics.