Deep Learning Toolkit: Shallow Geohazards Case Study

Abstract

Identification of shallow hazard features at the sub-surface are performed via seismic interpretation of high-resolution three-dimensional (3D) seismic surveys. Today, seismic interpretation is a fastidious and time-consuming process. Usually an interpreter must spend weeks or months to fully interpret this type of survey and its volume. Deep learning methodologies utilizing neural networks are currently revolutionizing seismic interpretation by accelerating the speed in which geologic features such as faults, salt diapirs, and gas chimneys can be mapped. It has become an industry goal to utilize neural networks for all uses of seismic interpretation, including shallow hazard identification. There have been challenges to the adoption of machine learning, to the interpretation of shallow hazards. The results of this interpretation have a direct impact on drilling safety and therefore its accuracy is paramount. Traditional machine learning technologies do not consider the human experience and abstract thinking required for seismic interpretation. For interpreters to fully trust the results of the neural networks, they must spend an extensive amount of time conducting quality control of the results and editing as needed. Traditional machine learning technology requires a significant amount of time to be spent on data preparation. Our project has two main goals. The first goal is to create technological improvements to enable humans to work interactively with neural networks. The second goal is to reduce the amount of time spent in data preparation to zero. We present a new interactive deep learning methodology, that differs from existing methodologies and achieves these two goals. We have built our methodology on top of our proprietary technology that provides random data access to seismic data, enabling interactivity between the neural network and the geoscientists, therefore putting geoscientists in complete control. Our methodology eliminates the burden of data preparation methodologies, because it does not require prior randomization of the seismic data like older technologies. We achieve acceleration, and improved results in the seismic interpretation of shallow hazards.