The broader impact/commercial potential of this Small Business Technology Transfer (STTR) Phase I project will result from a direct benefit to the energy sector of the U.S. economy, since seismic exploration will play an increasingly important role in meeting increasing energy demands and maintaining healthy oil and gas output. The goal of this project is to develop a software package for automated pattern recognition that can be used by seismic processing companies to automatically pick geological features from seismic data. Seismic data volumes have grown exponentially over the last three decades as the seismic exploration industry increases its survey coverage. Manual picking and geological pattern identification jobs, which depend on visual inspection, are labor intensive and cannot keep up with the growth in data generated by seismic surveys. In this project, the company will develop a machine vision enabled picking and identification tool trained by a deep learning network. Lessons learned in training an efficient deep learning network for pattern recognition have wide applications in other areas such as medical image analysis. This project will support the training of both graduate and undergraduate students in the areas of seismic exploration, machine learning and high-performance computing. This Small Business Technology Transfer (STTR) Phase I project aims to develop a deep learning network model to recognize unique patterns embedded in seismic data, which patterns are characteristic of the associated geological structures. Specifically, the project will demonstrate the feasibility of delivering a machine vision enabled inspection tool to relieve domain experts from labor-intensive visual examination activities. Various automatic picking approaches currently exist, with differing degrees of success. Nonetheless, the uncertainty involved in these tools is still too high for them to be widely adopted by the industry. Recent advances in the area of deep learning make it possible to surpass human-level visual recognition performance in some applications. High performance deep learning network models, however, require a large amount of high quality training data. In this project, the company proposes to use a novel self-taught deep transfer learning approach to overcome the data shortage problem resulting from proprietary rights associated with the data. The new training workflow is adaptive to the domain of seismic data processing. It will also minimize the training effort and deliver a robust system with guaranteed performance for new and unseen datasets.
