Big Data-Driven Advanced Analytics: Application of Convolutional and Deep Neural Networks for GPU Based Seismic Interpretations

Abstract

Abstract The democratization of artificial intelligence began with the collection of large datasets and the ability to consume them for inferences and prediction by leveraging exponentially increasing computational power. This was further enhanced by the ability to parallel process datasets by breaking them up into independent units and applying fast computation to those units using Graphical Processing Units (GPUs). We discuss one such application in the area of seismic interpretation in the oil and gas sector. Seismic interpretation is particularly suitable since seismic datasets’ characteristics make them inherently amenable to parallel processing in post-stack format. True big data in the oil and gas sector exists in the seismic arena, where up to one terabyte of data or more can be collected per hour. We show that while preprocessing is required for cleansing and quality checking the data, novel techniques can be applied from the medical and healthcare sectors, namely radiology, for image processing and anomaly detection in images. Further, we also show methods of preprocessing the seismic data, development of seismic images, and novel denoising techniques that lead to the construction of seismic cubes. 3D and 4D seismic cubes post-stack are today amenable to a plethora of neural network based parallel processing methods for anomaly detections. Using such techniques, the achievable speeds to detect anomalies are effectively above ten times what an experienced human being can do. These methods, at their very best, go far beyond human capabilities in terms of processing terabytes to petabytes of data. Such image processing is in use in image recognition, for example, in border control for person identification and verification, and satellite image analysis to discover the minutest of details. We showcase novel techniques based on Convolutional Neural Networks and Deep Neural Networks being utilized for subsurface geological and geophysical properties identification. Further developments of our current and future work are also discussed. Our presentation specifically describes methods based on Convolutional and Deep Neural Networks to predict faults and salt domes in seismic images. The ability of Deep Neural Networks to continuously learn and self-optimize is the basis of our novel approach. A common criticism of machine learning methods is that most reported results describe results on field data where part of the field data has been used in training the neural network algorithms being used. In the work reported here, we describe the results of our algorithm on two completely blind field data sets – where none of the field data has been used in training the algorithms.