Bayesian deep learning and uncertainty quantification applied to induced seismicity locations at the Groningen gas field in the Netherlands –What do we need for safe AI?


Title

Bayesian deep learning and uncertainty quantification applied to induced seismicity locations at the Groningen gas field in the Netherlands –What do we need for safe AI?

Publication Type
Conference Paper
Year of Publication
2019
Conference Name
SEG Technical Program Expanded Abstracts
Date Published
09/2019
Publisher
Society of Exploration Geophysicists
Publication Language
eng
Citation Key
3265
Abstract

Recently, with the increase of dense seismic monitoring networks and the resultant massive seismic data all over the world (e.g., Groningen gas reservoir), the traditional geophysical algorithms have faced the “big data” and high dimensionality issues and become inefficient and costly. Deep learning, as a candidate to mitigate the “big data” and high dimensionality problems, has started to be applied to many geophysical problems; however, previous deep learning studies seldom consider the model uncertainty, which may seriously impair reservoir production estimates, ground motion predictions, and seismic early warnings. Bayesian neural networks provide a practical solution to solve the uncertainty quantification problems in deep learning, i.e., to make AI safe. In this paper, we construct a Bayesian convolutional neural network and implement a stochastic regularized technique – dropout – to quantify the uncertainty of seismic location. This method has been applied to induced seismicities in the Groningen gas field in the Netherlands.