2019 AAPG Annual Convention and Exhibition:

Datapages, Inc.Print this page

Deep Learning Applied to Fault Interpretation and Attribute Computation

Abstract

An artificial intelligence technique called Deep Learning (DL) is shown to perform the seismic interpretation tasks of picking faults and computing attributes much faster than conventional methods. This increase in speed frees interpreters to concentrate on more valuable prospect de-risking tasks.

These results are enabled by 1) overcoming the prohibitive memory requirements typical of 3D Convolutional Neural Networks (CNNs) for segmentation and regression by implementing a novel, memory-efficient 3D-to-2D convolutional architecture and 2) including synthetically generated labeled examples to enhance DL network training. The synthetic examples are advantageous because: a) they can be generated in huge quantities, b) they can include a wide diversity of examples, and c) the labels are always accurate. We have trained our DL networks using tens of thousands of input/output data pairs drawn from real and synthetically generated data.

Current fault interpretation workflows require labor-intensive picking and editing of fault segments often on a sparse set of vertical sections. This can take several weeks, provide results that lack detail and result in repetitive strain injuries for the interpreter. Increasingly, faults are left uninterpreted in 3D seismic due to these constraints unless they are recognized as critical trapping elements. Our DL powered fault interpretation tool uses a fully trained CNN to segment faults from 3D seismic in just a few hours. Initial fault picking results show an ability to pick dense fault networks with very few false positives. This result gives our business units quick access to a fault network across entire datasets which improves their block-wide understanding of structural configurations allowing them to more quickly focus on the trapping configurations of their prospects.

Similarly, it is a time-consuming activity to choose appropriate parameters for and to compute algorithmically complex seismic attributes. Our DL-based attribute computations increase computation speed up to 100x and they mimic conventionally computed attributes very closely. Numerous Shell interpreters have done blind-test comparisons and found it difficult or impossible to distinguish between the DL and conventionally computed versions.