Experiences with Machine Learning and Deep Learning Algorithms for Seismic, Wells and Seismic-to-Well Applications

Abstract

Summary Here we present our experience with the standard machine learning and the more advanced deep learning workflows in three different application domains: seismic interpretation, well data analysis, and integrated seismic-to-well inversion. Seismic interpretation is discussed on the basis of two deep Convolutional Neural Network (CNN) examples: a modified U-Net based fault prediction and a LeNet-5 based seismic chimney prediction. For the well data and integrated seismic and well applications, we present log-to-log and seismic-to-log prediction experiments with various standard machine learning algorithms (e.g. Random Forest, Deep MLPs and XGBoost). We conclude that machine learning and deep learning algorithms add value in all subsurface applications we studied. However, we learned that deeper and bigger does not necessarily mean better, i.e. we do not always need “deep learning”. For example, for log-log prediction, standard machine learning algorithms already do a good job as they can work well with the typical small amount of log data present for training – unlike the deep learning models, which require more data. In the case of the seismic chimney cube, the conventional shallow network result was preferred by the interpreter over the deep learning result, which was considered to be perhaps more accurate but less interpretable.