Deep Learning for Static Reservoir Modeling

Dr. Yogendra Narayan Pandey (Senior Technical Professional, Halliburton Landmark) presented at the LIFE2017 Data Science session on the application of deep learning for static reservoir modeling. During which, the development of a deep learning approach for gridless petrophysical properties simulation was discussed, providing both a portable and automated means of generating a higher resolution reservoir model.



Characterizing reservoirs and building 3D reservoir models is an extremely important step in the process of exploration of petroleum reservoirs. Generation of a reservoir model is based upon a stratigraphic grid in most existing approaches. A high-resolution model would translate to a stratigraphic grid with billions of grid-cells, which renders sharing the models across teams and locations problematic and well blocking step upscale of data obtained from the well logs can lead to loss of resolution. For uncertainty estimation, a number of realizations of the petrophysical properties are generated. Post-processing of these realizations provides the most likely realization(s) and estimates of associated uncertainty. If a higher resolution model is deemed necessary, the whole process starting from the generation of grid needs to be repeated.


We have developed a deep learning approach for gridless petrophysical properties simulation. In this approach, data from log curves at original resolution are used in conjunction with the structural framework to generate the input features, which are used for training a deep neural network. Part of the data set is kept aside as “hold-out” test data which is used to test the deep neural network. Following which, it is used for predicting properties on a point-cloud generated in the region of interest. The approach alleviates problem of portability; as the model is a collection of weight matrices and bias vectors, and provides a portable representation of the 3D reservoir model. This makes the model easier to share across the teams and locations, and a higher resolution model can be generated automatically by redefining the number of points in the point-cloud. A scalable implementation of the above methodology is enabled by distributed shared memory architecture with deep learning models trained on high performance batch systems.

This approach has shown the potential to significantly reduce the time and efforts required for static reservoir modeling, while providing high-resolution, data-driven, gridless, 3D models of reservoir properties.


The presentation can be found here.