Reducing uncertainty in geological modelling

Author: Anna Roubickova
Posted: 5 Dec 2019 | 11:57

Cognitive Geology Ltd is an independent technology company based in Edinburgh, UK. It delivers innovative geological modelling software to the oil and gas industry, with the goal of improving efficiency and reducing uncertainty in geological modelling workflows and the business decisions which are based on them. EPCC has worked with the company to investigate ways to reduce large ensembles of geological models while maintaining the range of plausible scenarios described by the set.

Cognitive Geology’s software, Hutton, allows the exploration of the thousands of possible ways sub-surface properties may be distributed through an oilfield, with all the generated scenarios being based on geological first principles. For example, deciding the production quality of a reservoir under a specific scenario is computationally challenging and, given the high number of possible scenarios, investigating all possibilities is simply not feasible. The collaboration between EPCC and Cognitive Geology looked at machine learning approaches to select a few scenarios that would truthfully represent the original ensemble.

During the first part of the collaboration we tested whether Cognitive Geology’s novel reduced-basis representation of a complex 3D geological model of the Earth still contains enough information to derive properties of the sub-surface that are of interest to the petroleum industry. Using supervised learning techniques we confirmed that the relevant patterns can be learned from a rather small sample of 2000 geological models, which corresponds to less than 15% of the possible models.

The second part of the work looked at grouping the geological models based on their associated outcome to identify a representative subset. Such a grouping cannot rely on statistically significant similarities, as such an approach disregards rare and anomalous scenarios which in turn might lead to wrong decisions regarding the reservoir development. Clustering techniques generally allow identification of such outliers, but need a user to define a metric that describes similarity of different models. We have shown how the machine learning techniques developed in the previous step can be applied to this task. The choice of such a metric was further validated by experiments using various clustering algorithms and different metrics which demonstrated the importance of the metric over the specific algorithm.

The collaboration led to a reduction of 13824 models to 64 representative models, and this reduction was done based on a detailed evaluation of less than 15% of the original model ensemble. This work also provided new insights into the properties of the reduced basis representation used in this work as well as in Cognitive Geology’s software.

From EPCC’s perspective, the project was a great opportunity to deepen our experience with oil and gas exploration thanks to excellent domain specific support from Cognitive Geology. The findings of this work provide a proof of concept for the techniques developed by Cognitive Geology and highlighted several new lines of enquiry to follow up in the future.

Cognitive Geology: www.cognitivegeology.com

Authors

Lucy MacGregor, Cognitive Geology
lucy.macgregor@cognitivegeology.com

Anna Roubíčková, EPCC