A hyperbox classifier model for identifying secure carbon dioxide reservoirs

Abstract

Carbon management technologies such as carbon dioxide capture and storage and direct air capture systems will be needed to mitigate climate change in the coming decades. Both of these technologies will depend on the availability of secure geological storage sites that can permanently hold carbon dioxide with minimal risk of leakage. Machine learning tools that can characterize candidate storage sites based on geological data can aid decision-makers in planning carbon management networks. In this work, a mixed integer linear programming model is developed to generate a binary hyperbox classifier for determining the integrity of a candidate storage site. The model is calibrated and validated using literature data on natural carbon dioxide reservoirs, resulting in a set of IF-THEN rules that are readily interpreted by decision-makers. The approach developed here also includes rule simplification features and the capability to account for statistical Type I (false positive) and Type II (false negative) errors. Different sets of rules can be generated using the model based on user-defined number of hyperboxes. The best set of rules can be selected based on a combination of its performance with the validation data and consistency with expert knowledge. Using the case study for identifying secure CO2 reservoirs, the set of rules which resulted in zero false positives using the validation data was generated using three hyperboxes. However, an alternative set of rules which falsely predicted two out of three insecure sites as positive provides simpler rules indicating CO2 density and reservoir depth as the most important criteria.