Underground hydrogen storage: A recovery prediction using pore network modeling and machine learning

Abstract

Understanding the hydrogen-brine transport properties and the hydrogen trapping rate (i.e., ratio of residual hydrogen saturation after recovery to initial hydrogen saturation) is critical to the site selection for underground hydrogen storage (UHS). In this study, a three-dimensional pore network model (PNM) was used to simulate hydrogen-brine two phase flow in various porous media, including sandstone, carbonate, and sand packs. Surface contact angles measured from previous experiments were used to study the influence of the wettability on hydrogen transport in porous media. Many studies have investigated the impact of these factors on carbon dioxide sequestration. However, because of the difference in the thermal dynamic properties of the fluids and the purpose of UHS and carbon dioxide sequestration, it is still essential to analyze the UHS performance under the different rock and fluid properties. PNM simulations showed that a relatively larger contact angle with low water affinity was more suitable for UHS due to its low hydrogen trapping rate. Two machine learning methods, the least square fitting and the support vector machine (SVM), were developed to classify the ability of a rock to trap hydrogen and to predict hydrogen trapping rates. Hydrogen trapping rates simulated using the PNM were used as the training data in the machine learning models. The SVM classified rock samples into two groups which had high hydrogen trapping rates (>50 %) and low hydrogen trapping rates (<50 %). The machine learning results showed that rock samples with a low ratio of pore size to throat size and high pore connectivity (i.e., average number of throats connected to a given pore) were favorable for a low hydrogen trapping rate. This study illustrated that the impact of both rock surface wettability and pore structural geometry should be accounted for when evaluating a hydrogen-brine two-fluid system in porous media. This work is the first study that combined PNM simulation with machine learning to investigate the influence of rock surface wettability and pore structure features on the hydrogen trapping rate, which will provide insights into site selections and decision making in large-scale UHS projects.