Date of Award

May 2026

Document Type

Dissertation

Degree Name

Doctor of Philosophy (PhD)

Department

Energy Engineering

First Advisor

Olusegun S. Tomomewo

Abstract

Global CO2 emissions significantly exceed the capacity of planned sequestration projects. Among subsurface storage sites, deep saline aquifers offer the largest theoretical capacity because they occur naturally across most sedimentary basins. However, major geological uncertainties, such as permeability variations and sand continuity, complicate these projects due to sparse well data. Consequently, analyzing CO2 sequestration projects in saline aquifers require robust reservoir simulation which are unfortunately, computationally complex with long runtimes, presenting severe bottlenecks for extensive risk assessment.

Machine learning algorithms can overcome these limitations by drastically reducing computational runtimes, enabling rapid scenario testing and optimization. Despite this potential, existing literature predominantly applies machine learning to simplified synthetic datasets or idealized homogenous grids, failing to capture the complex heterogeneity and massive scale of real-world geological formations.

This dissertation addresses this critical gap by leveraging real-world field data to evaluate the spatiotemporal predictive accuracy of machine learning architectures. The first phase utilizes Volve field data to predict bottomhole pressure. By applying algorithms such as linear regression, extreme gradient boosting (XGBoost), gradient boosting machines (GBM), bidirectional long short term memory networks (BiLSTM), and Kolmogorov Arnold networks (KAN), the research demonstrates that engineering features based on reservoir engineering knowledge improves model accuracy when compared to permanent downhole gauge data.

These insights were scaled to the Smeaheia CO2 storage project. This represents a pioneering application of big data machine learning in reservoir simulation, requiring the prediction of approximately 1.5 billion spatiotemporal data points. This introduced unique computational challenges, including GPU memory management, zero-inflated imbalanced datasets, and complex plume migration paths. This research documents the novel architectural modifications required to resolve these computational bottlenecks.

The model results were evaluated both qualitatively (eye-test) and quantitatively using metrics such as R2, MAE and RMSE. The machine learning models gave similar accuracy to the reservoir simulation while achieving speedups of up to 35x when making predictions compared to the reservoir simulation models. This work provides a scalable framework for deploying big data machine learning models to de-risk CO2 storage projects and serve as an operational tool through the project lifecycle for measurement, monitoring and verification.

Download

Share

COinS