IAP-25-142

Intelligent Geothermal Energy Development: AI-Enhanced Modelling and Decision-Making Under Uncertainty

Geothermal energy offers substantial untapped UK potential as a baseload renewable resource crucial for energy system transformation (Gluyas et al., 2018). Effective exploration and development of geothermal energy requires reliable geological modelling of the geothermal reservoir and simulation of subsurface heat and mass transfer. Due to the inherent complexity of geological formations and the scarcity of direct subsurface data, there is a large degree of uncertainty in both geological and subsurface flow modelling. These uncertainties can lead to increased operational risks, inaccurate forecasts, and suboptimal management.

Therefore, to advance geothermal energy development, it is critical to quantify geological uncertainty and its propagation through heat and mass transfer processes. This uncertainty can then be reduced by assimilating available measurement data to calibrate geological models. To further enhance model reliability, the placement of exploration wells should be strategically designed to gather data with the highest information value, thereby maximising uncertainty reduction. Furthermore, the development strategy should be optimized under geological uncertainty to ensure efficient, sustainable, and economically sound geothermal operations.

Traditional numerical models for geothermal systems are often computationally intensive and inefficient for these uncertainty-driven tasks. AI-driven approaches, by contrast, offer scalable and efficient alternatives capable of accelerating analysis and decision-making (Wang et al., 2023).
This research aims to integrate AI with domain knowledge to enhance geothermal system modelling and management under uncertainty, directly supporting UK Geothermal Strategy and AI sector objectives.

The specific objectives are:
1) Knowledge-informed AI modelling: Integrate geological models, physical laws of heat and mass transfer, realistic geological heterogeneity, engineering controls, empirical relationships, and expert knowledge into deep learning models to develop knowledge-enhanced surrogate models for geothermal systems. These surrogates will enable efficient and accurate uncertainty quantification.
2) Intelligent exploration design: Employ deep reinforcement learning and Partially Observable Markov Decision Processes (POMDPs) to design exploration well sequential placements that maximize the information value of exploration data and minimize the geological uncertainty.
3) Optimized geothermal management: Develop geothermal system management strategies through optimization algorithms under uncertainty to achieve maximum economic benefits, reservoir sustainability, and environmental safety.

This research will develop intelligent frameworks that quantify uncertainty efficiently, monitor subsurface dynamics with enhanced accuracy, and enable informed and optimized decision-making in geothermal energy production. This research also drives broader AI adoption across geosciences, with applications extending beyond geothermal to other critical subsurface technologies, such as carbon storage, underground hydrogen systems, and mineral exploration, creating a transformative AI–geoscience interface for the global energy transition.

This research will be implemented under the supervision of Dr. Nanzhe Wang (Primary Supervisor, AI and Data Science for Geoenergy, Heriot-Watt), Dr. Zhiwei Ma (Second Supervisor, Heat and Mass Transfer, and Heat Utilization, Durham), Prof. Uisdean Nicholson (Third Supervisor, Earth and Environmental Science, Heriot-Watt), and Prof. Cédric John (Collaborative Supervisor, Data Science for the Environment and Sustainability, QMUL). This interdisciplinary research project will integrate geoscience knowledge, AI techniques, numerical simulation, probabilistic algorithms, and optimization methods to advance geothermal energy development. The potential methods corresponding to each project objective are outlined below.

Methods for objective 1: Numerical simulation of heat and mass transfer processes would be utilized for geothermal system modelling, generating training datasets and benchmarks for model validation. Various deep learning models (e.g., Convolutional Neural Networks, Graph Neural Networks, Transformers) will be explored to capture the spatial-temporal dynamics of geothermal reservoirs. Scientific machine learning strategies will then be utilized to incorporate domain knowledge (physical laws, engineering controls, empirical relationships) into the deep learning models to construct knowledge-enhanced deep learning models for higher accuracy and physical consistency (Wang et al., 2020). These surrogates will subsequently be integrated with probabilistic methods (e.g., Monte Carlo and ensemble-based approaches) for efficient uncertainty quantification of geothermal systems.

Methods for objective 2: To determine the optimal placement of exploration wells, the project will employ Partially Observable Markov Decision Processes (POMDPs) (Lauri et al., 2023) and deep reinforcement learning techniques for sequential decision-making under uncertainty. Additionally, Value of Information (VoI) analysis (Dutta et al., 2019) will be performed to assess the informational contribution of potential data acquisitions. The previously developed knowledge-enhanced models will be incorporated to accelerate forward simulations, enabling efficient evaluation of multiple exploration scenarios.

Methods for objective 3: Once geological uncertainty has been quantified and reduced, the research will focus on optimizing geothermal reservoir management strategies (such as well controls, well trajectories, injection strategies) to achieve optimal economic benefits, reservoir sustainability, and environmental safety (Wang et al., 2023). Both gradient-based and evolutionary optimization algorithms will be investigated to develop intelligent decision-making systems for operational strategies optimization under uncertainty. The knowledge-enhanced surrogate models will again be employed to accelerate the optimization process while maintaining physical reliability. Case studies based on synthetic forward models of geology will be used as ground truth to test the approaches.
In addition to computational and methodological training, the PhD student will have the opportunity to undertake field trips to geothermal sites operated by existing academic and industry partners, engaging directly with professionals to understand real-world operational challenges and explore pathways for translating research findings into practical solutions for geothermal development.

Year 1

– Literature review about geothermal energy systems and AI/deep learning techniques.
– Gain proficiency in numerical simulation tools for geothermal systems, such as (Settgast et al., 2024), Open Porous Media (OPM) (Rasmussen et al., 2021), and OpenGeoSys (OGS) (Kolditz et al., 2012), and construct a numerical model for a benchmark geothermal case study.
– Begin developing and training deep learning models for the selected case.
– Investigate strategies for incorporating domain knowledge into the deep learning models.

Year 2

– Performing uncertainty quantification for the studied case by integrating the developed deep learning model with probabilistic methods.
– Develop a strong understanding of Value of Information analysis.
– Implement reinforcement learning and POMDP frameworks to design sequential exploration well placement strategies that maximise information gain and reduce geological uncertainty.
– Present at a national conference.

Year 3

– Prepare and submit a journal manuscript based on the findings from Year 1 and 2.
– Work on the geothermal reservoir management optimization under uncertainty.
– Present at an international conference.

Year 3.5

– Prepare and submit a journal manuscript based on the findings from Year 3.
– Complete the PhD thesis and prepare for viva examination.

This interdisciplinary project will provide the student with comprehensive training across AI, geoscience, and numerical modelling, and equip the student with a unique combination of computational, analytical, and domain-specific expertise.
The student will receive advanced training in AI, machine learning, and data science techniques, including model development, coding, debugging, data analysis, and algorithm evaluation. This training will be primarily supported by the primary supervisor (Dr. Nanzhe Wang) and collaborative supervisor (Prof. Cédric John), who will provide hands-on guidance in the application of cutting-edge AI methods to subsurface systems.
In parallel, the student will gain strong proficiency in numerical simulation of heat and mass transfer within geothermal systems, guided by the primary supervisor (Dr. Nanzhe Wang and Dr. Zhiwei Ma).

The student will also develop broad knowledge in subsurface geoscience and geological modelling of geothermal systems, supported by the secondary supervisor (Prof. Uisdean Nicholson) and the collaborative supervisor (Prof. Cédric John).
Professional development will be further strengthened through opportunities to present research findings at national and international conferences (at least one of each during the PhD). The student will also have the chance to participate in field trips to geothermal energy sites and engage with industry professionals, gaining insights into the practical implementation of AI-enabled geothermal technologies. Additional fieldwork and case study opportunities—focused on aquifer geothermal systems in Scotland, Lithuania, and Ireland—could be available through the GeoGUARD geothermal project hosted by the Institute for GeoEnergy Engineering at Heriot- Watt University, providing valuable exposure to diverse geological and operational contexts.
Throughout the project, the student will also cultivate a range of transferable skills, including project management, scientific writing, peer-reviewed publication, oral presentation, communication, teamwork, leadership, and networking.