Source-linked AI summary
Decision-Support and Modeling with Large Language Models for Geothermal Well Arrays
Edwin Ouko, Emmanuel Lujan, Alan Edelman, Robert Metcalfe
TL;DR
Geothermal applications need reliable decision support and efficient numerical modeling, while LLM outputs can be biased, hallucinated, or difficult to verify. This study uses NotebookLM-generated benchmarks, LLM interviews, and an auto-parallelization case study to assess leading general-purpose and geothermal-specific models. Gemini, Claude, and Grok achieved a highest minimum accuracy of 85%, while ChatGPT-o1 and Claude 3.5 Sonnet generated accurate, executable parallel code for the coaxial-well model.
Problem
Geothermal decision support requires complex data analysis, informed recommendations, and modeling, but LLM reliability is limited by hallucinations and the lack of current standardized geothermal benchmarks.
Method
The study combines NotebookLM-generated geothermal multiple-choice benchmarks, LLM-based interviews, and an auto-parallelization case study using an in-house coaxial-well model.
Results
Gemini, Claude, and Grok achieved a highest minimum accuracy of 85%, while ChatGPT-o1 and Claude 3.5 Sonnet accurately identified the CPU-threading macro placement and produced executable parallel code.
Takeaways & Limitations
LLMs can support geothermal data interpretation, design exploration, and numerical-model enhancement, but their outputs require validation before serving as foundational decision support.
Takeaways & Limitations
ChatGPT and Gemini often omit citations for data, facts, calculations, and assumptions, requiring external validation of their generated information.
Abstract
from arXiv · showhide
Geothermal well arrays, which organize multiple geothermal wells into carefully planned geometric configurations, provide opportunities to enhance energy production capacity and increase fault tolerance. The development and adoption of these emerging geothermal technologies could be accelerated through the recent advances in large language models (LLMs) and high-level high-performance languages. A challenge in LLM-based applications is the reliability of the generated outputs, as they can be prone to subjective biases and hallucinations. This study assesses the potential of cutting-edge LLMs - such as ChatGPT, Gemini, Claude, Grok, and domain-specific models like AskGDR - as expert assistants that can synthesize insightful interpretations of complex geothermal data, as well as improve feature capabilities of geothermal models and numerical software. We developed a novel approach, leveraging Google's recently introduced AI assistant, NotebookLM, to accelerate the generation of unpublished quantitative geothermal benchmarks. The rapid generation of these evaluation instruments is essential for assessing the swiftly evolving capabilities of emerging language model technologies. In particular, we use these benchmarks and LLM-based interviews to analyze opportunities and limitations of two promising technologies: geothermal well arrays and closed-loop coaxial wells. Furthermore, we present a case study illustrating how LLMs can facilitate auto-parallelization of geothermal numerical models. Our analysis emphasizes their application in digital twins and underscores the importance of high-level, high-performance code generation. This line of research could play a transformative role in the geothermal sector by enabling the next-generation of decision-support applications, integrating data analysis, informed recommendations, and more dynamic numerical modeling workflows.
1. INTRODUCTION
The introduction frames geothermal well arrays and closed-loop coaxial wells as promising technologies whose development may benefit from LLM-based decision support, modeling, and high-performance code generation. It outlines reliability challenges, digital-twin extensions, and the study’s benchmark and auto-parallelization contributions.
- Geothermal technologies: Geothermal well arrays organize multiple wells geometrically to increase energy-production capacity and fault tolerance, while coaxial closed-loop systems may reduce drilling costs.These technologies are presented as responses to the need for dependable, cleaner, and economically viable energy.
- Research gap: LLM applications in geothermal energy remain insufficiently established, with no adequate strategies or methodologies yet available for integrating them into decision-support applications.The introduction also identifies hallucinations, subjective evaluation, and outdated or absent standardized benchmarks as reliability challenges.
- LLM applications: The study examines LLMs in two roles: providing data-informed answers to decision-makers and enhancing geothermal models, especially digital twins.The proposed digital-multiplet concept connects multiple models for complementary analysis, mutual validation, new physical dynamics, geometry changes, model reduction, or parallelization.
- Digital twins: Digital twins are high-fidelity virtual counterparts of physical systems that share real-time data for monitoring, analysis, optimization, and predictive insights.The paper applies this framing to coaxial wells and links it to anomaly detection, operational-cost reduction, and geothermal energy management.
- Computational modeling: High-level, high-performance languages such as Julia are positioned as a way to combine development productivity with efficient execution in LLM-generated geothermal code.The introduction connects this capability to digital twins and digital multiplets.
- Study contributions: NotebookLM is used to accelerate geothermal benchmark creation, while LLM-based interviews and an auto-parallelization case study assess decision support and numerical-model enhancement.The study specifically evaluates benchmark questions on geothermal energy, well arrays, and closed-loop coaxial wells, alongside high-level, high-performance code generation.
2. METHODOLOGY
The methodology combines quantitative multiple-choice evaluation of geothermal-focused LLMs with LLM-assisted geothermal modeling and auto-parallelization experiments. The modeling implementation uses a full 3D heat equation, finite differences, and a uniform mesh.
- LLM benchmark methodology: The study evaluates LLM accuracy quantitatively using multiple-choice questions rather than relying only on subjective human review.The questions are designed for geothermal research.
- LLM benchmark methodology: Google NotebookLM is used to accelerate development of geothermal-specific LLM benchmarks from uploaded multimodal information.The tool supports organizing, understanding, questioning, explaining, and summarizing the provided material.
- LLM-enhanced modeling: The auto-parallelization experiment assesses the software capabilities of LLMs using an in-house closed-loop coaxial geothermal model.The experiment prioritizes software performance rather than physically realistic predictions.
- LLM-enhanced modeling: The geothermal implementation represents the heat equation in 3D and computes it with finite differences on a uniform mesh.The model uses subdomains for working fluid, rock formation, and coaxial well, with fixed velocities and no convection in the rock or well subdomains.
3. RESULTS AND DISCUSSION
The study evaluates LLM accuracy, auto-parallelization, geothermal design brainstorming, and associated limitations using newly generated benchmarks and model-assisted experiments.
- LLM Accuracy of Geothermal Energy Q&A: New geothermal multiple-choice benchmarks evaluated ChatGPT o1, Gemini Advanced, Claude 3.5 Sonnet, Grok 2, and AskGDR across four geothermal categories.The categories were general geothermal energy, well arrays, closed-loop systems, and coaxial wells.
- LLM Accuracy of Geothermal Energy Q&A: 85% was the benchmark accuracy threshold, with Gemini, Claude, and Grok identified as the most reliable models by minimum accuracy.AskGDR accuracy was calculated both by excluding “I do not know” responses and by treating them as incorrect.
- LLM-based Auto Parallelism: ChatGPT-o1 and Claude 3.5 Sonnet correctly identified the @threads placement, producing parallel, compilable, and executable coaxial-well code without errors.The experiment used a serial in-house coaxial-well model as its software testbed rather than prioritizing physically realistic predictions.
- LLM-based Auto Parallelism: Gemini Advanced applied @threads to all loops, creating excessive spatial partitions, high parallelization overhead, and an inefficient strategy for this case.The excessive partitioning reduced computational workload per worker.
- LLM-Assisted Design and Cost Modeling: ChatGPT and Gemini proposed cost-saving geothermal design strategies, including modular assembly, just-in-time delivery, and networked redundancy.ChatGPT o1 also outlined trade-offs between well parameters, but without citing supporting references.
- LLM-Assisted Design and Cost Modeling: LLM-generated design and cost-modeling responses often lacked references, numerical-data sources, and underlying assumptions, requiring external validation.The quality of responses also depended on the prompter providing well-stated questions with substantial context.
4. CONCLUSION
The study combines LLMs with high-performance languages to support emerging geothermal technologies, benchmark model reliability, and improve geothermal numerical modeling. Its results include reliable benchmark performance from several models and successful auto-parallelization by ChatGPT-o1 and Claude 3.5 Sonnet.
- The study leverages LLM technologies and high-level, high-performance languages to accelerate adoption of geothermal well arrays and closed-loop coaxial wells.
- Gemini, Claude, and Grok achieved a highest minimum accuracy of 85% on the multiple-choice geothermal benchmarks, while models differed in result dispersion.
- ChatGPT-o1 and Claude 3.5 Sonnet identified the optimal CPU-threading macro placement, producing parallelized, error-free, executable code for an in-house coaxial well model.