From natural description to simulation: Large language models for geothermal engineering

Abstract

Computational modelling in geothermal engineering presents many opportunities for improved automation and validation. For example, manually editing configuration files and meshes for computational modelling is time-consuming and error-prone. In this study, we explore the potential of using large language models (LLMs) to automate typical tasks in geothermal engineering, focusing on generating and modifying structured simulation inputs (in JavaScript Object Notation or JSON format) for Waiwera, a modern geothermal simulator. However, these artificial intelligence (AI) tools come with risks in an engineering context, where precision, reliability and accuracy are crucial. Hence, we consider both syntactic validity and semantic accuracy of LLM generated configuration files. This article first reviews various core concepts of LLM workflows and AI engineering relevant to geothermal engineering. We then present preliminary results of experiments using LLMs for various geothermal modelling tasks, in particular knowledge base retrieval and JSON generation. We consider both open-source and commercial LLMs (e.g., OpenAI GPT-4o, Meta LlaMA 3.3, DeepSeek R1) and discuss progress towards developing a validation framework and implementing AI engineering techniques to improve performance. Initial results indicate improved knowledge retrieval through retrieval-augmented generation (RAG) and consistent generation of structurally valid JSON inputs by LLMs, with minor sensitivity in smaller files. Ultimately, we aim to develop a "virtual geothermal assistant" that improves accessibility and efficiency in geothermal simulation, while documenting best practices for LLM integration in engineering workflows.