Capabilities of text-to-text large language models in seismic velocity estimation of CO2- or H2-saturated rocks

Abstract Since the release of ChatGPT in 2022, generative artificial intelligence powered by large language models (LLMs) has gained significant traction, demonstrating its potential across various fields. This study explores the capability of pretrained text-to-text LLMs in rock physics modeling, focusing specifically on estimating seismic velocities in CO2- or H2-saturated rocks. Such estimations are essential for subsurface gas storage, a crucial component in advancing the energy transition and achieving decarbonization goals. These gas injections alter rock elastic properties, such as seismic velocities and density, necessitating accurate monitoring. While existing methods rely on theories, empirical relations, and machine learning, these methods face challenges like data scarcity and model uncertainty. We evaluate four popular language models — GPT-4o, GPT-4o mini, GPT-4 turbo, and Claude 3.5 Sonnet — in blind tests using experimental benchmark data to assess their ability to optimize model selection and parameter determination in rock physics.