209-9 LLM for Efficient Ontology Construction in Petroleum Source Rock Evaluation

Session: Deep-Time Earth and the AI Revolution

Presenting Author:

Authors:

Abstract:

To address the severe fragmentation of geological knowledge and the lack of structured ontologies hindering knowledge discovery and intelligent decision-making, this paper proposes an intelligent ontology construction method integrating Large Language Model (LLM) technology with the Ontology Learning Layer Cake framework. This method aims to achieve semantic integration of multi-source heterogeneous geological knowledge and automate the generation of complex domain logic rules.Specifically, after domain experts define core concepts, the approach employs customized LLM prompt engineering to extract standardized terms from unstructured and semi-structured sources (e.g., geological reports, academic literature) and accurately define semantic relationships. Leveraging the deep semantic reasoning capabilities of LLMs, it assigns confidence levels and introduces a dual-evidence verification mechanism to build a logically self-consistent ontological relationship network. Web Ontology Language (OWL) axioms are used to constrain properties (e.g.,numerical ranges, enumerated values), while iterative refinement is supported through a closed-loop process involving dynamic generation and expert feedback.Experimental results show that the constructed source rock ontology establishes a three-dimensional evaluation system—"Organic Composition – Geochemical Indicators – Thermal Evolution Stage"—through logical associations of key parameters such as organic matter type, Total Organic Carbon (TOC), and Vitrinite Reflectance (Ro). This approach interlinks term definitions and indicator logic, overcoming the limitations of fragmented knowledge. The resulting knowledge graph integrates heterogeneous sources (reports, literature, databases), enabling semantic interoperability across structured values (e.g., TOC = 5.7 wt%), semi-structured descriptions (e.g., "oil window"), and coded classifications (e.g., KerogenType = Type II).Moreover, by dynamically generating OWL axioms (e.g., "TOC < 0.5 wt% → Non-effective Source Rock"), the system can automatically trigger conflict alerts for parameter combinations (e.g., "Shale Source Rock + TOC = 0.3 wt% + Ro = 0.8%"), enabling computable reasoning in hydrocarbon assessment.This work addresses the limitations of traditional manual ontology construction by providing an automated, scalable, and standardized framework for geological knowledge engineering. Future research will focus on enhancing LLM-domain integration to further improve ontology quality and reasoning performance.

Geological Society of America Abstracts with Program. Vol. 57, No. 6, 2025

doi: 10.1130/abs/2025AM-9708

© Copyright 2025 The Geological Society of America (GSA), all rights reserved.