From Unstructured Geological Data to 3D Models: A Human-in-the-Loop LLM assisted Workflow for Automated Geological Model Building

Building implicit 3D geological models requires the detailed integration of diverse data sources, including legacy drill logs, technical reports, and stratigraphic descriptions. While this process is fundamental to understanding the subsurface, the manual translation of unstructured text into quantitative model inputs is a time-intensive task. Large Language Models (LLMs) offer promising capabilities to assist in processing data presented as text, but their application requires rigorous control to ensure geological validity. We present an ongoing research project developing a Human-in-the-Loop (HITL) workflow that leverages uses a collaborative human-AI approach to structure raw descriptions into inputs that will be used for implicit modeling.

The proposed workflow grounds the LLM in a formal Axiom-Based reasoning framework designed to minimize hallucinations and ensure consistency. The process begins with entity extraction, where the LLM parses depths and lithological descriptions from raw logs, followed by an axiomatic reasoning phase where units are categorized based on standardized rules (e.g., the Lithotectonic Framework). Crucially, the workflow integrates a dedicated validation interfaces that empowers geologists to go beyond simple verification. Experts use this environment to contextualize interpretations, test different stratigraphic hypotheses, and inject external knowledge such as fault definitions or regional correlations, before the structured output is finalized. This effectively translates text into the specific geometric parameters and interface points required to initialize the GemPy modeling engine.

We are applying this workflow to legacy data from the Campine Basin. The objective is to demonstrate how AI can function as a reliable assistant for data structuring, potentially reducing the time required for model initialization. Our workflow shifts the priority from slow data processing to critical validation; we aim to allow geologists to focus more on conceptual definitions and uncertainty analysis rather than data management. Ultimately, this research seeks to facilitate the creation of self-updating geological models that can continuously ingest and interpret new textual data as it becomes available.