GeoDecider: LLM-Based Lithology Classification Workflow

A new research paper introduces GeoDecider, a coarse-to-fine agentic workflow for explainable lithology classification using large language models (LLMs) without training. Lithology classification infers subsurface rock types from well-logging signals, crucial for reservoir characterization. Existing methods typically treat it as a single-pass task, but GeoDecider mimics expert geologists by incorporating geological principles, external knowledge, and tool use. The workflow is organized into multiple stages, starting with a base classifier-guided coarse classification that provides a rough estimate. This approach aims to improve accuracy and explainability in geoscience applications.