💻 computer science

GeoDecider: A Coarse-to-Fine Agentic Workflow for Explainable Lithology Classification

GeoDecider is a training-free, coarse-to-fine agentic workflow that leverages large language models and geological tools to achieve accurate, explainable, and geologically consistent lithology classification from well-logging signals, outperforming existing single-pass methods.

Original authors: Jiahao Wang, Mingyue Cheng, Yitong Zhou, Qingyang Mao, Xiaoyu Tao, Qi Liu, Enhong Chen

Published 2026-05-06

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Jiahao Wang, Mingyue Cheng, Yitong Zhou, Qingyang Mao, Xiaoyu Tao, Qi Liu, Enhong Chen

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to identify different types of rocks deep underground by looking at a long, continuous strip of data called a "well log." This data is like a heartbeat monitor for the Earth, showing how the rock changes as you drill deeper. The goal is to label every single point on this strip as a specific rock type (like sand, shale, or oil-bearing rock).

For a long time, computers have tried to do this automatically. Some methods are fast but make mistakes when the data is messy. Others are very smart but take forever to run, like trying to solve a complex puzzle for every single inch of the drill hole.

The paper introduces GeoDecider, a new system that acts like a team of expert geologists working together to solve this puzzle efficiently and accurately. Here is how it works, broken down into simple steps:

1. The "Quick Scan" (Coarse Classification)

Imagine a security guard at a busy airport. Instead of stopping every single passenger for a full body scan, the guard uses a quick metal detector.

How it works: GeoDecider starts with a "Base Classifier," which is a fast, lightweight computer program. It scans the rock data quickly.
The Logic: If the computer is very confident (e.g., "99% sure this is Sand"), it accepts the answer immediately. This saves a huge amount of time and money because it doesn't need to overthink easy cases.

2. The "Expert Panel" (Tool-Augmented Reasoning)

If the security guard isn't sure (e.g., the metal detector beeps, but it's faint), the passenger is sent to a special room with a panel of experts.

The Problem: Some rock layers look very similar, or the data is noisy. The fast scanner gets confused here.
The Solution: GeoDecider calls in a "Large Language Model" (a super-smart AI) to act as the expert. But instead of just guessing, this AI uses special tools to gather evidence, just like a detective:
- The Knowledge Book: It checks a digital encyclopedia of geological rules (e.g., "Sand usually sits on top of Shale").
- The Trend Analyzer: It doesn't just look at one point; it looks at the "neighborhood." It checks the rocks above and below to see if the pattern makes sense (e.g., "Rock types usually change gradually, not instantly").
- The History Book: It looks at past similar cases to see how experts classified them before.
- The Neighbor Check: It finds other drill holes nearby with similar data to see what they found.

3. The "Debate" (Multi-Perspective Reasoning)

Once the expert AI gathers all this evidence, it doesn't just give one answer. It simulates a debate between three different "personalities":

The Data Analyst: "The numbers look like Oil."
The Geologist: "But the trend above and below suggests it's actually a weak layer."
The Rule-Follower: "According to the physics of the rocks, it can't be Oil."
This ensures the AI doesn't just rely on one way of thinking, reducing the chance of a "hallucination" (a confident but wrong guess).

4. The "Final Judge" (Geological Refinement)

After the debate, a final judge steps in. This judge looks at the whole picture to ensure the story makes sense geologically.

The Metaphor: Imagine a movie editor. If a scene shows a character suddenly teleporting from a forest to a city without walking through a door, the editor cuts it out because it breaks the story's logic.
The Action: If the AI predicts a rock type that is physically impossible to exist right next to the previous rock (like a sudden, impossible jump), the Refinement Module fixes it. It smooths out the results so the final map of the underground looks realistic and continuous.

Why is this better?

Efficiency: It doesn't waste time using the "super-computer" for easy rocks. It only uses the heavy brainpower when it's actually needed.
Explainability: Unlike black-box AI that just says "Sand," GeoDecider can tell you why: "I chose Sand because the neighbor rocks are Sand, the trend is smooth, and the physics rules match."
Accuracy: By combining fast scanning with deep, context-aware reasoning, it catches mistakes that other methods miss, especially in tricky, messy data.

In short, GeoDecider is a smart, two-step workflow that uses a fast filter for easy jobs and a detailed, tool-using expert team for the hard ones, ensuring the final result is both accurate and geologically logical.

Technical Summary: GeoDecider

Problem Statement

Lithology classification, the task of inferring subsurface rock types from well-logging signals, is critical for reservoir characterization and geological modeling. While data-driven methods (machine learning and deep learning) offer efficiency, they often struggle with ambiguous lithological intervals, noisy data, and the integration of geological principles such as stratigraphic continuity. Consequently, these models frequently produce geologically implausible predictions. Conversely, human experts utilize external knowledge, tool-use capabilities, and geological reasoning to achieve accuracy, but this process is labor-intensive and difficult to scale. Furthermore, while Large Language Models (LLMs) show promise in integrating domain knowledge, their computational cost makes large-scale, point-wise application impractical. The core challenge is to achieve accurate, interpretable, and geologically consistent lithology classification while managing the computational overhead of advanced reasoning.

Methodology: GeoDecider

The authors propose GeoDecider, a coarse-to-fine agentic workflow designed to enable accurate and explainable lithology classification using training-free LLMs. The framework reformulates the classification task as an expert-like, multi-stage process that balances efficiency with deep reasoning.

1. Framework Overview

GeoDecider operates as a confidence-aware inference pipeline. It begins with a lightweight base classifier that provides initial point-wise predictions. Based on the confidence of these predictions, the system employs an adaptive routing mechanism:

High-confidence samples: Directly accepted to minimize computational cost.
Low-confidence samples: Routed to a sophisticated reasoning module for further analysis.

2. Core Components

The workflow consists of three primary stages:

A. Base Classifier-Guided Coarse Classification
A pre-trained lightweight classifier (e.g., XGBoost, GBDT) generates initial probability distributions for lithology classes. A routing decision function, $D(x)$ , utilizes a confidence threshold $\tau$ to determine whether a sample requires further reasoning. This threshold is calibrated via a validation set to optimize the trade-off between computational efficiency (coverage) and predictive reliability.

B. Tool-Augmented Reasoning with LLMs
For routed samples, the system constructs a structured "Evidence Profile" by augmenting raw numerical logs with outputs from specialized tools:

Knowledge Base Tool: Provides structured domain priors, including feature descriptions, label definitions, and expert heuristics, grounding numerical observations in geological semantics.
Contextual Trend Analysis Tool: Analyzes the target window alongside surrounding intervals to extract stratigraphic context, capturing trends like stable regimes and gradual transitions.
Neighbor Retrieval Tool: Retrieves $k$ -nearest neighbors from a reference dataset to provide empirical exemplars for case-based comparison.
Historical Classification Tool: Incorporates predicted labels from adjacent upper intervals to ensure temporal coherence without imposing hard constraints.

These tools feed into a Multi-Perspective Reasoning module, where three specialized reasoning personas (Data-Centric Analyst, Context-Aware Stratigrapher, and Rule-Based Physicist) independently infer candidate labels based on the Evidence Profile.

C. Geological Refinement
The final stage employs an LLM to perform a refinement step. It takes the raw logs, the upstream multi-perspective predictions, and explicit geological guidelines (e.g., stratigraphic continuity) as input. The LLM resolves conflicting local evidence and enforces geological consistency, producing a final lithology sequence that aligns with subsurface structural expectations.

Key Contributions

Framework Proposal: The introduction of GeoDecider, a coarse-to-fine agentic reasoning framework that aligns lithology classification with geological principles while utilizing training-free LLMs.
Modular Instantiation: The realization of the framework through three key components: a lightweight base classifier for coarse prediction, a tool-augmented LLM module for handling ambiguous cases, and a geological refinement module for enforcing stratigraphic consistency.
Empirical Validation: Demonstration of the framework's effectiveness through large-scale experiments on four public benchmarks, showing competitive performance and highlighting practical challenges in industrial deployment.

Experimental Results

The authors evaluated GeoDecider on four datasets: SEAM, Facies, Force, and Geolink.

Performance: GeoDecider consistently outperformed representative baselines, including traditional machine learning models (XGBoost, GBDT), deep learning architectures (LSTMFCN, InceptionTime), and recent LLM-based time-series models (TableTime, GPT4TS, UniTS). On datasets with high lithological ambiguity, GeoDecider achieved F1-score improvements of 3–5% and Recall improvements of 2–6% over its base models.
Efficiency: The adaptive routing mechanism significantly reduced the number of samples requiring expensive LLM reasoning while maintaining or improving overall accuracy.
Ablation Studies:
- Removing the Knowledge Base or Geological Refinement modules resulted in measurable performance drops, confirming the necessity of domain knowledge and consistency constraints.
- Contextual Trend Analysis was identified as the most critical component; its removal led to the largest performance decline, underscoring the importance of depth-wise trends.
- Multi-Perspective Reasoning proved superior to single-perspective approaches, particularly on complex datasets like Force.
Geological Consistency: The framework significantly reduced the "Flying Point Ratio" (isolated, geologically implausible predictions), producing smoother and more coherent lithology sequences.

Significance and Claims

The paper claims that GeoDecider offers a practical solution to the trade-off between classification performance and inference efficiency in geoscience. By mimicking the expert workflow—combining rapid screening with deep, tool-augmented reasoning and geological refinement—the framework achieves geologically interpretable predictions without the prohibitive cost of applying LLMs to every data point.

The authors emphasize that the framework is backbone-agnostic, performing well across different LLM backbones (DeepSeek-R1, GPT-5, Gemini 3 Pro), suggesting that the system-level design is the primary driver of success. Furthermore, the paper highlights the practical challenges of deploying such systems in highly regulated industrial environments, noting that strict data governance and privacy requirements often necessitate the use of public benchmarks for validation rather than proprietary production data. The work positions itself as a step toward reliable, automated subsurface modeling that respects geological constraints.