GeoMind: An Agentic Workflow for Lithology Classification with Reasoned Tool Invocation

Note: This repository contains the official implementation of the paper "GeoMind: A Agentic Workflow for Lithology Classification with Reasoned Tool Invocation".

📖 Introduction

GeoMind is a novel agentic framework that treats lithology classification not as a static pattern-matching task, but as a sequential multi-step reasoning process.

Traditional methods (Classical ML or standard LLMs) often struggle with noisy well-log data ("salt-and-pepper" noise) or fail to adhere to stratigraphic consistency. GeoMind bridges this gap by coordinating specialized tools through a Planner-Executor-Reflector architecture, optimized via a novel Module-Aware Group Relative Policy Optimization (MA-GRPO) algorithm.

🚀 Key Features

• 🤖 Agentic Workflow:

  • Planner: Dynamically decomposes tasks and orchestrates tool invocation based on log characteristics.
  • Executor: Utilizing a hierarchical toolkit including Trend Pattern Extractor, Neighbor Vote Aggregator, and Neural Probability Interpreter.
  • Reflector: Performs self-correction and resolves conflicting signals using geological constraints (Stratigraphic Sequence Validator).

• 🧠 MA-GRPO Training:

  • Implements Module-Aware Group Relative Policy Optimization to solve the credit assignment problem in multi-step reasoning.
  • Optimizes intermediate reasoning steps (trend analysis, reasoning traces) alongside final prediction accuracy.

🛠️ Environment Setup

Prerequisites

• Linux OS
• Python 3.8+
• CUDA-capable GPU (Recommended: A800/H800 for training)

Installation

1. Clone the repository

2. Create a virtual environment

   python -m venv venv
   source venv/bin/activate

3. Install dependencies

   pip install -r requirements.txt

📂 Data Preparation

We support four standard benchmark datasets. Please ensure your data is preprocessed (windowing, normalization, and outlier removal) as described in the paper.

• Facies: Council Grove gas reservoir logs.
• FORCE: 2020 Lithology prediction challenge data.
• GeoLink: Large-scale North Sea dataset.
• SEAM: SEG Wiki open data.

Structure your data directory as follows:

data/
├── facies/
├── force/
├── geolink/
└── seam/

🚅 Training

To train the GeoMind agent using the MA-GRPO strategy with process rewards:

python train_well_log.py train \
    --exp_name geomind_kdd_experiment \
    --model qwen3-4b \
    --model_name gpt4ts \
    --datasets facies \
    --epochs 10 \
    --runners 64 \
    --gpus 8 \
    --nodes 1 \