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

1. Problem Definition

Lithology classification from well-log data is a critical sequence-labeling task in geoscience, aiming to infer rock types (e.g., sandstone, shale, limestone) from multi-dimensional geophysical measurements (e.g., Gamma Ray, Resistivity, Sonic) along a depth axis.

Key Challenges:

• Noisy & Ambiguous Data: Well logs contain significant noise, "salt-and-pepper" artifacts, and ambiguous formation boundaries.
• Stratigraphic Inconsistency: Existing models often produce geologically implausible predictions, such as rapid, unrealistic switches between rock types that violate stratigraphic continuity.
• Limitations of Current Approaches:
  • Traditional ML/Deep Learning: Rely on statistical associations, often overfitting local noise and lacking domain-specific reasoning or global consistency constraints.
  • Large Language Models (LLMs): Possess geological knowledge but struggle with numerical precision, fine-grained signal patterns, and handling noisy multichannel time-series data.
• Static Paradigm: Most existing methods treat classification as a single-step, static mapping, lacking the ability to iteratively gather evidence, self-correct, or verify predictions against geological constraints.

2. Methodology: The GeoMind Framework

GeoMind reframes lithology classification as a sequential, multi-step agentic workflow coordinated by a Planner-Executor-Reflector architecture. It integrates specialized numerical predictors with LLM-based reasoning, guided by a process-supervised training strategy.

A. Agentic Workflow Architecture

1. Planner (Global Strategy):
   • Analyzes input signal characteristics (variability, depth continuity).
   • Dynamically generates an execution plan, selecting which tools to invoke based on the specific well-log segment rather than using a fixed pipeline.
2. Executor (Local Analysis):
   • Executes a modular toolkit organized into three layers:
     • Perception: Converts raw numerical logs into semantic trend narratives (e.g., "stable segment," "turning point") and retrieves similar historical cases.
     • Reasoning: Fuses multi-source evidence (numerical trends, neighbor voting, neural classifier outputs) to generate initial lithology hypotheses and reasoning traces.
     • Analysis: Diagnoses inconsistencies using a Consensus Conflict Scanner (comparing predictors) and a Stratigraphic Sequence Validator (checking against Markov transition models).
3. Reflector (Consistency Check):
   • Synthesizes outputs from the Executor (candidate labels, confidence scores, conflict reports).
   • Resolves conflicts between predictors (e.g., when a neural net says "Shale" but trends suggest "Sandstone") to produce a final, geologically plausible label and a traceable justification.

B. Training Strategy: Process Supervision & MA-GRPO

To ensure the agent learns not just the final answer but the correct reasoning path, GeoMind introduces a novel training framework:

1. Domain-Adaptive SFT: The base LLM (Qwen3-4B) is fine-tuned on lithology-specific instruction data (QA pairs, contrastive examples) to align with geological conventions.
2. Process Reward Design: Unlike standard methods that reward only the final outcome, GeoMind injects rewards at intermediate stages:
   • Trend Analysis Reward: Rewards the quality of the semantic trend narrative (clarity, accuracy, depth alignment).
   • LLM Accuracy Reward: Rewards the correctness of the intermediate LLM prediction.
   • Reflection Correction Reward: Rewards the Reflector for successfully resolving conflicts and selecting the correct label when evidence is available.
3. Module-Aware Group Relative Policy Optimization (MA-GRPO):
   • Addresses the "credit assignment" problem in multi-step workflows.
   • Instead of a single global reward, MA-GRPO computes advantages locally for each module (Trend, Reasoning, Reflector) within a group of sampled trajectories.
   • This decouples learning signals, ensuring that specific modules are optimized for their specific contributions without interference from other stages.

3. Key Contributions

1. Agentic Paradigm: Proposes GeoMind, the first framework to model lithology classification as a plan-execute-reflect process, enabling iterative evidence gathering and self-correction under stratigraphic constraints.
2. Process-Supervised Training: Introduces a fine-grained supervision strategy using module-specific process rewards and MA-GRPO to optimize intermediate reasoning steps, ensuring logical consistency and geological plausibility.
3. Transparent Decision Making: The workflow produces explicit, traceable reasoning paths (narratives, conflict reports, validation signals), making the classification process interpretable for geoscientists.
4. Robust Generalization: Demonstrates that the framework acts as a "plug-and-play" enhancer, improving the performance of various lightweight time-series predictors (XGBoost, InceptionTime, GPT4TS) across diverse geological settings.

4. Experimental Results

The authors evaluated GeoMind on four public benchmarks: SEAM, Facies, FORCE, and GeoLink.

• Performance: GeoMind consistently outperformed state-of-the-art baselines (including classical ML, deep learning, and LLM-based methods) in Weighted F1 scores.
  • Example: On the SEAM dataset, GeoMind achieved an F1 of 0.8584, surpassing the best baseline (InceptionTime at 0.8537).
  • Complex Settings: Gains were most pronounced in datasets with high noise and distribution shifts (e.g., FORCE, Facies).
• Stratigraphic Consistency: GeoMind significantly reduced the Fragmentation Rate (percentage of geologically impossible thin layers) by up to 27.9% compared to XGBoost, proving its ability to enforce global stratigraphic constraints.
• Ablation Studies:
  • Removing the Agentic Workflow RL caused the largest performance drop, confirming the necessity of the multi-step optimization.
  • Removing Process Rewards or MA-GRPO led to unstable training and lower accuracy, validating the importance of intermediate supervision.
  • The Reflector module was crucial for reconciling conflicting evidence; replacing it with simple majority voting degraded performance.
• Training Stability: MA-GRPO demonstrated faster convergence and lower gradient oscillation compared to standard GRPO, indicating more stable optimization.

5. Significance

GeoMind represents a significant shift in geoscience data mining from static discriminative mapping to dynamic, evidence-based reasoning.

• Scientific Impact: It bridges the gap between numerical data processing and geological logic, producing predictions that are not only accurate but also geologically defensible.
• Methodological Impact: It demonstrates the efficacy of process-supervised agentic workflows for complex scientific tasks where intermediate reasoning quality is as important as the final output.
• Practical Application: By providing transparent, traceable decision paths, GeoMind builds trust with domain experts, facilitating its adoption in reservoir evaluation, mineral exploration, and geological modeling.