GeoMind: An Agentic Workflow for Lithology Classification with Reasoned Tool Invocation
Lithology classification is a critical task in geoscience that involves identifying rock types from complex, multi-dimensional well-log data. Traditionally, this has been treated as a static, single-step mapping problem, where models attempt to predict labels directly from raw data. This approach often fails to account for geological context, leading to predictions that are inconsistent with stratigraphic reality or overly sensitive to noise. GeoMind addresses these issues by reframing lithology classification as an agentic, multi-step reasoning process that mimics how human experts analyze geological data.
A Structured Reasoning Workflow
Instead of a single forward pass, GeoMind uses a "Planner-Executor-Reflector" architecture to manage the classification process. A global planner first analyzes the characteristics of the well-log data and determines the best sequence of tools to use. The executor then carries out the plan using a hierarchical toolkit that includes perception tools to identify trends, reasoning engines to form hypotheses, and analytical validators to ensure the results align with known geological constraints. Finally, a reflector module synthesizes these various inputs to produce a final, geologically plausible prediction.
Integrating Tools and Logic
The framework bridges the gap between numerical sensor data and geological knowledge by using specialized tools. For example, a "Case Retriever" finds similar historical well-log patterns to provide context, while a "Trend Pattern Extractor" converts complex numerical curves into natural language descriptions. By combining these with a "Stratigraphic Sequence Validator"—which checks for impossible transitions between rock types—the system ensures that its decisions are not just statistically likely, but also physically and geologically sound.
Process-Supervised Learning
A key innovation in GeoMind is its training strategy. Rather than only rewarding the final classification accuracy, the researchers introduced a "fine-grained process supervision" strategy. This method provides rewards at intermediate steps, such as the quality of the trend narrative or the ability of the reflector to resolve conflicts between different predictors. By using a technique called module-aware group relative policy optimization, the model learns to improve its internal reasoning steps, resulting in a system that is not only more accurate but also provides a transparent and traceable path for how it reached its conclusions.
Performance and Transparency
Experiments conducted on four benchmark datasets show that GeoMind consistently outperforms traditional data-driven models and standard language-model-based approaches. By moving away from static, "black-box" mappings, the framework provides explicit reasoning traces that explain the logic behind each classification. This makes the system more robust against the "salt-and-pepper" noise often found in well logs and ensures that the final output remains consistent with the broader stratigraphic context of the subsurface environment.

Comments (0)
to join the discussion
No comments yet
Be the first to share your thoughts!