GeoMind: A New Method for Lithology Classification Based on Tool-Enhanced Agent Workflow

GeoMind is a tool-enhanced agent framework for well logging lithology classification, which transforms traditional static discrimination into a sequential reasoning process. Through the collaborative work of three modules—perception, reasoning, and analysis—combined with a process supervision strategy, it achieves geologically reasonable and interpretable decisions. Experiments show that this method consistently outperforms strong baselines on four benchmark datasets and provides a transparent and traceable reasoning path, offering a new direction for the credible development of geological AI.

Well logging lithology classification is a fundamental and critical task in oil and gas exploration and geological research. It infers underground rock layer types through geophysical logging data, directly related to reservoir evaluation and oil-gas resource estimation. Traditional machine learning methods treat it as a static single-step discrimination task, inputting multi-dimensional logging sequences and directly outputting lithology labels, which has fundamental flaws: lack of evidence-based diagnostic reasoning based on geological standards, prediction results often disconnect from geological reality due to the absence of domain prior knowledge, and exhibit a "black box" characteristic, making it impossible to explain the basis for judgment or verify using knowledge such as sequence constraints and sedimentary environments.

GeoMind's core innovation lies in modeling lithology classification as a sequential reasoning process, drawing on the working methods of human geologists: observing trends → proposing hypotheses → verifying constraints. Its architecture includes three modules:

1. Perception Module: Converts raw logging curves into semantic trend descriptions (e.g., "A rising resistivity interval may indicate improved permeability"), elevating low-level numerical signals to high-level geological concepts.
2. Reasoning Module: Based on semantic trends, infers lithology hypotheses with confidence levels (e.g., 70% fine sandstone, 25% siltstone) from multi-source evidence including current depth features, adjacent segment context, and prior constraints from the geological knowledge base.
3. Analysis Module: Verifies whether the prediction results comply with geological constraints (e.g., sandstone should not appear below a certain depth in marine sedimentary layers). If violated, it triggers re-reasoning or reduces the confidence of the hypothesis.

Additionally, the global planner adaptively coordinates the module workflow based on input data characteristics: using a simplified process for regular data, and enabling a complete verification cycle or even iterative reasoning-analysis for complex geological environments to balance efficiency and accuracy.

The core challenge of training multi-step reasoning agents is ensuring the effectiveness of intermediate steps. Traditional supervised learning only focuses on the correctness of the final output and ignores the quality of the reasoning path. GeoMind introduces a fine-grained process supervision strategy: not only optimizing the final result but also providing feedback on each intermediate step—whether the trend description of the perception module is accurate, whether the hypothesis of the reasoning module is supported by sufficient evidence, and whether the constraint check of the analysis module correctly identifies geological irrationality. This strategy significantly improves the model's reliability and interpretability, allowing users to trace the entire reasoning chain to understand the decision logic.

The research team evaluated GeoMind's performance on four well logging benchmark datasets covering different geological environments and data characteristics:

• Dataset A: Conventional sand-shale profile with standard logging series
• Dataset B: Complex carbonate reservoir with secondary pores
• Dataset C: Thin interbedded deposits with high vertical resolution requirements
• Dataset D: Low-contrast lithology with high classification difficulty

The results show that GeoMind consistently outperforms strong baseline methods on all datasets. More importantly, it provides a transparent and traceable decision process, with each prediction accompanied by a complete reasoning path, allowing geologists to review and understand the AI's "thinking" process.

The interpretability of GeoMind has core application value:

1. Quality Control: When predictions are inconsistent with geologists' judgments, the source of divergence can be quickly located through the reasoning path (e.g., the perception module misunderstood the trend, the reasoning module ignored key evidence), enabling human-machine collaboration.
2. Knowledge Discovery: Analyzing reasoning patterns can reveal new geological laws (e.g., frequent use of a certain logging combination to identify lithology suggests its diagnostic value).
3. Education and Training: The reasoning process can serve as teaching material to help novices learn expert-level well logging lithology inference methods.

Additionally, this agent paradigm can be extended to other fields of earth science: seismic interpretation, reservoir characterization, geological hazard assessment, etc., providing a feasible path for "AI for Science".

Current GeoMind has the following limitations:

• Computational Cost: Multi-step reasoning is more time-consuming than single-step prediction, which may affect real-time applications.
• Knowledge Dependence: Performance is limited by the completeness of the built-in geological knowledge base.
• Generalization Ability: Performance may decline in geological environments outside the training data distribution.

Future research directions include:

• Integrating online learning to continuously improve the system from new logging data.
• Developing more efficient reasoning strategies to reduce unnecessary computations.
• Expanding the knowledge base to cover a wider range of sedimentary environments and lithology types.