GeoAgent-MVP

A human-supervised geological modelling agent that inspects borehole data, validates it with deterministic geological rules, pauses for human approval, and then builds an implicit 3D model with GemPy.

Example cross-section generated from nine synthetic boreholes. The workflow converted logged intervals into contact points, fitted surface orientations, and constructed the model only after human approval.

Scope: GeoAgent-MVP is a research and portfolio prototype. It demonstrates a supervised agentic workflow over a real geological modelling engine; it is not a production geomodelling platform.

Why this project exists

Engineering agents should not allow a language model to invent calculations, silently repair data, or bypass consequential decisions.

GeoAgent-MVP separates orchestration, explanation, and computation:

The LLM does not calculate geological values or decide whether invalid data should be accepted. The complete workflow runs without an API key by using a deterministic narration fallback.

Related MCP server: DeepMap AI MCP Server

Workflow

Borehole CSV
     |
     v
Inspect data -------------------- unreadable input ----> halt
     |
     v
Validate geology ---------------- blocking errors -----> halt + report
     |
     v
Propose configuration
     |
     v
+--------------------------------------------------+
| HUMAN APPROVAL GATE                              |
| approve / modify / cancel                        |
| LangGraph interrupt() persists the graph state   |
+--------------------------------------------------+
     |
     | approve
     v
Revalidate before build --------- new errors ----------> halt + report
     |
     v
Build GemPy model --------------- tool failure --------> report; invent nothing
     |
     v
Optional uncertainty ensemble
     |
     v
Markdown report + figures + audit trail

The approval gate is a real LangGraph interrupt(). The graph pauses before model construction, and its state is stored by a checkpointer. Nothing downstream runs until a human decision resumes the same thread.

Geological validation

The validator distinguishes between a blocking data fault and a geological warning requiring interpretation.

Blocking errors

Warnings for human review

Lithology identifiers are normalized internally. For example, Unit-A, UNIT_A, and unit a map to the stable key unit_a.

From boreholes to GemPy

GemPy requires surface points and orientations rather than interval tables. GeoAgent-MVP performs this conversion explicitly.

1. For consecutive intervals where unit T overlies unit U, the shared boundary is extracted at:

   Z = z_collar - contact_depth

2. The contact is assigned to the upper unit T, because it represents the base boundary of that unit in the GemPy structural model.

3. The deepest unit is represented as basement rather than as another interpolated surface.

4. Where measured structural orientations are unavailable, a plane is fitted through each surface's contact points and its pole is used as the model orientation.

The fitted orientation is an assumption, not a field measurement. It is shown in the approval configuration and recorded in the final report.

Uncertainty analysis

The optional ensemble perturbs contact elevations using Gaussian noise, rebuilds the model for multiple realizations, and calculates normalized information entropy for each grid cell:

H = -sum_k(p_k * log(p_k)) / log(K)

• H = 0: all realizations agree.

• Larger H: greater disagreement between realizations.

• The entropy figure highlights the areas most sensitive to contact-elevation uncertainty.

This analysis propagates uncertainty in contact elevations only. It does not represent uncertainty in structural interpretation, fault geometry, or model topology.

Example entropy section. The displayed section is selected where the ensemble contains the strongest uncertainty signal.

MCP interface

mcp_server/server.py exposes GeoAgent through the Model Context Protocol.

Resources

• borehole://datasets

• borehole://dataset/{name}

Tools

• inspect_dataset

• validate_dataset

• propose_modelling_config

• build_model

• analyse_uncertainty

Prompt

• supervised_validate_and_model

The MCP layer confines file access to the data/ directory, rejects unsafe filenames, validates datasets before modelling, and requires explicit approval for consequential modelling and uncertainty operations.

Launch the MCP Inspector with:

npx -y @modelcontextprotocol/inspector python mcp_server/server.py

On Windows, using the virtual-environment interpreter explicitly is also supported:

npx -y @modelcontextprotocol/inspector .venv\Scripts\python.exe mcp_server\server.py

Quick start

Windows Command Prompt

git clone https://github.com/<your-username>/geoagent-mvp.git
cd geoagent-mvp

python -m venv .venv
.venv\Scripts\activate.bat

python -m pip install --upgrade pip
pip install -r requirements.txt
python data\make_datasets.py

Linux or macOS

git clone https://github.com/<your-username>/geoagent-mvp.git
cd geoagent-mvp

python -m venv .venv
source .venv/bin/activate

python -m pip install --upgrade pip
pip install -r requirements.txt
python data/make_datasets.py

Run the workflow

Interactive approval:

python run_demo.py --dataset data/valid_boreholes.csv

Approved run with uncertainty:

python run_demo.py \
  --dataset data/valid_boreholes.csv \
  --approve \
  --realisations 10

Approved run without uncertainty:

python run_demo.py \
  --dataset data/valid_boreholes.csv \
  --approve \
  --no-uncertainty

Blocking-data demonstration:

python run_demo.py --dataset data/incomplete_boreholes.csv

Cancellation demonstration:

python run_demo.py \
  --dataset data/valid_boreholes.csv \
  --cancel

An ANTHROPIC_API_KEY is optional. Without it, the explanation node uses a deterministic fallback while all modelling behaviour remains unchanged.

Outputs

A successful run writes artifacts under outputs/, including:

outputs/
├── figures/
│   ├── model cross-section
│   └── uncertainty entropy section
└── reports/
    └── timestamped Markdown modelling report

The report records:

• dataset summary;

• validation errors and warnings;

• normalized stratigraphic order;

• proposed modelling extent and resolution;

• assumptions presented to the human supervisor;

• approval decision;

• GemPy model metadata;

• uncertainty statistics;

• generated figure paths;

• tool-call audit history and timings;

• limitations and outstanding warnings.

Testing

Run the complete suite:

pytest -q

The current test suite contains 29 tests covering:

• valid and invalid schema handling;

• missing coordinates and depths;

• duplicate and overlapping intervals;

• lithology normalization;

• unknown units;

• stratigraphic conflicts and sparse support;

• contact extraction and fitted orientations;

• proposed extent;

• workflow halting;

• genuine human interruption;

• approval, modification, invalid modification, and cancellation;

• audit history;

• repeated-run reproducibility;

• explicit model failure handling;

• rejection of unrecognized approval decisions.

Before publishing a release, record the actual local result rather than a hard-coded historical timing:

29 passed in <environment-dependent time>

Repository layout

geoagent-mvp/
├── run_demo.py
├── requirements.txt
├── README.md
├── SETUP.md
├── data/
│   ├── make_datasets.py
│   ├── valid_boreholes.csv
│   ├── incomplete_boreholes.csv
│   └── inconsistent_boreholes.csv
├── geoagent/
│   ├── graph.py
│   ├── state.py
│   ├── schemas.py
│   └── explain.py
├── tools/
│   ├── data_inspection.py
│   ├── geological_validation.py
│   ├── gempy_modelling.py
│   ├── uncertainty.py
│   └── report_generator.py
├── mcp_server/
│   └── server.py
├── outputs/
│   ├── figures/
│   └── reports/
└── tests/
    ├── test_validation.py
    └── test_agent_workflow.py

Limitations

• The prototype assumes one conformable stratigraphic series.

• Faults, unconformities, intrusions, and overturned structures are not modelled.

• Orientations are inferred from contact geometry when measurements are unavailable.

• The included datasets are synthetic.

• Real projects commonly require separate collar, survey, and lithology tables.

• Deviated boreholes require desurveying, which is not implemented.

• The uncertainty ensemble perturbs contact elevation independently and is not a Bayesian inversion.

• Model quality still depends on data coverage, geological assumptions, and human interpretation.

Planned extensions

• ingest separate collar, survey, and lithology tables;

• desurvey deviated boreholes;

• accept measured dip and azimuth;

• use fitted orientations only as a documented fallback;

• add spatially correlated contact perturbations;

• support faults and unconformities through GemPy structural groups;

• add a lightweight web interface for reviewing the approval payload;

• persist production runs in a durable database-backed checkpointer.

Responsible-use statement

GeoAgent-MVP is a decision-support prototype. Human approval does not guarantee geological correctness, and model outputs should not be used for engineering design, safety-critical decisions, resource estimation, or investment decisions without qualified professional review and validation against project data.

Licence

MIT.