AI-Assisted Mine Planning Prototype

The client is referred to throughout as "the mining company".

Project Status

Prototype accepted, pending validation against real-world site data.

Executive Summary

This project explored how artificial intelligence can accelerate early-stage mine planning for a mining company working with large geological block-model datasets.

The objective was to test whether an AI-assisted workflow could ingest a full mining dataset, convert it into a queryable planning model, generate candidate mining paths, and produce prototype life-of-mine planning outputs suitable for review by an experienced domain expert.

The prototype demonstrated that AI can support mine planning when paired with a structured data layer and deterministic engineering logic. Rather than asking a language model to read a large spreadsheet directly, the system converts the raw dataset into a database, performs calculations through controlled Python functions, and uses an LLM to interpret planning options, constraints, and trade-offs.

The current result is not a replacement for a geologist or mining engineer. It is a working proof of concept showing how AI can reduce manual analysis time, expose planning alternatives, and create a repeatable framework for expert review.

Business Problem

Mine planning workflows often rely on large spreadsheets, specialist software, and expert judgement built up over many years. A typical planning exercise can involve:

Reviewing large geological datasets
Identifying mineable material
Testing alternative mining sequences
Balancing grade, tonnage, access, constraints, and plant feed
Revising plans repeatedly as new assumptions are introduced

This creates a productivity challenge. Much of the work is highly technical, data-heavy, and repetitive, yet still requires expert oversight. The opportunity for AI is not to remove the expert, but to make expert review faster, more systematic, and easier to scale.

Project Objective

The goal was to build an MVP system capable of:

Loading a full mining dataset from a large CSV file
Storing the data in a fast local analytical database
Querying mineable blocks without exposing the entire raw dataset to an LLM
Generating provisional mining path options
Producing life-of-mine reports for expert review
Capturing domain feedback and turning it into explicit planning rules

The target use case was early-stage scenario generation: helping the mining company quickly see possible planning directions before committing significant human time to manual modelling.

Technical Approach

The system was designed around a simple principle: the AI should reason over structured results, not raw mining spreadsheets.

The prototype architecture includes:

Layer	Purpose
CSV ingest	Read the source mining dataset and load it into a structured database
DuckDB data layer	Store and query the block model locally without requiring a database server
Query functions	Calculate filtered inventories, sector summaries, candidate mining rectangles, and adjacent moves
Path planner	Generate short candidate mining paths using explicit rules and scoring
Life-of-mine planner	Extend path planning into full prototype schedules
Report generation	Produce Markdown and JSON outputs for expert review and future UI integration

This architecture keeps the large geological dataset out of the language model context window. The LLM only needs to interact with summarised, validated outputs.

Dataset Handling

The system loaded a full block-model CSV into DuckDB and created working views and summary tables.

The imported dataset contained:

67,765 block records
21,557 available blocks
13,948 topsoil blocks
Spatial coordinate fields for mining geometry
Heavy-mineral, slimes, oversize, density, elevation, and classification fields

For prototype planning, the active inventory was filtered to include material that satisfied the current planning assumptions:

Excludes already-mined material
Excludes low-confidence RESCAT = 3 material (resource category 3)
Excludes topsoil
Includes only blocks above the current elevation threshold

These rules can be changed as domain requirements are refined.

Planning Logic

The first planning model used a sector-based approach. Each candidate mining step was represented as a rectangular mining sector approximately 100 metres by 200 metres.

The planner evaluated each sector by aggregating the blocks inside it and calculating:

Number of blocks
Number of columns
Estimated tonnes
Weighted heavy-mineral grade
Average heavy-mineral grade
Elevation range
Movement distance
Planning score

Candidate paths were then generated by extending from one sector to nearby connected sectors. The planner also supported relocations, but domain feedback showed that this capability needed much stronger control.

Prototype Outputs

The system produced two main classes of output.

Candidate Path Outputs

The initial path planner generated multiple short-path alternatives. Each path included:

Starting location
Mining direction
Sector-by-sector movement
Tonnes per step
Weighted heavy-mineral grade per step
Elevation range
Movement distance
Overall path score

This demonstrated that the system could convert geological data into readable planning options.

Life-of-Mine Prototype Outputs

The life-of-mine planner expanded the process into full schedule-style reports.

For each scenario, the generated report included:

Run summary
Active assumptions
Plan-level summary
Sequence summary
Period-by-period mining schedule
Movement and relocation report
Feed quality profile
Inventory depletion summary
End-of-plan notes
JSON data output for later integration

The latest run generated five alternative life-of-mine schedules in approximately 35 seconds on a local workstation.

Domain Review and Lessons Learned

Expert review was critical. The system successfully produced mining schedules, but the review identified important gaps between computationally valid paths and operationally realistic plans.

Issues identified:

The alternative scenarios were too similar and mined much of the same material.
Some paths mined drainage-channel areas that are only available during dry months.
The planner allowed too many pit changes within a short mining period.

Operating rules learned:

New pit starts are expensive and require machine time that may not be available.
A mining unit is expected to mine approximately 12.5 million tonnes per annum.

This feedback is exactly the kind of domain intelligence the system is intended to capture. The first version proved the data and planning pipeline. The next version needs stronger operational constraints.

Productivity and Cost-Saving Potential

The project demonstrates several ways AI can create value for mining companies.

Faster Scenario Generation

A structured AI workflow can produce candidate mining sequences in seconds or minutes, allowing experts to review more options than would usually be practical.

Better Use of Expert Time

The system can handle repetitive filtering, aggregation, scoring, and report generation, allowing senior technical staff to focus on judgement, constraints, and strategy.

Repeatable Decision Logic

Tacit expert rules can be converted into explicit logic. This improves consistency and makes planning assumptions easier to inspect, refine, and transfer.

Reduced Manual Spreadsheet Work

Large spreadsheets can be moved into a database-backed workflow, reducing the risk of manual handling errors and making calculations easier to audit.

Improved Review Cycles

Each run creates structured outputs that experts can critique. Their feedback can then be folded back into the rules, improving the system over time.

Current Limitations

The current system is a proof of concept, not a production-grade mine planning system.

Known limitations include:

Drainage-channel seasonality is not yet modelled
Pit boundaries and starter-pit constraints require clearer input data
Batter angles are not yet incorporated
Plant throughput and production schedules are not fully modelled
Stockpiling and blending are not yet implemented
Economic cut-offs and cashflow modelling are not yet included
The planner still needs stronger scenario-diversity logic

These are solvable development tasks, but they require validated site data and expert-defined operating rules.

Outcome

The project reached a useful prototype milestone.

The system successfully:

Loaded and queried a full mining dataset
Converted raw block-model data into a local analytical database
Generated candidate mining paths
Produced prototype life-of-mine reports
Captured expert feedback about operational realism
Identified the next set of domain rules needed for a more useful production scenario planner

The prototype has been accepted as a valid starting point, pending validation against real-world site data.

Strategic Significance

This project shows how mining companies can begin using AI without handing critical planning decisions over to a black box.

The most promising model is a hybrid workflow:

Databases handle the data
Deterministic functions handle calculations
Domain experts define the rules
AI assists with scenario generation, explanation, reporting, and iteration

That approach can improve productivity while keeping expert control where it belongs.

For mining companies, the opportunity is not simply to automate planning. The larger opportunity is to make high-value technical expertise more scalable, auditable, and productive.

Recommended Next Phase

The next version should move from a broad life-of-mine demonstration toward a more operationally realistic first-year scenario planner, applied to validated real-world site data with known pit boundaries, drainage-channel definitions, production periods, and operating constraints.

The proposed changes are:

Generate genuinely different scenarios by starting in different pits or materially different mining zones
Cap each annual scenario at approximately 12.5 million tonnes
Keep each 12-month scenario within the starting pit unless explicitly allowed otherwise
Model starter-pit costs and machine-time limitations
Restrict drainage-channel mining to the dry-season window from 1 April to 30 November
Treat new pit starts as long-term decisions, more likely every 3 to 6 years rather than within a single annual plan
Add a clearer distinction between connected sector moves and major relocations

This would align the AI planner more closely with how mining operations actually make decisions, and refine the system into a first-year scenario planner that generates operationally realistic alternatives for expert review.