AI for Mine Planning & Production Optimization

Block Model Optimization

Every open pit mine plan starts with a block model and a question: what is the optimal pit shell? The Lerchs-Grossmann algorithm and its floating cone variants have solved this for decades — given block values (grade × recovery × price − mining cost − processing cost), find the ultimate pit limit that maximizes undiscounted profit.

AI enters when the problem gets more realistic. The ultimate pit is a theoretical endpoint. What matters operationally is the sequence of nested pit shells (pushbacks) that maximizes NPV over a 15-30 year mine life, subject to:

This is a combinatorial optimization problem that grows exponentially with the number of blocks and time periods. Classical mixed-integer programming (MIP) solves small cases but hits computational limits on real block models (500K-5M blocks).

AI-Driven Approaches

The practical hybrid: use a GNN to learn block value dependencies and reduce the problem dimensionality, then solve the reduced MIP. This has shown 95-98% of optimal NPV at 10-50× faster solve times on real-world block models.

Open data/block-model.csv in the code panel. Each row is a mining block with coordinates (easting, northing, elevation), tonnage, grades (Fe, SiO2, Al2O3), rock type, and density. This is the raw input for pit optimization.

Production Scheduling Under Constraints

Seasonal and Wet-Season Impact

Mining operations face seasonal constraints that generic mine planning software often ignores. In tropical and subtropical operations (Pilbara cyclone season, West African and South American wet seasons, Canadian/Nordic winter freeze), production can drop sharply or halt for weeks to months — waterlogged pits, impassable haul roads, slope stability risk, or sub-zero equipment limits. This is rarely a gradual slowdown — it is often a hard stop driven by safety and environmental permit conditions.

A production schedule that assumes a steady 12-month profile will fail. AI-based schedulers must encode the seasonal envelope as a hard constraint:

For each year y, for each month m:
  if m in seasonal_shutdown_window:
    production_tonnes[y][m] = 0  (or reduced_capacity for covered/protected operations)
    waste_stripping[y][m] = 0
  Constraint: annual_target must be met in available operational months

This concentrates production into fewer months, requiring higher monthly throughput than a flat schedule — which cascades into equipment sizing, workforce planning, and processing plant capacity. Australian WHS and Resources Safety regulators, and Canadian provincial authorities, treat wet-season and winter haul-road controls as mandatory.

Grade Blending

Open data/production-schedule.json — it contains a multi-year schedule with monthly targets, source benches, and grade specifications.

Processing plants need consistent feed. An iron ore beneficiation plant designed for 62% Fe feed cannot handle 55% Fe ore without losing recovery, or 66% Fe ore without overwhelming the concentrate dewatering circuit. The blending problem:

Minimize: deviation from target feed grade
Subject to:
  - Total tonnes from all sources = plant capacity
  - Grade_min ≤ weighted_average_grade ≤ Grade_max
  - Each source bench has limited exposed inventory
  - Haul distances vary by source (affects cost and fleet requirement)

AI-based blending optimizers solve this in real-time, adjusting shovel-truck allocation as actual grades deviate from the block model estimate. At major thermal coal and metallurgical coal operations (Glencore, BHP, Teck), real-time blending has reduced product quality variance by 30-40%, directly improving customer specification compliance and power plant efficiency.

Equipment Fleet Optimization

Open data/equipment-fleet.json — it contains fleet composition, maintenance schedules, fuel consumption curves, and availability data.

Dispatch Optimization

Shovel-truck dispatch is the minute-by-minute decision engine of an open pit mine. Classical dispatch sends the next available truck to the best shovel based on a priority rule (shortest queue, highest priority material). AI dispatch considers the full system state:

Reinforcement learning dispatch trains an agent in a simulation of the mine's truck-shovel system. The agent learns policies that outperform heuristic dispatch by 8-15% in tonnes moved per shift. The key insight: RL discovers non-obvious strategies like intentionally idling a shovel to prevent downstream bottlenecks, or routing trucks on longer paths to avoid congestion at a junction.

Autonomous Haulage

The leading edge of fleet optimization is fully autonomous haulage. Rio Tinto's Autonomous Haulage System (AHS) operates one of the world's largest driverless truck fleets across the Pilbara, and BHP, Fortescue, Suncor (oil sands), and Vale run autonomous fleets built on Caterpillar Command and Komatsu FrontRunner platforms. Autonomy turns dispatch into a pure optimization problem — no operator variability, continuous operation across shift changes, and telemetry on every vehicle. AI dispatch and autonomous haulage compound: an RL dispatcher controlling an autonomous fleet can hold tighter cycle times than any human-operated equivalent.

Fuel Efficiency

Haul trucks are the largest single operating cost in open pit mining — fuel alone can be 25-35% of total mining cost. AI-based fuel optimization targets:

At a large iron ore complex operating a fleet of 35 Caterpillar 793F trucks and 8 Komatsu PC4000 shovels, implementing AI-based dispatch and fuel optimization improved fleet productivity by 12% and reduced diesel consumption by 9% — a saving of several million USD annually.

Haul Route Optimization

Haul roads evolve as the pit deepens. The in-pit road network must be re-designed every 6-12 months as new benches open and old ones are mined out. AI-based route optimization considers:

Graph-based optimization with learned edge costs (accounting for gradient, surface quality, and traffic density) produces road networks that reduce average cycle time by 10-20% compared to manually designed roads.

Global Operational Context

Major Diversified Miners (BHP, Rio Tinto, Glencore, Vale, Anglo American)

The world's largest miners move billions of tonnes annually across hundreds of operations. The scale makes even small efficiency gains enormously valuable:

Iron Ore Majors (Pilbara & Labrador Trough)

Rio Tinto and BHP operate the world's largest mechanized iron ore mines in the Pilbara; Vale, Champion Iron, and Tacora operate in the Labrador Trough:

Caterpillar and Komatsu Fleets

Global mines predominantly run Caterpillar (dump trucks, dozers) and Komatsu (excavators, shovels, trucks), with Liebherr and Hitachi in the mix. Fleet-specific AI models must account for:

Key Takeaways