What's the lowest-cost path to an electric mining haul fleet?

A fleet transition is a capital sequence. Order and timing decide the outcome: chargers have to exist before the trucks that need them, a truck type that is right in year five can be stranded by year fifteen, and a diesel unit retired early leaves salvage value on the table while one retired late keeps burning fuel against the interim target.

A per-truck net present value in a spreadsheet cannot see any of that. Queues at a charging station, the interaction between haul cycle length and battery range, the way a trolley segment changes which routes are worth electrifying: these are behaviors of the whole system, and a spreadsheet row has no place for them.

The company wanted the combination of four electric truck types, three recharging technologies, rail, and conveyor that reached its emissions targets at the lowest acquisition and lifecycle cost over 30 years, with the diesel retirement rate and the electric introduction rate as inputs its own planners set.

The pilot had three parts, and the first two existed only to make the third possible.

• Speed. Cut the existing model's run time to under two minutes per simulated year on average, across a 30-year horizon, without losing the fidelity that made the model credible.
• Integration. Put an AI optimizer in the loop with the model. Microsoft's Project Bonsai would propose a transition plan, the model would run it and return cost, emissions, and haulage, and the optimizer would learn from the result and propose again.
• The answer. Run the combinations of truck types, charging technologies, rail, and conveyor against the company's chosen retirement and introduction rates, and find the plan that minimized lifetime cost while hitting the carbon targets.

The company's technology and innovation team owned the model throughout. SimWell's job was to make it fast enough to optimize and to wire it to the optimizer.

Speed came from subtraction. The team deactivated model logic block by block, measured the effect on run time, and kept only the detail that moved the answer. Truck transport times and counts, the diesel-to-electric transition logic, truck additions and removals, loading and hauling unit capacity, speed, number, and location, and maintenance and downtime all stayed. Detail that cost seconds without changing a decision went.

Integration meant teaching the model to talk. After each iteration the model handed the optimizer four numbers for the plan it had just run: capital cost, operating cost, annual greenhouse gas emissions, and total haulage for the year. The optimizer compared plans on cost outside the model and sent back the next plan to test. The company's team selected the scenarios; the loop did the searching.

Microsoft retired Project Bonsai in late 2023. The method it served does not depend on the product. A simulation fast enough to act as a training environment lets a reinforcement learning agent try thousands of plans at a cost of seconds each, which is what makes a 30-year, multi-technology capital decision searchable at all. The same loop runs today on current reinforcement learning tooling.

Run time per simulated year fell from 160 seconds to 9 seconds, a 30-year scenario from more than an hour to under five minutes. The integration with the optimizer was validated, with the model returning CapEx, OpEx, annual emissions, and haulage after every iteration. Against the company's chosen diesel retirement rate and electric introduction rate, the loop returned an optimized count and type of electric trucks and charging stations across the 30-year horizon.

The pilot did what a pilot is for. The company went on to commission a broader open-pit emissions reduction study on the same foundation, covering alternatives across multiple mines.