Open-Pit Ultimate Pit Limit and Production Scheduling Using Max-Flow, Genetic Algorithm and Reinforcement Learning

Open-pit planning first finds the ultimate pit limit - the most valuable set of blocks that respects the wall slopes

- and then decides in which year each block is mined to maximise the net present value (NPV) under mining and processing capacities. This project solves four real MineLib block models (newman1 (1,060 blocks), zuck_small (9,400 blocks), kd (14,153 blocks) and marvin (53,271 blocks)) on a standard laptop CPU, 16 GB RAM: the exact ultimate pit by max-flow and pseudoflow, nested pits with Whittle-style best and worst-case schedules, a Lagrangian upper bound by max-flow, the TopoSort heuristic, a genetic algorithm and a PPO reinforcement-learning scheduler on about 300 bench-polygons, every schedule checked by one evaluator against MineLib's published LP bound. Max-flow = pseudoflow = MineLib on 4 of 4 instances. Gap to MineLib's LP bound, GA / PPO / best case: newman1 3.84 / 3.85 / 4.54 %; zuck_small 9.78 / 10.65 / 4.52 %; kd 3.28 / 3.77 / 3.30 %; marvin 8.73 / 8.82 / 5.31 %. Best own method: newman1 GA (best seed) (3.62 %), zuck_small Best case, UPIT (4.52 %), kd TopoSort (1.72 %), marvin Best case, UPIT (5.31 %). The literature schedule SD-CPIT (2026) stays best on every instance (newman1 1.26 %, zuck_small 0.83 %, kd 0.10 %, marvin 0.89 %). 13 of 20 spec targets met: max-flow = pseudoflow = MineLib on 4 of 4 instances; worst case 10-30 % below best NOT MET on newman1 and kd; GA within 5 % of best case NOT MET on zuck_small; PPO 1-5 % gap NOT MET on zuck_small and marvin. PPO never beat the fixed GA plan on the 20 test scenarios of any instance. Research and teaching prototype on public benchmark block models: not a mine plan for a real operation.

Category: Mining Engineering

Mining Engineering Project MN07 from ProTechSC Academy, with graphical abstract, video demonstration and project guidance.