WA12 - Digitalization and Intelligent Production Planning Systems Transforming Industrial Mining Complexes and Mineral Value Chains

This study proposes an evolutionary framework that uses large language models (LLMs) to generate heuristics for mining truck dispatching. The framework evolves interpretable heuristics through LLM-driven crossover and mutation operators. Preliminary results show the effectiveness of LLMs for automated heuristic design, which can directly support broader optimization of extraction and destination scheduling in mining operations.

This work presents an AI-powered metaheuristic to solve stochastic mining complex production scheduling and sequencing extraction problems. The framework generates operationally viable diglines and utilizes simulation-based optimization to find near-optimal solutions. A detailed mine operations simulator evaluates these solutions, maximizing expected economic value while strictly enforcing physical and operational constraints.

The open-pit mine production scheduling problem is a large-scale, highly constrained mixed-integer optimization problem for which obtaining a high-quality feasible solution within reasonable computation time remains a major challenge in practice. This paper proposes a relax-and-repair framework designed to rapidly generate a feasible initial solution that can subsequently be improved by advanced heuristic methods. The proposed approach first computes a fast initial solution through a constraint relaxation strategy, where selected hard constraints are softened by introducing slack variables that are penalized in the objective function. This relaxation allows the model to be solved efficiently while preserving the core structure of the scheduling problem. To further accelerate solution construction and enhance temporal consistency across the planning horizon, a rolling-horizon heuristic is applied, progressively solving overlapping time windows and propagating decisions across the planning horizon. The resulting solution is then processed by a repair mechanism that systematically removes constraint violations and restores feasibility with respect to all original constraints. The repaired feasible solution serves as a high-quality starting point for improvement heuristics. In particular, we embed the proposed initialization within a large-neighborhood search (LNS) framework, where the availability of a strong initial solution significantly improves convergence speed and final solution quality. Computational experiments on large-scale industrial instances demonstrate that the proposed relax-and-repair framework produces feasible solutions orders of magnitude faster than direct exact approaches, while substantially improving the effectiveness of subsequent heuristic optimization. These results highlight the value of fast feasibility-driven initialization strategies for solving real-world open-pit mine production scheduling problems.

Effective mine planning and design are pivotal to project value, yet traditional methodologies often oversimplify critical elements like spatial cost variations and the opportunity cost of time. Standard sequential approaches rely on fixed costs per ton, failing to account for shared infrastructure or varying transport distances, which leads to suboptimal designs and inaccurate reserve estimations.
To address these limitations in underground mining, we propose an integrated framework coupling a Decomposed Cost Model (DCM) with a Reinforcement Learning (RL) algorithm. By optimizing stopes and developments simultaneously, our approach ensures cost estimations converge reliably. Building on this, we introduce a scheduling framework that learns to sequence extraction to maximize Net Present Value (NPV) while strictly adhering to capacity and operational constraints. This adaptable system can be tailored to various mining methods and specific site requirements.
Our results demonstrate that this AI-driven methodology effectively navigates complex planning problems. By transitioning to more accurate cost modeling and leveraging machine learning, we achieve a substantial increase in both recoverable reserves and NPV. This research highlights the potential of RL to overcome the rigidity of traditional models, delivering high-quality, fully feasible schedules in reasonable time.