Decision-focused neural adaptive search and diving for optimizing mining complexes

Abstract

Optimizing industrial mining complexes, from extraction to end-product delivery, presents a significant challenge due to non-linear aspects and uncertainties inherent in mining operations. The two-stage stochastic integer program for optimizing mining complexes under joint supply and demand uncertainties leads to a formulation with tens of millions of variables and non-linear constraints, thereby challenging the computational limits of state-of-the-art solvers. To address this complexity, a novel solution methodology is proposed, integrating context-aware machine learning and optimization for decision-making under uncertainty. This methodology comprises three components: (i) a hyper-heuristic that optimizes the dynamics of mining complexes, modeled as a graph structure, (ii) a neural diving policy that efficiently performs dives into the primal heuristic selection tree, and (iii) a neural adaptive search policy that learns a block sampling function to guide low-level heuristics and restrict the search space. The proposed neural adaptive search policy introduces the first soft (heuristic) branching strategy in mining literature, adapting the learning-to-branch framework to an industrial context. Deployed in an online fashion, the proposed hybrid methodology is shown to optimize some of the most complex case studies, accounting for varying degrees of uncertainty modeling complexity. Theoretical analyses and computational experiments validate the components’ efficacy, adaptability, and robustness, showing substantial reductions in primal suboptimality and decreased execution times, with improved and more robust solutions that yield higher net present values of up to 40%. While primarily grounded in mining, the methodology shows potential for enabling smart, robust decision-making under uncertainty.