Deep Learning Decision Support System for Open-Pit Mining Optimisation: GPU-Accelerated Planning Under Geological Uncertainty

📝 Abstract

This study presents Part II of an AI-enhanced Decision Support System (DSS), extending Rahimi (2025, Part I) by introducing a fully uncertainty-aware optimization framework for long-term open-pit mine planning. Geological uncertainty is modelled using a Variational Autoencoder (VAE) trained on 50,000 spatial grade samples, enabling the generation of probabilistic, multi-scenario orebody realizations that preserve geological continuity and spatial correlation. These scenarios are optimized through a hybrid metaheuristic engine integrating Genetic Algorithms (GA), Large Neighborhood Search (LNS), Simulated Annealing (SA), and reinforcement-learning-based adaptive control. An ε-constraint relaxation strategy governs the population exploration phase, allowing near-feasible schedule discovery early in the search and gradual tightening toward strict constraint satisfaction. GPU-parallel evaluation enables the simultaneous assessment of 65,536 geological scenarios, achieving near-real-time feasibility analysis. Results demonstrate up to 1.2 million-fold runtime improvement over IBM CPLEX and significantly higher expected NPV under geological uncertainty, confirming the DSS as a scalable and uncertainty-resilient platform for intelligent mine planning.

💡 Analysis

This study presents Part II of an AI-enhanced Decision Support System (DSS), extending Rahimi (2025, Part I) by introducing a fully uncertainty-aware optimization framework for long-term open-pit mine planning. Geological uncertainty is modelled using a Variational Autoencoder (VAE) trained on 50,000 spatial grade samples, enabling the generation of probabilistic, multi-scenario orebody realizations that preserve geological continuity and spatial correlation. These scenarios are optimized through a hybrid metaheuristic engine integrating Genetic Algorithms (GA), Large Neighborhood Search (LNS), Simulated Annealing (SA), and reinforcement-learning-based adaptive control. An ε-constraint relaxation strategy governs the population exploration phase, allowing near-feasible schedule discovery early in the search and gradual tightening toward strict constraint satisfaction. GPU-parallel evaluation enables the simultaneous assessment of 65,536 geological scenarios, achieving near-real-time feasibility analysis. Results demonstrate up to 1.2 million-fold runtime improvement over IBM CPLEX and significantly higher expected NPV under geological uncertainty, confirming the DSS as a scalable and uncertainty-resilient platform for intelligent mine planning.

📄 Content

Long-term open-pit mine planning seeks to optimize the extraction and processing sequence throughout a mine’s life while managing geological, operational, and market uncertainties (Armstrong et al., 2021;Quelopana & Navarra, 2024;Tolouei et al., 2021). Metallurgical plants serve as critical components in the mineral value chain, as their efficiency directly influences the overall economic performance of mining operations. Nevertheless, complete recovery of valuable minerals such as gold and silver is rarely achievable. Attaining total extraction would require excessive use of chemical reagents, particularly cyanide, leading to higher operational expenses and diminishing overall efficiency, as some valuable minerals inevitably remain in the tailings (Ordenes J et al., 2021). Additionally, the comminution stage, which involves crushing and grinding the ore to achieve finer particle sizes (Gupta CK, 2003), is among the most energy-demanding phases of mineral processing. It typically accounts for 30%-60% of a mine’s total energy consumption and can reach up to 80% in certain operations (Wang C et al., 2013;Lois-Morales P et al., 2022).

Plant reconfiguration should be treated as a strategic decision aligned with future processing requirements, not as a reaction to minor ore feed fluctuations (Navarra A et al., 2017). An operational mode represents a defined set of parameters that outline a metallurgical plant’s configuration, including its metal recovery rates and processing capacities (Quelopana A et al., 2023). Each operational mode is tailored to maximize metal or alloy output while maintaining safe and efficient processing for distinct ore blends or rock types. Given the significant variability in metal grades across geological materials, operating multiple modes, where technically and economically viable, can substantially enhance plant adaptability. This operational flexibility allows for better adjustment to fluctuating ore characteristics and provides a more robust mechanism for managing uncertainty associated with grade variations (Quelopana A et al., 2023).

The surge in geological, operational, and market data demands innovative methods for efficient processing to support more accurate and adaptive mine planning. Traditional mine planning frameworks relying on deterministic models and fixed inputs fail to capture the complexity and variability of real-world geological and operational conditions. Modern approaches that employ a limited number of geological scenarios-typically between 10 and 20 static realizations-offer an insufficient statistical basis to capture the virtually infinite range of possible grade distributions within actual ore deposits. This limitation results in overly simplified representations of uncertainty and weakens the reliability of risk assessments. As the mining sector advances toward digital transformation and deeper integration of artificial intelligence, it becomes increasingly essential to adopt advanced optimization frameworks that harness machine learning for geological modeling, ensuring both computational efficiency and responsiveness to dynamic, evolving conditions. Multi-agent scheduling has been applied widely in optimisation problems. Zhang et al. (2022) presents a multi-agent scheduling framework that integrates Deep Reinforcement Learning (DRL) with Proximal Policy Optimization (PPO) to handle dynamic and uncertain manufacturing environments. Experiments shows that the proposed PPO-based system outperforms traditional genetic programming and DQN approaches in convergence speed, workload balance, and scheduling efficiency for flexible job-shop operations. Recent advances in deep learning, particularly Variational Autoencoders (VAEs), offer transformative potential for geological uncertainty modelling by learning complex spatial patterns in geological data and generating realistic scenario sets that preserve geological continuity and spatial correlations. Combined with Graphics Processing Unit (GPU)accelerated computing, these AI-enhanced approaches can address the fundamental limitations of static scenario-based planning while maintaining real-time computational performance. Barkalov&Lebedev (2017) proposed a nested optimization approach for solving multidimensional global optimization problems using both CPU and GPU resources. A complex serial algorithm runs on the CPU for global control, while a simple parallel algorithm executes on the GPU for local search. Implemented in the ExaMin solver, this hybrid scheme achieves significant computational speedup and efficiency in benchmark tests. GPU-accelerated optimization has gained momentum across various domains, from mining optimization to medical applications such as federated learning frameworks for disease detection (Mukhopadhyay et al., 2025), though application to AIenhanced geological modeling in constrained mine scheduling remains underexplored. With the capability to execute thousands of threads in parallel,

This content is AI-processed based on ArXiv data.