Sim2Schedule: A Simulator-Guided LLM Framework for Autonomous Open-Pit Mine Scheduling

Researchers introduce Sim2Schedule, an LLM-based framework that uses a simulator to guide autonomous decision-making for open-pit mine scheduling, achieving 94-99% of optimal performance compared to traditional MILP optimization while scaling linearly in computation time and operating entirely offline without fine-tuning.

Sim2Schedule represents a significant shift in how computationally intensive industrial optimization problems can be solved. Traditional MILP approaches, while mathematically optimal, struggle with real-time adaptation in dynamic environments due to exponential computational complexity. This research demonstrates that LLMs can function as practical decision-making agents when properly constrained by domain-specific simulators that encode hard operational rules directly into action generation, rather than relying on learned patterns.

The framework's key innovation lies in its hybrid approach: the simulator enforces geotechnical precedence, extraction-processing coupling, and capacity constraints at each decision step, while the LLM generates interpretable schedules without requiring cloud-based inference, domain-specific fine-tuning, or retraining. This zero-shot capability is particularly valuable for sensitive industrial operations requiring data security and regulatory compliance.

The performance metrics are compelling—recovering 94-99% of MILP-optimal net present value while scaling linearly represents a substantial practical improvement for resource-intensive industries. The novel MILP formulation developed as a benchmark adds methodological rigor. For mining operations, this translates to faster scheduling decisions that can adapt to real-time operational changes while maintaining near-optimal economic returns.

The broader implications extend beyond mining. This simulator-constrained LLM agent pattern could address similar long-horizon scheduling problems across manufacturing, logistics, and energy sectors. The research validates that LLMs excel not as standalone solvers but as reasoning engines when paired with deterministic constraint systems, offering a scalable alternative to computationally prohibitive optimization methods.

• →LLM-based framework recovers 94-99% of optimal performance from traditional MILP with linear scaling instead of exponential complexity
• →Simulator-constrained approach encodes hard operational rules directly into action generation without requiring fine-tuning or cloud inference
• →Zero-shot capability enables deployment in data-secure, offline environments suitable for regulated industrial operations
• →Hybrid architecture combining LLM reasoning with deterministic constraint systems offers practical alternative to classical optimization
• →Framework demonstrates broader applicability to long-horizon scheduling problems across manufacturing, logistics, and resource extraction