The Challenge
A major mining operator sought to evaluate reinforcement learning (RL) as a next-generation control strategy for their crushing circuit. While they had a functioning Simulink model of their crusher control system, it was not yet ready for integration with reinforcement learning platforms. The goal was to adapt the model for use within BONS.AI, enabling the development and testing of AI-driven control policies that could outperform traditional approaches.
However, the existing model lacked compatibility with RL workflows and required enhancements to reflect realistic plant behaviour. It also needed to incorporate an existing Model Predictive Controller (MPC) while maintaining site-specific control characteristics. Technical uncertainties included ensuring simulation fidelity during RL training and preserving key performance dynamics from the original control strategy.
The Solution
Opti-Num re-engineered the provided crusher simulation model to serve as a high-fidelity test bed for reinforcement learning experimentation. This was achieved through a structured development and integration process:
- Model Cleanup and Modularisation - The original Simulink model was reviewed, cleaned, and refactored using Variant Subsystems and Library blocks to support reusability and ease of experimentation.
- MPC Implementation - A Model Predictive Controller was implemented and calibrated to accurately reflect site control behaviour. This allowed the model to simulate realistic responses while also serving as a benchmark for RL comparison.
- Reinforcement Learning Integration - The final model was prepared to interface directly with BONS.AI, enabling it to serve as a training environment for reinforcement learning algorithms. Special care was taken to ensure compatibility while keeping the original model structure intact.
- Best Practice Simulink Design - Modular design principles were applied throughout to support maintainability and scalability. Generalised components were introduced for flexibility in future use cases or control strategies
The Outcome
The upgraded simulation model enabled the mining operator to explore AI-driven control development in a realistic, risk-free environment:
- Accelerated Innovation – Reinforcement learning algorithms could be tested without impacting live operations, supporting faster exploration of advanced control methods.
- Benchmarking Against Existing Control – With the MPC integrated into the model, RL performance could be directly compared to current site-level strategies.
- Enhanced Simulation Fidelity – The updated model provided a closer match to real-world behaviour, increasing confidence in experimental results.
- Reusable and Scalable Framework – The modular structure allows the model to be adapted to future scenarios, control techniques, or machine learning environments.
- Stakeholder Confidence – The delivery included technical documentation and structured components, ensuring transparency, interpretability, and ease off future handover.
By aligning control engineering with AI experimentation, Opti-Num delivered a simulation platform that bridges the gap between traditional process control and modern reinforcement learning—paving the way for more intelligent, adaptive operations in industrial mining environments.
.png)
.png)
