For a PDF of the article, click here.
How AI and Reinforcement Learning Transforms Ore Blending for Smarter Gold Processing
By Tenzin Worden, Data Scientist | Eclipse Mining Technologies
Declining grades and rising processing costs are reshaping the economics of gold mining. Management across the industry faces a pressing question: how can operations ensure maximum recovery without driving up energy consumption or overextending margins? Traditional methods of mine planning and grade control help, but when it comes to ore blending and stockpile management – the processes that directly affect plant performance – most operations still lack real-time optimisation tools.
Ore blending isn’t just a logistical necessity; it directly determines mill throughput, recovery efficiency and energy use, according to Tenzin Worden, Data Scientist at Eclipse Mining Technologies. The right blend can boost grinding efficiency and improve leaching performance, whereas poor blending can result in lower recoveries, wasted energy and lost revenue. Similarly, well-managed stockpiles act as buffers between the mine and the mill, smoothing out variability in feed quality and increasing operational stability.
Historically, blending decisions have relied on average grades or simplified models, leaving operations vulnerable to variability in mineralogy and ore quality, Worden says. The complexity of interactions between gold grade, sulphur content and recovery efficiency mean averages alone are rarely accurate predictors of these relationships. To establish a solid baseline, the process must begin with trusted sampling at the plant entry point. Yet, even when good samples exist, data integrity is not guaranteed. It is common to encounter problematic data, such as samples collected during plant downtime, misread instruments, or values that fall well outside the expected range. Without careful validation, such data can distort reconciliation and undermine decision making.
This is where AI-enabled validation becomes essential, according to Worden.
“The rise of sensors, cloud computing and AI has opened a new frontier in mining,” she says. “Instead of relying solely on static models, operations can now access real-time predictive insights that adapt to changing conditions. Big data and AI don’t just provide numbers; they create generalised solutions from patterns in the data, helping miners optimise decisions as ore properties shift.”
But not all AI is created equal. Generalpurpose models, such as large language models, may dominate headlines, yet they often lack the precision required for domain-specific challenges in mining. Instead, the real breakthroughs lie in combining statistical modelling, simulation and reinforcement learning (RL), a trio of methods that together can represent complex geometallurgical systems and optimise them in real time.
Worden explains: “Statistical modelling captures the relationships between variables, such as the interaction between sulphur content and gold recovery. Rather than treating grades as static, it uses probability distributions to reflect the real variability in ore properties.
“Simulation builds on these models, creating digital twins of processing systems where operators can test thousands of ‘what if’ scenarios. How would a different blend impact cyanide leaching efficiency? What would happen to recovery if a certain stockpile were prioritised? Simulation provides answers without the costly trial and error that often occurs in the real world.
“RL takes the final step, using trialand-error feedback to continuously refine blending strategies. Much like a plant operator learning through experience, RL agents experiment with different blending decisions, receive input on recovery and energy outcomes, and improve their strategies over time. Unlike traditional optimisation methods that rely on fixed averages, RL adapts dynamically as grades, mineralogy and plant conditions evolve.”
Conventional tools, such as linear programming, are rigid in their application, according to Worden, requiring explicit mathematical models, assuming steadystate conditions and often overlooking the inherent variability in mining. A weekly blending plan may appear optimal on paper, but, in practice, it often fails to account for deviations in ore grade, unexpected equipment downtime, or shifts in operational productivity.
“RL, by contrast, thrives in dynamic environments,” she says. “It doesn’t need explicit instructions on the ‘right’ action; it learns from experience, guided by rewards and penalties. This evaluative approach makes it uniquely suited for mining, where real-world conditions rarely align with theoretical models.”
In gold mining, blending is particularly complex because recovery is not “additive” in nature. Different mineralogical and grade combinations can interact in ways that dramatically change leaching performance. Statistical modelling and simulation allow operators to explore these non-linear relationships, while RL identifies strategies that maximise recovery over time rather than in isolated scenarios.
“By combining these techniques, miners can shift from static, average-based planning to adaptive, real-time decision making,” Worden says. “The result is a system that not only predicts recovery outcomes with greater accuracy but also adjusts strategies as conditions change – something traditional methods cannot achieve.”
The implications for gold processing are significant. Optimised blending can increase recovery, lower energy use and reduce costs per ounce. More importantly, it transforms blending and stockpile management from underused processes into strategic levers for profitability.
“At the centre of this transformation is the integration of statistical modelling, simulation and reinforcement learning into unified knowledge systems,” Worden says. “By bridging geology and metallurgy, these systems enable miners to predict, adapt and optimise across the entire value chain.”
She concluded: “One such solution, SourceOne EKPS, applies these advanced AI methods to gold operations. By turning raw data into actionable intelligence, it enables predictive modelling and optimisation that closes the loop between mine and mill. The result: higher recovery rates, lower costs and a more resilient operation in the face of declining grades and rising energy demands.”