Improving milling throughput performance at a large-scale copper mine through the development of a mine-to-mill analytics solution
Rovjok led the co-development of a mine-to-mill data analytics solution with a large-scale, top-tier copper mine located in Latin America. The project consisted of a large data integration and standardisation exercise, combining systems from across mine and process plant into a single analytical model. On top of this, Rovjok led the generation and deployment of machine learning mill throughput models, which achieved 90%+ accuracy and were successfully used to forecast future mine operational performance, and provide detailed reviews into historical impacts on throughput from across geology, mine, and process plant operations. The project led to a 5-10% improvement in mill throughput and ~10% reduction in mill power draw, boosting the mine’s revenue significantly. The project also provided lasting improvements to in-house analytical capabilities across multiple teams, including D&B, and mill process control.
Rovjok was chosen as a partner to co-develop a mine-to-mill analytics solution with a large-scale tier 1 copper miner operating in Latin America. The recently commissioned operation is continually scaling to become one of the largest copper mines globally by processing capacity.
Plant performance is greatly impacted by ore properties (for example, it is well understood that milling rates are highly dependent on factors such as ore hardness and fragmentation). To analyse and optimise operational performance requires a rigorous understanding of the relationship between ore feed properties and rates. Yet, mine and processing data is usually contained in many disparate data systems, with poor lineage between them. Further, even if lineage exists between the data, the interdependent relationships can be very complicated and difficult to effectively analyse by normal Excel desktop methods. These make it challenging to draw a lineage between ore properties and plant performance.
At our client’s Latin American site, the processing plant operators experienced high variability in daily mill throughput rates, but had limited understanding of the underlying reasons. The upstream process flowsheet is shown in Figure 1 and consisted of in-pit explosive blasting, in-pit primary crushing, followed by optional secondary crushing, which led to a large run-of-mine stockpile. The stockpile fed directly into the semi-autogenous (SAG) mills, which each fed two ball mills. Pebbles filtered at the mills were returned to the secondary crusher circuit for either re-crushing or direct return to the stockpile.
Our client’s data across the mine and process plant were contained in six different systems (covering the geological block model, the designed and actual drill & blast patterns, load & haul tracking, site weather stations, and hundreds of process plant sensors), making any analysis a very time-consuming and one-off process, which resulted in very little progress being made without external assistance.
The myriad of variable parameters upstream of the mills, from ore geology, to blast parameters and resulting particle size distributions (PSDs), to crushing settings, to stockpile blending, meant that simple analyses such as the existing linear regression model, were incapable of fully disentangling the multidimensional relationships between upstream decisions and mill throughput performance. The client therefore wanted to explore more powerful analytical capabilities, including the use of machine learning models, to illuminate causal relationships and therefore drive more informed decision making to increase mill throughput.
Rovjok’s chosen framework for this mine-to-mill co-development was the team data science project, which is designed for efficiently managing advanced analytics projects and facilitates collaboration between all project stakeholders. It is designed specifically for predictive analytics data science projects, and covers: (i) business needs, (ii) data acquisition, (iii) modelling, and (iv) deployment. This process is summarised in Figure 2.
Each component of the framework was tackled in turn:
- The key business questions that the project aimed to answer were clearly defined and agreed upon, as well as the key deliverables to be generated through the project,
- The data systems were fully mapped, with thorough data quality analysis reports produced, and additional data requirements determined before any engineering and analysis were begun,
- The data modelling stage included all data integrations as well as the extensive, iterative machine learning model development process,
- Once models were prepared and tested, the deployment of all systems and deliverables into the client’s production environment was executed, and full documentation prepared for handover.
The tangible outcomes for the client were to be as follows:
- A master data table of all parameters located upstream of the mills, to be used as input to the machine learning modelling process.
- An accurate (~90-95%) machine learning throughput prediction model for each SAG mill,
- An automated pipeline for producing ML predictions for the weekly mine plan and comparing to actual performance within a waterfall chart.
Building the system
To facilitate the above deliverables, and encourage further analytics efforts by stakeholder teams after project-end, it was decided to create an automated data model that would act as a foundational semantic layer for all other project components. Data from the six source systems were all integrated from a SQL database into the tabular data model. Within the data model, time resolution standardisation and data lineages were established to link the material flow from in-pit grade block through the circuit to the mills – a complete digital twin from pit to process plant at daily granularity.
Hundreds of calculations were written within the data model to correctly track all the necessary parameters, including equipment runtimes, average power draw, mill feed fragmentation, and many more. These produced the master data table of all key parameters, from which the machine learning modelling could be performed. The full system development workflow is shown in figure 3.
Implementing machine learning
The machine learning modelling was an iterative process combining advanced data science capabilities with in-depth understanding of the mine’s operational flowsheet. The target for each model was the daily effective milling throughput rate. A large starter set of hundreds of parameters were successively reduced to a collection of 16 that demonstrated powerful insights into milling performance. These parameters were carefully selected to represent key points within the flowsheet: multiple geological parameters (material hardness, weathering fragmentation), primary crushing settings, secondary crushing performance, pebble crushing performance, rainfall levels, SAG mill operation, and ball mill operation. Two ‘generations’ of ML modelling were performed. The first generation was put into production quickly to start producing information and allow the client to fully understand how they wanted to interpret the prediction results. After several weeks, a second generation of ML models were created, which incorporated several important pieces of feedback through user testing of the first generation models.
The ML models put into production achieved consistent 90%+ accuracy and were able to accurately breakdown contributions to the milling rate by each stage of the flowsheet: from ore characteristics, to fragmentation, to crushing and milling control settings. To provide clear and regular distillation of results and feedback opportunities after the system deployment stage, a weekly mine-to-mill meeting was established for all project stakeholders where the main insights were presented and discussed. Rovjok built a waterfall analysis report (Figure 4), showing the impacts of each model parameter on the resulting mill throughput performance.
The waterfall chart compares actual data to the mining plan to see where where performance was good and/or where improvements can be made.
The ML models provided the client with new understanding of their weekly mining plan and its forecasted impacts on milling throughput, which they could then review using the waterfall analysis report, once the week had finished. With the analytical dashboard and regular review in place, the mine and process plant teams were in a position to clearly understand the impacts of their upcoming weekly mining plans, discuss, and make changes as necessary to seek to improve mill throughput.
Through the second half of the project (Jan22 – Jul22), improvements were made across the operation. Secondary crusher operators used the new fragmentation data visibility to crush greater amounts of coarse material. Mill operators reduced the variability of load, throw and toe to improve performance. These and other changes led directly to a 5-10% improvement in average monthly milling rate. Combined with this, a ~10% reduction in average monthly mill power draw was observed, providing a clear double benefit to the client. These improvements are presented graphically in Figures 5 and 6. Figure 5 presents the headline result of the mine-to-mill project, showing a distinct 5-10% improvement in monthly milling throughput through the duration of the project, which corresponds to over $200 million in additional value gained. Also achieved, as shown in Figure 6, is a clear reduction in mill power usage, with an approximate 10% reduction even at the increased milling rates. This corresponds to 1000 MWh of saved energy through 2022.
The client has continued utilising the capabilities generated by the mine-to-mill project. Further advanced analytics initiatives, building upon this foundation, are now taking place in specific departments across the operation.
Our collaboration with the Rovjok team provided us with invaluable insights. Their expertise in data analytics and the mining industry greatly enhanced our analytical capabilities. By working alongside our mining and process plant engineers in the data discovery process, the Rovjok team significantly improved our teams’ critical data literacy. The knowledge transfer process was meticulously planned and executed, resulting in a seamless transition. Thanks to this model, bolstered by machine learning, we can now forecast and analyse operational performance with a high degree of accuracy, leading to tangible improvements in both throughput and energy savings. The knowledge and skills imparted by the Rovjok team have been instrumental in our ongoing efforts to leverage data more effectively in our operations. The impact of their work extends beyond immediate operational improvement, contributing to our goal of making better data-driven decisions.
IT, BI & Analytics Manager; First Quantum Minerals Cobre Panama