AI-ready digital twin for mineral processing plants: Maximizing copper production with dynamic water management

Copper concentrators are required to increase metal production under stricter constraints on water, energy and operational stability. The main technical limitation is not the absence of data, but the limited conversion of historian tags, laboratory measurements, process events and operating records into validated process information. For a digital twin to support plant decisions, time-series data must be associated with plant topology, operating modes, event frames, performance indicators and dynamic process models. This article presents an AI-ready digital twin plant model for mineral processing plants with dynamic water management. The architecture combines process flow diagrams, PI System infrastructure, PI Asset Framework structures, event framing, time-derived analytics and hybrid process models. The objective is to classify plant operation, quantify hidden production and water-recovery losses, and provide a governed information layer for predictive analytics, model-based recommendations and language-model query interfaces. The technical connection between metal production and water recovery is established through the grinding product and the particle size distribution shape. Grinding and classification determine flotation response, the amount and type of particles reporting to tailings, tailings flocculation behavior, thickener stability and recovered water availability. The digital twin plant model therefore evaluates grinding, flotation, thickening, tailings and water recovery as a coupled plant system rather than as independent unit operations. Figure 1: Plant digital twin summary: The digital twin plant model integrates process engineering, operational data, predictive analytics, contextual events, visualization tools, and continuous improvement. (Adapted from the original OSB Digital Twin framework) This article also incorporates a feed-forward thickener characterization layer based on automated sedimentation measurement at the thickener feed or at a representative sampling point. Sedimentation rate and interface-profile measurements characterize the incoming slurry before delayed feedback indicators such as torque, bed level, overflow turbidity, underflow density and water recovery dominate the response. These measurements can be integrated with OPC/PI infrastructure and with thickener feedback variables to support earlier diagnosis and more consistent flocculant and operating recommendations. An industrial implementation in a Lundin Mining concentrator in Chile is used as the reference case. The reported results include increased water recovery, higher copper production and improved process stability without major capital investment. The case shows that the practical value of artificial intelligence (AI) in mineral processing depends on the prior construction of operational intelligence (OI): validated data, classified operating states, process models and plant-wide performance context. Figure 1 summarizes the digital twin plant model used in this work. The model is an operational information architecture that combines process engineering, process control, safety, sustainability, historian data, predictive analytics, contextual events and visualization tools. Its technical value depends on the continuous conversion of raw measurements into plant objects, operating time event intervals, model variables and performance indicators. The AI-readiness of architecture is treated as a data-governance and process-modeling requirement. Predictive models and language-model interfaces require plant data organized by topology, operating mode, event frame and performance indicator. Without this structure, the digital layer may display or summarize large volumes of information while providing limited process diagnosis. Water management is evaluated jointly with metal production. In low-grade copper concentrators, the grinding product affects mineral liberation, flotation recovery, tailings particle size, flocculation response, thickener overflow clarity, underflow solids concentration and recovered water availability (Trahar, 1981; Concha and Bascur, 2024). Particle size distribution shape therefore provides a technical link between mineral processing and dynamic water recovery. Figure 2 illustrates this plant-wide view. Operating-mode classification and time-derived analytics identify hidden losses across grinding, flotation, thickening, tailings and water recovery. The same structure allows metallurgical performance and water management to be evaluated against the production plan. This article also reflects the technical discussion developed during review meetings: A dynamic model that behaves correctly in simulation cannot be transferred directly to plant operation unless the data have been cleaned, classified and interpreted within the correct operating context. The implementation must use the existing historian infrastructure when available, preferably through PI System and PI Asset Framework, to avoid creating disconnected data silos and additional maintenance burdens. Figure 2: Maximizing copper production and water management: The figure connects operating modes, grinding product quality, flotation response, tailings thickening, water recovery and economic production. The industrial reference case is the digital twin plant implementation in Chile, which reported significant improvements in water recovery and copper production without major capital investment (Bascur et al., 2022). The case is used to show how structured historian data, operating-mode classification, dynamic models, and water-balance evaluation can improve plant performance. Digital twin integrated mine-plant-water treatment and tailings architecture A digital twin integrated plant model starts with the conversion of historian signals into operationally interpretable information. A single measurement can represent different conditions depending on ore type, plant load, operating mode, upstream state, downstream constraint and production target. A mill torque increase, cyclone-pressure change, flotation air-hold-up shift or thickener water-recovery loss requires process context before it can be diagnosed. Figure 3: Digital twin processing plant object model: The plant is represented as an integrated system connecting rock processing, mineral processing and water processing. The integrated plant object model shown in Fig. 3 organizes the concentrator into rock processing, mineral processing and water processing. Crushing, grinding, classification, flotation, thickening, tailings, recovered water and freshwater addition are represented as connected elements of the same plant system. This structure is required for plant-wide diagnosis because tailings thickener performance depends on upstream particle size distribution, flotation response, solids loading, reagent strategy and hydraulic balance. Figure 4 shows the corresponding data architecture. Sensors, weather data, PLC/SCADA systems, mining systems, utilities, laboratory databases, instruments, manual records, text files and cloud sources feed the process historian. PI Asset Framework, PI Analytics, event frames, and dynamic or empirical models then convert disconnected time-series signals into process information associated with plant assets and operating states. The critical step is contextualization. Each process unit is evaluated against its expected operating condition, production target, quality requirement, energy use, water use, environmental constraint, equipment condition and safety limit. Plant data are classified into operating modes such as running according to plan, running with hidden losses, operating under constraint, idle, downtime, maintenance and invalid-data condition (Bascur, 2020; Goldratt, 2014). Figure 4: OSB digital twin plant architecture: Data sources are transformed into operational information through the process historian, asset framework, analytics, event frames, dynamic models and visualization/AI tools. Event framing provides the time basis for this classification. Continuous historian data are divided into operational intervals, and each interval is characterized by time-derived variables such as average, minimum, maximum, rate of change, standard deviation, totalized production and deviation from plan. These calculations indicate when the plant met the production plan, when it lost production, when a unit was constrained, and when data quality was insufficient for model calibration. Figure 5 shows an example of this information in PowerBI. The dashboard summarizes the time spent by each process unit in each operating mode and relates those intervals to overall production effectiveness. The resulting language is common to operations, metallurgy, maintenance, process control and management. Figure 5: Overall production effectiveness: PowerBI advanced analytics. Validated and classified operating intervals are the proper basis for predictive analytics and hybrid modeling. Machine-learning models fitted to mixed data that include downtime, sensor failure, maintenance and abnormal states tend to reproduce artifacts rather than process behavior (Raschka, 2015; Kelleher et al., 2015; Brunton and Kutz, 2019). The data-preparation layer is therefore a modeling requirement. Figure 6: PI Vision display for real-time and historical plant KPIs: The display relates the process topology, operating state and production/water indicators. The language-model layer is positioned as a query and synthesis interface over structured operational information. It operates over plant objects, event frames, operating modes, model outputs, residuals, historical deviations and documented operating responses. This approach is consistent with the meeting discussion: The plant historian and asset model should carry the process context, while the AI layer queries that context instead of becoming a parallel data silo. Net metal production rate and dynamic water management Plant performance in a copper concentrator cannot be evaluated from throughput, recovery or water consumption as isolated indicators. A higher feed rate increases metal production only when grinding, classification, flotation, thickening and water recovery remain within their operating envelopes. Recovered water also affects the amount of water available for grinding and flotation. The digital twin plant model therefore evaluates production and water management as a plant-wide performance problem. Figure 6 shows the real-time visualization layer used to represent current and historical plant conditions. The display connects crushing, grinding, classification, rougher flotation, tailings thickeners, recovered water and freshwater consumption. Color-coded operating states indicate whether each section is running according to plan, operating under constraint, idle, down or requiring diagnosis. The objective is to locate the source of production, energy or water losses. Figure 7: Industrial visualization and optimization of net copper production rate in rougher flotation. Dynamic water management is evaluated through recycled water, freshwater flow, grinding water additions, flotation feed conditions, thickener operation and tailings discharge. Recovered water is an operational variable because it affects plant stability, achievable throughput, solids concentration and freshwater demand. A reduction in recovered water may force lower throughput, suboptimal density control or increased freshwater intake. A central indicator in this framework is the net metal production rate (NMPR). Instead of analyzing feed rate, copper grade and recovery separately, NMPR combines them into a single real-time production indicator: NMPRCu = (F · ttCu · RCu)/100 where F is the plant feed rate in tonnes per hour, ttCu is the copper grade, and RCu is the copper recovery according to the plant convention used for metallurgical reporting. When plotted against the production plan, NMPR quantifies the gap between actual and expected metal production. Figure 7 illustrates the use of NMPR in the rougher flotation circuit. The visualization compares actual net copper production rate with the predictive hybrid model and the expected operating condition. Deviations can then be interpreted together with grinding product quality, flotation hydrodynamics, water additions and thickener water recovery. NMPR creates a common performance indicator across operations, metallurgy, maintenance, process control and management. Operators evaluate alignment with the plan; metallurgists relate deviations to grade, recovery, particle size distribution and flotation response; process engineers evaluate grinding, classification, water availability and thickening; managers quantify the economic effect of operation outside the target window. Figure 8: Rougher flotation bank process profiles required for coarse particles, fines and slow-floating species. The dynamic water balance provides the complementary view. The digital twin plant model tracks freshwater addition, recovered water, grinding water, flotation water, thickener overflow, underflow solids concentration and tailings discharge. These variables are evaluated together because water recovery affects plant capacity and upstream particle size distribution affects thickener performance. The combined evaluation of NMPR and water balance answers two operational questions: how much metal is being produced relative to plan, and how much water is required or recovered to sustain that production. This basis supports the model hierarchy described in the next section. Hybrid process models for grinding, flotation and thickening The digital twin plant model uses dynamic process models calibrated with industrial data. These models preserve mass, water, energy and metallurgical structure, while the historian provides the data required for parameter estimation, operating-state evaluation, and comparison between measured and predicted performance. This formulation is required because plant response depends on ore type, particle size distribution, water addition, solids concentration, residence time, equipment condition and operating mode (Bascur, 2020; Concha and Bascur, 2024). Grinding and classification are the main upstream sources of variability transmitted to flotation and thickening. Throughput and P80 are insufficient descriptors when the complete particle size distribution shape affects mineral liberation, flotation kinetics, water entrainment, tailings flocculation, thickener overflow clarity and underflow solids concentration. The digital twin evaluates the grinding product through size-distribution parameters, cyclone behavior, water addition, circulating load and solids concentration (Trahar, 1981; Concha and Bascur, 2024). The flotation model links the grinding product to rougher bank performance through particle size distribution, pulp hydrodynamics, residence time, air hold-up, reagent addition and metallurgical recovery. Deviations in copper recovery are interpreted as coupled effects of upstream conditions and flotation operating state, rather than as isolated changes in a single control variable (Lynch et al., 1981; Mika and Fuerstenau, 1968). Figure 8 shows rougher flotation bank profiles for coarse particles, fine particles and slow-floating species. These profiles provide the operational structure for evaluating the distribution of air, pulp level, frother addition, power intensity and residence time across the bank. The same framework supports bank-by-bank recovery estimation and identification of hydrodynamic conditions that are inconsistent with the incoming particle size distribution. The thickening model represents the downstream water-recovery response of the plant. Tailings thickener performance depends on solids flow from flotation, tailings particle size distribution, flocculant dosage, feed solids concentration, bed inventory, underflow withdrawal, overflow clarity and residence time. These variables determine recovered water, underflow stability and the operating margin available before torque, bed level or overflow turbidity constraints are reached. Figure 9 shows the tailings thickener modeling and tracking layer. The model compares measured water recovery with the expected recovery calculated from the thickener operating state. Relevant inputs include feed tonnage, feed solids concentration, underflow solids concentration, flocculant rate, inventory level, underflow flowrate and particle size distribution shape. The residual between measured and calculated water recovery provides an operational indicator of thickener performance. Residence time is central to thickener interpretation. Rake torque, bed level, underflow density, overflow turbidity and recovered water flowrate are delayed responses of material already inside the thickener. A torque increase may be associated with underflow withdrawal, solids inventory, particle size distribution, flocculation response, or disturbances generated in grinding and flotation. The interpretation of feedback variables therefore requires upstream context. Figure 9: Tailings thickener water-tonnage modeling and tracking. The hybrid modeling layer uses plant data for parameter estimation, model validation, residual analysis and state classification. Data-driven calculations are applied to operating intervals classified as valid for model evaluation. Downtime, maintenance, sensor failure, frozen signals, unstable operation and invalid tags are excluded from model calibration (Raschka, 2015; Kelleher et al., 2015; Brunton and Kutz, 2019). The model hierarchy is organized as follows. The grinding and classification model estimates the particle size distribution delivered to flotation and tailings. The flotation model estimates metallurgical recovery and tailings characteristics as a function of particle size, hydrodynamics, reagent addition and residence time. The thickening model estimates water recovery, solids inventory, underflow behavior and overflow clarity. These models are connected through a common plant data structure, operating-mode classification, event frames and performance indicators. This hierarchy supports coordinated diagnosis. A deviation in NMPR can be traced to grinding quality, flotation hydrodynamics, water availability or thickener water recovery. A reduction in recovered water can be evaluated against upstream particle size distribution, flocculation response, solids loading, inventory control and underflow discharge capacity. Feed-forward thickener characterization and model-based decision support Tailings thickeners are commonly monitored through rake torque, bed level, underflow density, underflow flowrate, overflow turbidity, flocculant dosage and recovered water flowrate. These variables are necessary for supervision and diagnosis, although they describe the delayed response of material already inside the thickener. Their interpretation must account for residence time, solids inventory, compression behavior, underflow withdrawal capacity, and the upstream operating condition of the concentrator. The review meeting on thickener models emphasized this limitation. Existing model predictive control (MPC), advanced process control (APC) and expert-system approaches often use feedback indicators such as torque, bed level, underflow density or overflow turbidity. Those signals are affected by the internal residence time of the thickener. A disturbance may already have propagated through the unit before the feedback variable reaches an alarm or control threshold. The resulting operating logic becomes reactive when it lacks a direct measurement of the incoming slurry condition. Figure 10: Dynasoftpro running in Google Cloud for training and process-control analysis. A thickener operating deviation cannot be assigned to a single measurement without characterizing the material entering the unit. A torque increase, loss of overflow clarity, decrease in underflow density or reduction in recovered water may originate from grinding product size distribution, cyclone classification, flotation behavior, reagent carry-over, feed solids concentration, clay content or flocculation response. The digital twin plant model therefore includes a feed-forward characterization layer for the thickener feed. The proposed layer uses automated sedimentation measurement at the thickener feed or at a representative sampling point. The measurement system extracts a slurry sample, performs a controlled settling test, tracks the interface position as a function of time and generates a sedimentation curve. From this curve, the system estimates initial settling rate, interface stability, compression-zone behavior, abnormal settling patterns, and changes in flocculation response. These variables characterize the incoming material before the delayed thickener response dominates the measurement. When transmitted through OPC/PI infrastructure, the sedimentation signal can be combined with historian tags, operating modes, particle size distribution, feed solids concentration, flocculant dosage, torque, bed level, underflow density, overflow turbidity, underflow flowrate, inventory estimates and recovered water. The result is a combined feed-forward and feedback representation of thickener operation. The feed-forward component describes the expected settling behavior of material entering the thickener. The feedback component describes the actual response of the thickener inventory, underflow discharge and overflow quality. Their comparison supports mechanism-level diagnosis. For example, a reduction in water recovery preceded by a decrease in feed settling rate suggests a change in incoming slurry characteristics. A torque deviation without a corresponding change in settling behavior may indicate inventory accumulation, underflow withdrawal limitations or local mechanical constraints. This structure is directly relevant for flocculant recommendation. Conventional logic often adjusts flocculant from delayed indicators such as torque, bed level, overflow turbidity or underflow density. Incorporating settling rate and interface-profile information allows the recommendation calculation to include the current condition of the feed stream. Feed-forward variables include sedimentation rate, sedimentation profile, feed solids concentration, particle size distribution, upstream operating state, and available ore-type or mineralogical indicators. Feedback variables include torque, bed level, underflow density, overflow turbidity, recovered water, underflow flowrate and inventory estimates. The recommendation must be evaluated within the operating mode and production context of the plant. A flocculant adjustment has different implications when the plant is running according to plan, operating under a water-recovery constraint, experiencing abnormal fine-particle generation or operating with underflow discharge limitations. The digital twin therefore evaluates flocculant dosage, underflow withdrawal, dilution water and operating limits as coordinated actions rather than isolated set-point changes. Dynamic simulation and training tools provide an additional layer for model evaluation. Figure 10 shows Dynasoftpro running in a cloud environment for training and process-control analysis. Historical operating intervals, model predictions and alternative operating strategies can be evaluated outside the control system. This supports engineering analysis, operator training, model calibration and scenario comparison. The same structured information layer can be exposed through advanced visualization tools or language-model interfaces. In this architecture, the query layer interrogates validated plant information: operating modes, event frames, model residuals, historical deviations, similar events and documented operating responses. The reliability of this interaction depends on the data contextualization, model structure and event classification described in the previous sections. Industrial implementation and results The digital twin plant model was implemented in an industrial copper concentrator in Chile to support plant-wide performance evaluation, dynamic water management and operating-mode analysis. The implementation used existing process data infrastructure, including historian data, process tags, calculated variables, laboratory information and operating records. These data were structured according to the plant object model described in the previous sections, allowing grinding, flotation, thickening, recovered water and freshwater consumption to be evaluated as one operational system. The implementation followed a staged methodology. First, relevant plant units and process streams were mapped into an asset-based structure. Second, available tags were reviewed, validated, and associated with their process objects. Third, operating modes were defined for each relevant unit, distinguishing productive operation, constrained operation, idle periods, downtime, maintenance events and invalid-data intervals. Fourth, event frames transform continuous historian data into operational time windows. Finally, dynamic models and performance calculations were applied to classified operating mode time intervals. The main performance objective was to increase copper production while improving water use and recovery. These objectives were evaluated through NMPR and the dynamic water balance. NMPR measured metallurgical production against the operating plan, while the water balance quantified freshwater use, recovered water, thickener performance, and available process water for grinding and flotation. A central outcome was the identification of hidden losses: intervals in which the plant was operating but failing to meet expected production, water-recovery or operating-condition targets. The event-based structure quantified the duration and impact of these intervals and related them to specific process areas. The analysis therefore moved from trend visualization to operational diagnosis. The integrated model showed that deviations in water recovery were not explained only by thickener operation. In several intervals, thickener performance was affected by grinding quality, particle size distribution, flotation response, solids loading and water distribution. This result confirmed the need to evaluate the thickener as part of the complete concentrator. The reported industrial results include an increase of approximately 40 percent in water recovery and 32 percent in copper production without major capital investment (Bascur et al., 2022). These improvements were associated with operating visibility, classification of plant states, identification of hidden losses, coordinated interpretation of grinding-flotation-thickening interactions, and systematic comparison of actual performance against the production plan. The case shows that the main contribution of the digital twin plant model is the organization of plant information into a structure for diagnosis, performance evaluation and coordinated operational decision-making. Validated historian data, process models, event frames, operating-mode classification and plant-wide key performance indicators (KPIs) provide a consistent basis for identifying the origin of production and water losses. The implementation also clarifies the technical requirement for AI-ready infrastructure. Visualization dashboards, predictive models and language-model interfaces require information structured by process topology, operating state and business objective. The same tools add limited engineering value when they are connected only to unclassified tag collections. Implementation roadmap and conclusions Implementation must follow a staged sequence. The first requirement is a reliable operational information layer: plant topology, historian data, asset structures, validated tags, event frames, operating modes and performance indicators. Predictive models and language-model interfaces should be introduced after this layer is governed. Table 1 summarizes the proposed maturity roadmap. The sequence starts with visualization and contextualization, advances toward predictive analytics and model-based recommendations, and then supports closed-loop optimization under defined operating envelopes. The roadmap is an operational transformation sequence. A plant cannot progress reliably toward prediction, optimization or autonomy if its data structure, instrumentation, operating modes and process models are not governed. The digital twin plant model provides the technical framework for converting fragmented industrial data into a plant-level representation of production, water recovery, equipment condition and operating constraints.The work leads to six conclusions: Dynamic water management and copper production must be evaluated as coupled objectives. In low-grade ore concentrators, recovered water affects plant capacity, while grinding and flotation conditions affect thickener performance. Particle size distribution shape is a key variable connecting grinding, flotation and thickening. It affects mineral liberation, flotation response, tailings flocculation, overflow clarity, underflow solids concentration and recovered water.Historian data require validation, classification and contextualization before they are used for hybrid modeling or decision support. Operating-mode classification and event framing distinguish productive operation, constrained operation, downtime, maintenance, sensor failure and invalid-data periods. Hybrid process models connect process knowledge with plant data. Grinding, flotation and thickening models interpret measured performance, estimate difficult-to-measure states, calculate deviations from operating envelopes and relate local disturbances to plant-wide consequences. Feed-forward characterization of thickener feed improves diagnosis of delayed thickener responses. Torque, bed level, underflow density, overflow turbidity and recovered water remain essential feedback variables, while sedimentation-rate and interface-profile measurements provide upstream information for flocculant recommendation and operating-limit adjustment. The language-model interface should operate over validated operational information. Its practical value comes from access to plant objects, event frames, model outputs, historical deviations, performance indicators and documented operating responses. The industrial case indicates that a structured digital twin plant model can improve performance without relying exclusively on new capital equipment. The reported increase in water recovery and copper production was achieved through better use of existing data, operating-mode analysis, model-based interpretation, and coordinated evaluation of grinding, flotation, thickening and water management. The main contribution of this work is the integration of OI, hybrid process modeling, dynamic water management and AI-ready data structures into a plant-level framework for evidence-driven mineral processing operation. (References available from the author).

AI-ready digital twin for mineral processing plants: Maximizing copper production with dynamic water management

Authored by: Osvaldo A. Bascur

To view the past issues of the UCA magazine you must be a member of SME/UCA.

Mining Engineering’s Weekly Rundown: Episode 73

Get your SME Foundation VIP Access Conference Pass to MINEXCHANGE 2027

“Check the box” and help support the SME Foundation

Mining Engineering’s Weekly Rundown: Episode 73

Get your SME Foundation VIP Access Conference Pass to MINEXCHANGE 2027

“Check the box” and help support the SME Foundation

To view the past issues of the UCA magazine you must be a member of SME/UCA.

Roundtable discussion on mine finance

How GeologicAI uses high-resolution data to reduce uncertainty in mining

Related Articles

Get your SME Foundation VIP Access Conference Pass to MINEXCHANGE 2027

“Check the box” and help support the SME Foundation

The Art and Science of the Flowsheet

UCA Division officer nominations

Mining Engineering’s Weekly Rundown: Episode 73

Get your SME Foundation VIP Access Conference Pass to MINEXCHANGE 2027

“Check the box” and help support the SME Foundation

Mining Engineering’s Weekly Rundown: Episode 73

Get your SME Foundation VIP Access Conference Pass to MINEXCHANGE 2027

“Check the box” and help support the SME Foundation

Use Single Sign On