Blending Characterization for Effective Management in Mining Operations

Abstract

Ore blending plays a critical role in ensuring feed consistency and optimizing downstream processes in the mining industry. Despite its importance, effective blending remains challenging due to ore variability and operational constraints. This review focuses exclusively on modern, data-driven blending methodologies, with particular emphasis on the application of data science and machine learning (ML) in predicting key process variables and supporting real-time decision-making. It discusses core challenges such as data quality, feature engineering, and model generalization, alongside enabling technologies including sensor integration, automation platforms, and real-time data acquisition systems. By consolidating the recent literature and highlighting emerging trends, this work outlines future directions for advancing intelligent blending systems and underscores the importance of standardized, high-quality data in the development of robust digital solutions for mineral processing.

1. Introduction

Blending quality is a critical factor in mining operations, directly influencing the efficiency and effectiveness of downstream processes. It encompasses various characteristics, including ore grade, particle size distribution (PSD), moisture content, mineralogical composition, and mechanical properties such as hardness and abrasiveness. Variability in these attributes can disrupt process stability, leading to inefficiencies, increased energy consumption, equipment wear, and suboptimal product quality. For instance, ore grade determines the concentration of valuable minerals and impacts recovery, while PSD affects comminution efficiency and throughput [1]. Similarly, moisture content influences material handling and processing, potentially causing blockages in feeders or conveyors or affecting residency times [2]. Mineralogical composition dictates the selection of processing methods and reagents, while variations in ore competence or hardness can increase grinding energy demands and wear rates [3]. Managing these variations effectively is crucial for stabilizing operations, improving productivity, and ensuring sustainable resource extraction.

Beyond operational improvements, blending also contributes to sustainability efforts in mining. More consistent feed reduces energy consumption, equipment wear, and environmental impacts. The ability to predict and control ore variability also can improve resource planning, reducing risks associated with extraction and enhancing long-term viability. As mining operations strive to meet global sustainability goals, effective blending strategies are becoming essential for both operational and environmental objectives.

Recent technological advancements have transformed blending from a reactive process into a proactive, data-driven approach. Real-time monitoring and sensor-based systems provide continuous feedback on ore characteristics, allowing dynamic adjustments to blending ratios. Machine learning (ML) algorithms and data-driven methodologies further enhance blending operations by identifying patterns and predicting future behavior. By integrating these technologies, mining operations can optimize blending decisions, reduce variability, and improve overall process performance [4].

This work reviews and analyzes the most widely adopted data-driven blending techniques in the mining sector, with a particular focus on machine learning and artificial intelligence (AI)-based methodologies. The paper evaluates the challenges, effectiveness, and limitations of these approaches while highlighting opportunities to enhance blending strategies through emerging technologies. By concentrating on technologically enabled blending solutions and their influence on operational efficiency, this study provides a structured overview of current practices and identifies key areas for future research.

This article addresses a critical gap by synthesizing blending strategies through predictive modeling, variable control, and digital integration. Unlike broader reviews that discuss blending as part of general mine planning or geo-metallurgy, this work identifies how operational and data-driven blending strategies are currently deployed and evaluated, highlighting key trends, research gaps, and future opportunities in the field.

2. Background

Mining involves multiple stages to transform raw ore into a product suitable for downstream processing. After extraction, one of the most critical intermediate steps is ore blending, which aims to homogenize feed characteristics before the material enters the plant. Effective blending reduces variability in particle size distribution (PSD), grade, and hardness, factors that directly influence plant stability, energy efficiency, and recovery [5].

Blending practices can be implemented at various points across the value chain. However, this study focuses exclusively on mine-site and pre-processing blending, particularly upstream of SAG milling. While blending is also critical in product preparation or port-based operations (e.g., in coal or manganese industries), such downstream strategies are outside the scope of this manuscript. The focus here is on upstream blending decisions where operational teams can influence material properties in real-time or short-term planning windows.

Typical upstream blending strategies include the following:

• Wheeled Blending: This method is performed on the stockpile pad using heavy machinery such as loaders or bulldozers to physically mix ores from different trucks or shovels, reducing feed heterogeneity. This mechanical mixing creates a more uniform blend by homogenizing material layers [6,7]. The study by [7] further highlights the use of near real-time 3D ROM stockpile modeling to track and improve blending efficiency, emphasizing how loader operations directly influence grade distribution and blending outcomes.
• In-Pit Blending: Blending occurs at the mining face by selectively loading ores from distinct geological zones with varying characteristics into haul trucks. These materials are combined en route to the crusher or stockpile, enabling early-stage homogenization of feed properties. Often integrated with In-Pit Crushing and Conveying (IPCC) systems, this approach has demonstrated significant improvements in net present value and reductions in grade variability [6].
• Rotary Blending: Conducted near the processing plant using rotary drums or multi-axis mechanical mixers, this method homogenizes ore blends prior to grinding. Rotary mixers promote uniformity through tumbling and mixing actions. Advanced designs, such as multi-dimensional rotary mixers, can substantially enhance mixing efficiency by improving axial and radial blending [8].
• Conveyor Belt Blending: This method blends multiple ore sources fed at controlled rates onto conveyor belts, allowing inline mixing as materials move through conveyor chutes and transfer points. Discrete Element Method (DEM) simulations have shown how conveyor geometry and transfer chute design influence blending effectiveness, leading to a more homogeneous feed for downstream processing [9,10].

Due to confidentiality agreements with industrial partners, specific mine names or process configurations cannot be disclosed. Nonetheless, this study builds upon practical collaborations with three large-scale metallic mining operations, each employing multi-stage blending practices that combine in-pit, stockpile, and conveyor blending. These integrated approaches are essential for managing variability in feed properties such as PSD, hardness, and mineralogical composition, parameters that critically affect SAG mill performance and overall process stability. Figure 1 illustrates a generalized material flow within a mine-site and processing plant, highlighting key operational stages. The flow begins at the Mine, where ore is extracted and selectively loaded based on geological variability. Material then passes through the Primary Crusher for initial size reduction. The Stockpile stage serves as the critical blending step, where ores from different sources are homogenized to reduce variability in feed characteristics such as particle size distribution and grade. This blended material is subsequently fed into the Grinding Circuit, which includes high-pressure grinding rolls (HPGRs) and ball mills, representing metallurgical optimization steps focused on further size reduction, mineral liberation, and maximizing recovery downstream.

3. Methods

This section outlines the methodology followed for conducting the systematic review in alignment with the PRISMA (Preferred Reporting Items for Systematic Reviews and Meta-Analyses) guidelines [11]. The objective is to identify, select, and analyze relevant studies that apply blending strategies, ore characterization, and data-driven techniques in mining operations. The review incorporates both peer-reviewed journal articles and selected conference proceedings to enhance the practical context.

We conducted a structured literature search across major scientific databases, including ScienceDirect, Springer, and IEEE Xplore. The search targeted publications related to blending in mining, predictive modeling, and ore characterization. To guide this search, we developed a set of research questions (defined in Section 3.1) that focus on key aspects of ore particle size distribution, stockpile management, instrumentation, and predictive methods in mining operations.

3.1. Research Questions

These research questions provide a structured framework to analyze and extract meaningful insights from the existing body of literature. They address critical factors influencing particle size distribution (PSD) and their implications for mining efficiency and process characterization:

• How can we determine if the blending was successful, and what indicators help us evaluate it?
• What variables and models are used to determine blending criteria, and how do they influence the process?
• What methods are available to assess the impact of blending on downstream processes like milling and flotation?
• What are the typical challenges in blending operations?
• What machine learning models and predictive technologies can be applied to improve blending strategies?
• What types of sensors are used to monitor blending performance?
• What procedures are used to verify blending quality?

Guided by these questions, the literature search employed targeted search strings such as

• “blending" AND “data" AND “mining" (AND “analytics" OR “modeling" OR “machine learning" OR “artificial intelligence" OR “AI" OR “data analysis" OR “statistical analysis").
• “blending" AND “particle size" AND “distribution" AND “mining".
• “blending" AND “mineral" AND “characterization" AND “mining".

A total of 143 records were initially identified through database searching and additional sources. After removing 3 duplicate records, 140 records remained for screening based on titles and abstracts. During this screening, 78 records were excluded due to reasons such as lack of relevance or unavailability of full texts. Subsequently, 62 full-text articles were assessed for eligibility, of which 15 were excluded with specific reasons. Consequently, 47 studies were selected for inclusion in the final qualitative synthesis. Figure 2 illustrates the PRISMA flowchart summarizing this selection process.

3.2. Inclusion and Exclusion Criteria

To ensure alignment with the objectives of this study, the following inclusion criteria were applied:

• Studies focused on blending operations or ore characterization in metallic open-pit mining.
• Research involving particle size distribution (PSD), ore grade, moisture, and mineralogical variability as blending variables.
• Applications of machine learning (ML), statistical modeling, or sensor-based approaches in blending.
• Articles published between 2014 and 2025.

The exclusion criteria were as follows:

• Studies unrelated to blending or ore characterization in mining.
• Research exclusively focused on underground mining, non-metallic ores, or laboratory-scale experiments without operational relevance.
• Articles lacking methodological transparency or missing performance metrics.

3.3. Data Extraction and Coding

From the selected studies, we extracted the following categories of information:

• Variables used to define blending criteria (e.g., PSD, grade, and moisture).
• Types of sensors and data acquisition systems used, including the following:

  ‑

  Prompt Gamma Neutron Activation Analysis (PGNAA).

  ‑

  Laser-Induced Breakdown Spectroscopy (LIBS).

  ‑

  Smart Data Fusion (SDF).

• Modeling techniques such as regression, simulation, and classification.
• Performance metrics including root mean square error (RMSE), mean absolute error (MAE), and measures of grade variability.
• Reported outcomes and industrial relevance.

Each study was analyzed and categorized according to the research questions defined earlier. We note that while some studies consider mechanical properties such as abrasiveness (i.e., the ore’s tendency to wear down equipment), this characteristic primarily informs equipment design and maintenance decisions, rather than blending optimization. Therefore, abrasiveness is excluded as a blending variable in this review.

4. Results and Discussion

This section presents the results of our comprehensive review, addressing the research questions outlined in Section 3.1, and is followed by a detailed discussion. The reviewed studies are thematically grouped and summarized in the tables: Table 1, Table 2 and Table 3 organize machine learning models applied to ore blending and particle size distribution (PSD) prediction, categorized by methodology. Studies within each table are enumerated for clarity, and ordering reflects thematic relevance rather than chronological publication date.

It is important to note that performance metrics such as RMSE, MAE, R², and grade variability reported in the reviewed studies are not standardized across the literature. Variations in datasets, preprocessing techniques, and evaluation methodologies mean that only some studies report these metrics, and those that do often use different definitions or calculation methods. Consequently, direct comparisons of model performance between studies should be made with caution. This highlights the need for more consistent and transparent reporting standards in future research to enable meaningful benchmarking and evaluation of machine learning models applied to ore blending and PSD prediction.

Table 1. Summary of representative machine learning models for blending and PSD prediction.

Model/Method (with Citation)	Input Variables	Predicted Output	Performance Metrics
Artificial Neural Network (ANN, 3-10-1) [12]	Rod load (%), Feed load (%), Grinding time (min)	PSD (percent passing per sieve)	R = 0.999 (train), R = 0.987 (val), RMSE = 0.165–0.965
ANN + Genetic Algorithm [13]	Feed size, Ore hardness, Liner wear	Energy consumption	RMSE reduced vs. baseline; improved accuracy
Artificial Bee Colony (ABC) Optimization [14]	Ore grades, costs, recovery constraints	Optimized ore blending plan	Recovery: W +8.95%, Mo +12.2%, Bi +8.33%; Grade fluctuation <10%
Goal Programming Optimization [15]	Rock powder sampling, Grade priority constraints	Optimized ore blending schedule	Grade deviation minimized; rapid linear optimization
Stochastic Gradient Boosting (SGB) [16]	MIR/NIR spectra, soil C, pH, RGB, extractants	PSD (sand, silt, clay %)	R2 = 0.87–0.95 (MIR); R2 = 0.66–0.79 (NIR)
Deep LSTM (DRNN) [17]	Ore feed, Water flow (sumps), Control loops	6 KPIs (Throughput, Solids %, Size fractions, Recirculation)	Accuracy = 99%; outperforming wavelet models
CatBoost + Differential Evolution [18]	SAG mill weight, power draw, feed size, water rate	SAG mill throughput (t/h)	Highest prediction accuracy (CatBoost); best robustness (DE)
ANN, Random Forest, XGBoost, GBM [19]	SAG feed P80, granulometry (<30 mm, >100 mm), rotational speed, liner age, solids %, stockpile level	SAG mill production (TpH), energy consumption	ANN: R2 = 0.885, RF: R2 = 0.755, XGBoost: R2 = 0.772; Production increase 4.42%, Energy decrease 7.62%

Table 1. Summary of representative machine learning models for blending and PSD prediction.

Model/Method (with Citation)	Input Variables	Predicted Output	Performance Metrics
Artificial Neural Network (ANN, 3-10-1) [12]	Rod load (%), Feed load (%), Grinding time (min)	PSD (percent passing per sieve)	R = 0.999 (train), R = 0.987 (val), RMSE = 0.165–0.965
ANN + Genetic Algorithm [13]	Feed size, Ore hardness, Liner wear	Energy consumption	RMSE reduced vs. baseline; improved accuracy
Artificial Bee Colony (ABC) Optimization [14]	Ore grades, costs, recovery constraints	Optimized ore blending plan	Recovery: W +8.95%, Mo +12.2%, Bi +8.33%; Grade fluctuation <10%
Goal Programming Optimization [15]	Rock powder sampling, Grade priority constraints	Optimized ore blending schedule	Grade deviation minimized; rapid linear optimization
Stochastic Gradient Boosting (SGB) [16]	MIR/NIR spectra, soil C, pH, RGB, extractants	PSD (sand, silt, clay %)	R2 = 0.87–0.95 (MIR); R2 = 0.66–0.79 (NIR)
Deep LSTM (DRNN) [17]	Ore feed, Water flow (sumps), Control loops	6 KPIs (Throughput, Solids %, Size fractions, Recirculation)	Accuracy = 99%; outperforming wavelet models
CatBoost + Differential Evolution [18]	SAG mill weight, power draw, feed size, water rate	SAG mill throughput (t/h)	Highest prediction accuracy (CatBoost); best robustness (DE)
ANN, Random Forest, XGBoost, GBM [19]	SAG feed P80, granulometry (<30 mm, >100 mm), rotational speed, liner age, solids %, stockpile level	SAG mill production (TpH), energy consumption	ANN: R2 = 0.885, RF: R2 = 0.755, XGBoost: R2 = 0.772; Production increase 4.42%, Energy decrease 7.62%

Table 2. Machine learning applied to blending. Part 1.

N	Author	Model Type	Objectives	Methodology	Instrumentation	Data Used	Key Variables	Results
1	Zongnan Li [16]	Mathematical Model	Establish PSD model for mine tailings and validate with samples.	Mathematical PSD model with A, B, K coefficients; validated with twelve tailing materials.	PSD determined using Laser PSA	Twelve tailing samples from metal mines in China.	Model coefficients (A, B, K) for PSD.	R2 > 0.99 for all twelve PSD lognormal curves.
2	Srinivas Soumitri Miriyala [17]	Deep Learning	Develop a multi-objective optimization algorithm for LSTM networks to improve generalization	Focuses on nonlinear system identification in grinding circuits, comparing optimal LSTMs with traditional tools	Deep recurrent neural networks for system identification, highlighting data model accuracy needs	Lack of sensors in circuits, analyzing overflow slurry properties for KPIs	Data from PRBS signals; LSTM networks achieved 99% accuracy, tested on unseen signals	Emphasizes throughput as a productivity indicator and evaluates models using AIC and RMSE
3	Brian Lindner [20]	Simulation/ Diagnosis	Develop a fault diagnosis approach using process topology and feature extraction for fault detection	Applies methodology to industrial concentrator data to create a connectivity map	Combines process topology and feature extraction for effective diagnosis	Uses historical data to extract topology, generating a connectivity map for analysis	Key variables include Mill2 and Mill3; identifies root nodes influencing cyclone feed	AUCs improved from 0.85 to nearly 1; methodology emphasizes accurate fault detection and topology integration
4	Kumah, F.N. [21]	Regression	Create ML models to predict mine production and enhance efficiency.	ANN, RF, GBR, and DT models; MLR as baseline; regression analysis for feature selection.	Not addressed	126 historical datasets from Pit W of Mine X.	Daily trucks, excavator hours, and utilization.	ANN model best (R2 = 0.8003); MLR worst (R2 = 0.6044).
5	Bingyu Liu [22]	Optimization	Establish a quality model for ore blending to enhance product quality.	Intelligent ore blending method using NSGA-II and ABC algorithms.	Not addressed	Historical blending and dressing data.	Ore grade, high-quality content.	New method improved profits.
6	F Nakhaei [23]	Imaging/ Automation	State-of-the-art 3D imaging in mineral processing, focusing on automation and efficiency	3D imaging techniques (RhoVol, X-ray CT), high-speed scanning, integration for automation	RhoVol measures particle mass, XCT captures internal structures, improving mineral characterization	3D imaging for particle size, shape, and composition analysis, comparison with sieving methods	Key variables: size, shape, porosity, crack density, and intergranular fractures, essential for characterization	Not addressed
7	Shi Zhao [24]	Simulation	Develop a real-time 3D volumetric model and improve quality grade accuracy.	Triangular prism model, sub-prism partitioning, adaptive reclaiming, and blending.	LMS200 and O1D100 laser scanners	2D laser range finder data for stockpile profiling.	Stockpile shape, quality grades, angle of repose, density.	R2 for regression accuracy, SSE for model evaluation.
8	Ji Wu [14]	Optimization	Improve ore blending to improve recovery and cost efficiency.	ABC algorithm for multi-objective optimization; software developed for blending.	Algorithm implemented in MATLAB	Historical data from Shizhuyuan Mine, focusing on ore recovery.	Recovery, mining costs, production capacity.	Recovery rate increased from 0.45% to 0.491%.
9	Dahee Jung [25]	Review	Review ML applications in mining and analyze trends.	Systematic literature review of 109 papers; keyword search; selection criteria defined.	Not addressed	Various datasets, including open-source data from RapidEye.	Data sources, utilization types.	ML performance evaluated; deep learning average RMSE: 0.000608.
10	Shi Qiang Liu [26]	Review	Review ML applications in open-pit mining to improve predictability and feasibility	Statistical analysis of 100+ papers (2012–2023), categorized into four key areas	Sensors for real-time drilling, hyperspectral data analysis	Operational resources, equipment utilization, mine geology, production, environment	Block extraction, tonnage (tons), volume (cubic meters), destination assignment	Not addressed
11	Larissa Statsenko [27]	Simulation	Develop an info system for ore blending to stabilize output quality.	Imitation modeling and simulation for ore blending design; decision support system applied.	Not addressed	Datasets from four open-pits in Kazakhstan.	Stockpile dimensions, transport mass, Fe-grade.	Ore grades variance reduced  4 times; blending coefficient improved 1.5–2 times.
12	Parent, E. [28]	Regression	Evaluates infrared (IR) models for soil PSD using machine learning (SGB) to predict sand, silt, clay, and carbon content. The 4X gain method achieves R2 values between 0.87 and 0.95; NIR and MIR spectra give R2 between 0.40 and 0.94.	SGB applied to NIR and MIR spectra; R2 values: 0.40–0.94 for NIR, 0.45–0.80 for MIR, 4X gain method: R2 = 0.87–0.95.	Stochastic Gradient Boosting (SGB)	1298 soil samples; R2 for sand sieving: 0.66, sedimentation: 0.53, NIR: 0.69–0.92, MIR: 0.45–0.80, SDF and settling time: 0.84–0.92.	Sand, silt, clay, carbon content, SDF, settling time, NIR, MIR spectra.	4X gain: R2 = 0.87–0.95; NIR and MIR: R2 = 0.40–0.94; Settling time and SDF: R2 = 0.84–0.92. Machine learning enhances PSD prediction.
13	Otsuki, A. et al. [12]	Regression	Investigate ANNs for predicting PSD to reduce energy consumption.	Supervised ANN trained on experimental data.	Rod mill for grinding, PSD measurement of coal.	Experimental data with varying feed vol.%, rod load vol.%, and grinding time.	Feed volume%, rod load, grinding time.	ANN predicted PSD well; RMSE ranged from 0.165 to 0.965, showing its ability to predict PSD under varying conditions.

Table 2. Machine learning applied to blending. Part 1.

N	Author	Model Type	Objectives	Methodology	Instrumentation	Data Used	Key Variables	Results
1	Zongnan Li [16]	Mathematical Model	Establish PSD model for mine tailings and validate with samples.	Mathematical PSD model with A, B, K coefficients; validated with twelve tailing materials.	PSD determined using Laser PSA	Twelve tailing samples from metal mines in China.	Model coefficients (A, B, K) for PSD.	R2 > 0.99 for all twelve PSD lognormal curves.
2	Srinivas Soumitri Miriyala [17]	Deep Learning	Develop a multi-objective optimization algorithm for LSTM networks to improve generalization	Focuses on nonlinear system identification in grinding circuits, comparing optimal LSTMs with traditional tools	Deep recurrent neural networks for system identification, highlighting data model accuracy needs	Lack of sensors in circuits, analyzing overflow slurry properties for KPIs	Data from PRBS signals; LSTM networks achieved 99% accuracy, tested on unseen signals	Emphasizes throughput as a productivity indicator and evaluates models using AIC and RMSE
3	Brian Lindner [20]	Simulation/ Diagnosis	Develop a fault diagnosis approach using process topology and feature extraction for fault detection	Applies methodology to industrial concentrator data to create a connectivity map	Combines process topology and feature extraction for effective diagnosis	Uses historical data to extract topology, generating a connectivity map for analysis	Key variables include Mill2 and Mill3; identifies root nodes influencing cyclone feed	AUCs improved from 0.85 to nearly 1; methodology emphasizes accurate fault detection and topology integration
4	Kumah, F.N. [21]	Regression	Create ML models to predict mine production and enhance efficiency.	ANN, RF, GBR, and DT models; MLR as baseline; regression analysis for feature selection.	Not addressed	126 historical datasets from Pit W of Mine X.	Daily trucks, excavator hours, and utilization.	ANN model best (R2 = 0.8003); MLR worst (R2 = 0.6044).
5	Bingyu Liu [22]	Optimization	Establish a quality model for ore blending to enhance product quality.	Intelligent ore blending method using NSGA-II and ABC algorithms.	Not addressed	Historical blending and dressing data.	Ore grade, high-quality content.	New method improved profits.
6	F Nakhaei [23]	Imaging/ Automation	State-of-the-art 3D imaging in mineral processing, focusing on automation and efficiency	3D imaging techniques (RhoVol, X-ray CT), high-speed scanning, integration for automation	RhoVol measures particle mass, XCT captures internal structures, improving mineral characterization	3D imaging for particle size, shape, and composition analysis, comparison with sieving methods	Key variables: size, shape, porosity, crack density, and intergranular fractures, essential for characterization	Not addressed
7	Shi Zhao [24]	Simulation	Develop a real-time 3D volumetric model and improve quality grade accuracy.	Triangular prism model, sub-prism partitioning, adaptive reclaiming, and blending.	LMS200 and O1D100 laser scanners	2D laser range finder data for stockpile profiling.	Stockpile shape, quality grades, angle of repose, density.	R2 for regression accuracy, SSE for model evaluation.
8	Ji Wu [14]	Optimization	Improve ore blending to improve recovery and cost efficiency.	ABC algorithm for multi-objective optimization; software developed for blending.	Algorithm implemented in MATLAB	Historical data from Shizhuyuan Mine, focusing on ore recovery.	Recovery, mining costs, production capacity.	Recovery rate increased from 0.45% to 0.491%.
9	Dahee Jung [25]	Review	Review ML applications in mining and analyze trends.	Systematic literature review of 109 papers; keyword search; selection criteria defined.	Not addressed	Various datasets, including open-source data from RapidEye.	Data sources, utilization types.	ML performance evaluated; deep learning average RMSE: 0.000608.
10	Shi Qiang Liu [26]	Review	Review ML applications in open-pit mining to improve predictability and feasibility	Statistical analysis of 100+ papers (2012–2023), categorized into four key areas	Sensors for real-time drilling, hyperspectral data analysis	Operational resources, equipment utilization, mine geology, production, environment	Block extraction, tonnage (tons), volume (cubic meters), destination assignment	Not addressed
11	Larissa Statsenko [27]	Simulation	Develop an info system for ore blending to stabilize output quality.	Imitation modeling and simulation for ore blending design; decision support system applied.	Not addressed	Datasets from four open-pits in Kazakhstan.	Stockpile dimensions, transport mass, Fe-grade.	Ore grades variance reduced  4 times; blending coefficient improved 1.5–2 times.
12	Parent, E. [28]	Regression	Evaluates infrared (IR) models for soil PSD using machine learning (SGB) to predict sand, silt, clay, and carbon content. The 4X gain method achieves R2 values between 0.87 and 0.95; NIR and MIR spectra give R2 between 0.40 and 0.94.	SGB applied to NIR and MIR spectra; R2 values: 0.40–0.94 for NIR, 0.45–0.80 for MIR, 4X gain method: R2 = 0.87–0.95.	Stochastic Gradient Boosting (SGB)	1298 soil samples; R2 for sand sieving: 0.66, sedimentation: 0.53, NIR: 0.69–0.92, MIR: 0.45–0.80, SDF and settling time: 0.84–0.92.	Sand, silt, clay, carbon content, SDF, settling time, NIR, MIR spectra.	4X gain: R2 = 0.87–0.95; NIR and MIR: R2 = 0.40–0.94; Settling time and SDF: R2 = 0.84–0.92. Machine learning enhances PSD prediction.
13	Otsuki, A. et al. [12]	Regression	Investigate ANNs for predicting PSD to reduce energy consumption.	Supervised ANN trained on experimental data.	Rod mill for grinding, PSD measurement of coal.	Experimental data with varying feed vol.%, rod load vol.%, and grinding time.	Feed volume%, rod load, grinding time.	ANN predicted PSD well; RMSE ranged from 0.165 to 0.965, showing its ability to predict PSD under varying conditions.

Table 3. Machine Learning applied to blending. Part 2.

N	Author	Model Type	Objectives	Methodology	Instrumentation	Data Used	Key Variables	Results
14	Shubham Shrivastava [29]	Deep Learning	Develop an AI method to predict PSD from dump images.	Mask R-CNN with ResNet50 and FPN trained on dump images.	Mask R-CNN with ResNet50 and FPN for image segmentation.	Images from iron ore dump, particles 5 mm to larger.	Particle size distribution, image resolution, segmentation accuracy.	Mask R-CNN: 0.936 accuracy, 0.829 loss, predicts 250–300 particles/image, 1.274 mm error; predicts wide range of particles.
15	Both, Christian [13]	Regression	Predict throughput at Tropicana Gold Mining using machine learning.	Comparison of neural network models and linear regression for throughput prediction.	Data from blasthole drilling, comminution circuit, ball mill power, product particle size.	Production data including hardness proportions, ball mill power, particle size.	Hardness proportions, ball mill power, product particle size.	Neural network reduces RMSE by 10.6%, improves accuracy with ball mill power and particle size; hardness proportions enhance prediction by 6.3%.
16	Ghasemi, Zahra [18]	Regression	Develop a hybrid framework to optimize SAG mill throughput using machine learning and evolutionary algorithms.	Evaluated 17 models with 36,743 records, applying feature selection and three evolutionary algorithms (Differential Evolution, GA, PSO). CatBoost was the most accurate, and Differential Evolution was the best optimization algorithm.	SAG mill operational data including throughput and parameters.	36,743 records from SAG mill operations.	CatBoost, Differential Evolution, throughput, parameters.	CatBoost achieved R2 = 0.82, with Differential Evolution providing robust throughput predictions. Key metrics: RMSE = 49.6, MAE = 36.3, EVS = 0.645, R2 = 0.82.
17	Saldaña, Manuel [19]	Regression	Optimize SAG mill grinding in Chilean copper mining using machine learning (ANN, Random Forest, GBM) to analyze the impact of fragmentation, mill power, and liner age on throughput.	Compares multiple models (ANN, Random Forest, GBM, XgBoost, linear regression) for throughput prediction and mill parameter optimization.	Production data from Chilean copper mining, including mill power, fragmentation, and liner age.	Production data, mill power, fragmentation, liner age, energy consumption, throughput (TpH).	ANN, Random Forest, GBM, XgBoost, Linear Regression	ANN performed best with R2 = 0.89. GBM and XgBoost achieved R2 = 75.46% and 77.18%, while linear regression had R2 = 0.55. ANN increased production by 4.41% and reduced energy consumption by 7.62%. Machine learning models outperformed linear regression for throughput prediction.

Table 3. Machine Learning applied to blending. Part 2.

N	Author	Model Type	Objectives	Methodology	Instrumentation	Data Used	Key Variables	Results
14	Shubham Shrivastava [29]	Deep Learning	Develop an AI method to predict PSD from dump images.	Mask R-CNN with ResNet50 and FPN trained on dump images.	Mask R-CNN with ResNet50 and FPN for image segmentation.	Images from iron ore dump, particles 5 mm to larger.	Particle size distribution, image resolution, segmentation accuracy.	Mask R-CNN: 0.936 accuracy, 0.829 loss, predicts 250–300 particles/image, 1.274 mm error; predicts wide range of particles.
15	Both, Christian [13]	Regression	Predict throughput at Tropicana Gold Mining using machine learning.	Comparison of neural network models and linear regression for throughput prediction.	Data from blasthole drilling, comminution circuit, ball mill power, product particle size.	Production data including hardness proportions, ball mill power, particle size.	Hardness proportions, ball mill power, product particle size.	Neural network reduces RMSE by 10.6%, improves accuracy with ball mill power and particle size; hardness proportions enhance prediction by 6.3%.
16	Ghasemi, Zahra [18]	Regression	Develop a hybrid framework to optimize SAG mill throughput using machine learning and evolutionary algorithms.	Evaluated 17 models with 36,743 records, applying feature selection and three evolutionary algorithms (Differential Evolution, GA, PSO). CatBoost was the most accurate, and Differential Evolution was the best optimization algorithm.	SAG mill operational data including throughput and parameters.	36,743 records from SAG mill operations.	CatBoost, Differential Evolution, throughput, parameters.	CatBoost achieved R2 = 0.82, with Differential Evolution providing robust throughput predictions. Key metrics: RMSE = 49.6, MAE = 36.3, EVS = 0.645, R2 = 0.82.
17	Saldaña, Manuel [19]	Regression	Optimize SAG mill grinding in Chilean copper mining using machine learning (ANN, Random Forest, GBM) to analyze the impact of fragmentation, mill power, and liner age on throughput.	Compares multiple models (ANN, Random Forest, GBM, XgBoost, linear regression) for throughput prediction and mill parameter optimization.	Production data from Chilean copper mining, including mill power, fragmentation, and liner age.	Production data, mill power, fragmentation, liner age, energy consumption, throughput (TpH).	ANN, Random Forest, GBM, XgBoost, Linear Regression	ANN performed best with R2 = 0.89. GBM and XgBoost achieved R2 = 75.46% and 77.18%, while linear regression had R2 = 0.55. ANN increased production by 4.41% and reduced energy consumption by 7.62%. Machine learning models outperformed linear regression for throughput prediction.

4.1. Blending and Ore Variability

As shown in Table 4, blending plays a critical role in stabilizing ore composition and improving downstream processing performance. Studies summarized in Table 4 highlight the need to minimize variability in ore type, particle size distribution (PSD), and flow rates, all of which impact blend uniformity. For example, Bicak et al. [30] and Mkurazhizha [31] examine how fluctuations in feed composition affect flotation efficiency and comminution performance, reinforcing the importance of consistent blending to maintain process stability.

Importantly, PSD is measured upstream, prior to the SAG mill, and serves as a key feedforward control input rather than only as an output variable of milling performance. This proactive use of PSD allows blending and feed adjustments to reduce variability entering the grinding circuit, supporting enhanced mill efficiency and product quality.

Table 4. Key studies on blending and ore variability, including ore grade variability, homogenization, and blending optimization.

No.	Author	Year	Objective/Goal	Methodology	Instrumentation	Results	Conclusion
1	Ozlem Bicak [30]	2019	To study ore variability in sulfide ores and its impact on flotation.	Oxidation Index using EDTA extraction.	ToF-SIMS, XPS	High Oxidation Index correlates with copper-enriched ore contamination and poor flotation recovery.	Oxidation Index is a practical, cost-effective tool for geometallurgical mapping and flotation prediction.
2	Yue Xie, Frank Neumann, Aneta Neumann [32]	2021	To optimize stockpile blending under uncertainty using chance-constrained programming.	Nonlinear optimization model with chance constraints, Differential Evolution (DE) algorithm.	No specific instrumentation mentioned.	DE algorithm outperforms deterministic models, with chance constraints affecting success rates.	Chance-constrained optimization effectively manages uncertainties, with copper grade constraints significantly influencing results.
3	Diego Marques, João Felipe C. Costa [33]	2013	To simulate ore grade variability in blending and homogenization for improved processing efficiency.	Geostatistical simulations using the Turning Bands Method (TBA), 3D grade block model.	No specific instrumentation mentioned.	Homogenization reduces silica (SiO2) variability, with effectiveness tied to pile size and layer count.	Optimal homogenization parameters are selected based on pile mass and layer count for better processing results.
4	Shi Zhao, Tien-Fu Lu, Ben Koch, Alan Hurdsman [24]	2015	To develop a real-time method for stockpile quality calculation.	3D voxel model, laser scanning, BWR model.	Laser scanner, 3DOF scanner, field trials for UKF localization.	3D voxel model improves stockpile quality and geometry predictions.	Limitations: scanning accuracy, high computation time, and BWR reclaiming pattern. Real-time chemical data is often unavailable.
5	Yue Xie [32]	2021	To improve large-scale stockpile blending and metal grades using the DE algorithm.	Continuous improvement model with DE algorithm and repair operators.	DE algorithm, repair operators.	DE algorithm improves metal volume over traditional methods.	DE algorithm enhances stockpile blending and mine scheduling efficiency.
6	Mkurazhizha, H. Huggins [31]	2018	To study ore blending effects on comminution and product quality at LKAB.	Investigates blending effects on grinding and product quality, focusing on high tonnage operations.	Laboratory-scale milling, energy consumption, particle size distribution.	Binary blends more efficient than tertiary blends.	Ore blending improves energy use, reduces maintenance, and enhances product quality.
7	Mohamed Farghaly [34]	2021	To optimize blending and production processes using linear programming.	Linear programming model to optimize costs and production readiness.	Linear programming model, LINDO TM software.	Optimal model reduced costs to 194.183 million, highlighting key decisions for blending, batch size, and transportation.	Model improves profitability, verified with Cu and Co production data.

Table 4. Key studies on blending and ore variability, including ore grade variability, homogenization, and blending optimization.

No.	Author	Year	Objective/Goal	Methodology	Instrumentation	Results	Conclusion
1	Ozlem Bicak [30]	2019	To study ore variability in sulfide ores and its impact on flotation.	Oxidation Index using EDTA extraction.	ToF-SIMS, XPS	High Oxidation Index correlates with copper-enriched ore contamination and poor flotation recovery.	Oxidation Index is a practical, cost-effective tool for geometallurgical mapping and flotation prediction.
2	Yue Xie, Frank Neumann, Aneta Neumann [32]	2021	To optimize stockpile blending under uncertainty using chance-constrained programming.	Nonlinear optimization model with chance constraints, Differential Evolution (DE) algorithm.	No specific instrumentation mentioned.	DE algorithm outperforms deterministic models, with chance constraints affecting success rates.	Chance-constrained optimization effectively manages uncertainties, with copper grade constraints significantly influencing results.
3	Diego Marques, João Felipe C. Costa [33]	2013	To simulate ore grade variability in blending and homogenization for improved processing efficiency.	Geostatistical simulations using the Turning Bands Method (TBA), 3D grade block model.	No specific instrumentation mentioned.	Homogenization reduces silica (SiO2) variability, with effectiveness tied to pile size and layer count.	Optimal homogenization parameters are selected based on pile mass and layer count for better processing results.
4	Shi Zhao, Tien-Fu Lu, Ben Koch, Alan Hurdsman [24]	2015	To develop a real-time method for stockpile quality calculation.	3D voxel model, laser scanning, BWR model.	Laser scanner, 3DOF scanner, field trials for UKF localization.	3D voxel model improves stockpile quality and geometry predictions.	Limitations: scanning accuracy, high computation time, and BWR reclaiming pattern. Real-time chemical data is often unavailable.
5	Yue Xie [32]	2021	To improve large-scale stockpile blending and metal grades using the DE algorithm.	Continuous improvement model with DE algorithm and repair operators.	DE algorithm, repair operators.	DE algorithm improves metal volume over traditional methods.	DE algorithm enhances stockpile blending and mine scheduling efficiency.
6	Mkurazhizha, H. Huggins [31]	2018	To study ore blending effects on comminution and product quality at LKAB.	Investigates blending effects on grinding and product quality, focusing on high tonnage operations.	Laboratory-scale milling, energy consumption, particle size distribution.	Binary blends more efficient than tertiary blends.	Ore blending improves energy use, reduces maintenance, and enhances product quality.
7	Mohamed Farghaly [34]	2021	To optimize blending and production processes using linear programming.	Linear programming model to optimize costs and production readiness.	Linear programming model, LINDO TM software.	Optimal model reduced costs to 194.183 million, highlighting key decisions for blending, batch size, and transportation.	Model improves profitability, verified with Cu and Co production data.

4.2. Real-Time Monitoring and Sensor Technologies

Effective blending relies on continuous monitoring of ore properties. Table 5 summarizes studies focusing on sensor technologies and real-time monitoring methods for improving blend control. These works explore various instruments including laser scanners, ultrasonic sensors, machine vision, and geostatistical modeling, which provide critical data to enable dynamic feed adjustments. For instance, Zhao et al. [24] and Benndorf [35] demonstrate the value of integrating sensor data with modeling frameworks to reduce uncertainty and enhance operational decisions.

It is important to note that NIR spectroscopy is primarily utilized for downstream viscosity control by monitoring clay content and ore composition that affect slurry rheology. While NIR data can occasionally inform upstream decisions such as segregation or bypass routing under specific conditions, it is not typically employed as a direct sensor for blending control.

The author in [36] demonstrates this through a sensor-driven blending system that updates stockpile quality in real time. Similarly, Refs. [32,37] apply heuristic and differential evolution algorithms to maintain optimal blends under changing conditions, and ref. [34] uses linear programming to stabilize grade delivery.

Table 5. Key studies on real-time monitoring and sensor technologies for blending control.

No.	Author	Year	Objective/Goal	Methodology	Instrumentation	Results	Conclusion
1	Shi Zhao, Tien-Fu Lu, Ben Koch, Alan Hurdsman [24]	2015	To develop a real-time method for stockpile quality calculation.	3D voxel model, laser scanning, BWR model.	Laser scanner, 3DOF scanner, field trials for UKF localization.	3D voxel model improves stockpile quality and geometry predictions.	Limitations: scanning accuracy, high computation time, and BWR reclaiming pattern. Real-time chemical data is often unavailable.
2	Z. Ye, M. Hilden, M. Yahyaei [38]	2022	To develop a 3D cellular automaton model for simulating ore pile formation and segregation.	3D CA model for ore pile formation, size segregation.	No specific instrumentation.	Model simulates ore pile formation and segregation; validated with laboratory-scale data.	Limitations include inability to represent conical piles and neglect of compaction effects.
3	Z. Ye, M.M. Hilden, M. Yahyaei [39]	2023	To develop a fast 3D cellular automata-based stockpile model for process control.	3D CA model for real-time stockpile monitoring and size segregation.	Ultrasonic height sensors, machine vision-based particle-size analysis.	Model simulates stockpile behavior accurately, validated with industrial data.	Enhances process control, but assumptions about free-flowing ore and compaction limit its effectiveness.
4	Alexander Bowler, Serafim Bakalis, Nicholas Watson [40]	2020	To use ultrasonic sensors and machine learning for optimizing mixing processes.	Ultrasonic sensors and machine learning for non-invasive monitoring of mixing.	Ultrasonic sensors, machine learning, MATLAB R2019a.	High accuracy in classifying mixing states and predicting completion times.	Challenges in obtaining labeled data and achieving optimal performance across algorithms.
5	Jörg Benndorf [35]	2016	To improve mine production control with sensor data and geostatistical modeling.	Framework for reducing uncertainties and optimizing real-time decisions.	Sensor-based characterization, geostatistical modeling, Kalman Filter.	15%–40% improvement in production control and reduced uncertainty.	Reduces losses from discrepancies in production targets.
6	T. Wambeke, J. Benndorf [41]	2018	To study the influence of blending ratio and sensor precision on grade control models.	Simulation-based geostatistical approach for grade control model reconciliation.	Sensors for measuring blended material streams, varying measurement errors.	RMSE reductions of 20%–60% in grade control models.	Highlights the importance of sensor accuracy and blending ratios.

Table 5. Key studies on real-time monitoring and sensor technologies for blending control.

No.	Author	Year	Objective/Goal	Methodology	Instrumentation	Results	Conclusion
1	Shi Zhao, Tien-Fu Lu, Ben Koch, Alan Hurdsman [24]	2015	To develop a real-time method for stockpile quality calculation.	3D voxel model, laser scanning, BWR model.	Laser scanner, 3DOF scanner, field trials for UKF localization.	3D voxel model improves stockpile quality and geometry predictions.	Limitations: scanning accuracy, high computation time, and BWR reclaiming pattern. Real-time chemical data is often unavailable.
2	Z. Ye, M. Hilden, M. Yahyaei [38]	2022	To develop a 3D cellular automaton model for simulating ore pile formation and segregation.	3D CA model for ore pile formation, size segregation.	No specific instrumentation.	Model simulates ore pile formation and segregation; validated with laboratory-scale data.	Limitations include inability to represent conical piles and neglect of compaction effects.
3	Z. Ye, M.M. Hilden, M. Yahyaei [39]	2023	To develop a fast 3D cellular automata-based stockpile model for process control.	3D CA model for real-time stockpile monitoring and size segregation.	Ultrasonic height sensors, machine vision-based particle-size analysis.	Model simulates stockpile behavior accurately, validated with industrial data.	Enhances process control, but assumptions about free-flowing ore and compaction limit its effectiveness.
4	Alexander Bowler, Serafim Bakalis, Nicholas Watson [40]	2020	To use ultrasonic sensors and machine learning for optimizing mixing processes.	Ultrasonic sensors and machine learning for non-invasive monitoring of mixing.	Ultrasonic sensors, machine learning, MATLAB R2019a.	High accuracy in classifying mixing states and predicting completion times.	Challenges in obtaining labeled data and achieving optimal performance across algorithms.
5	Jörg Benndorf [35]	2016	To improve mine production control with sensor data and geostatistical modeling.	Framework for reducing uncertainties and optimizing real-time decisions.	Sensor-based characterization, geostatistical modeling, Kalman Filter.	15%–40% improvement in production control and reduced uncertainty.	Reduces losses from discrepancies in production targets.
6	T. Wambeke, J. Benndorf [41]	2018	To study the influence of blending ratio and sensor precision on grade control models.	Simulation-based geostatistical approach for grade control model reconciliation.	Sensors for measuring blended material streams, varying measurement errors.	RMSE reductions of 20%–60% in grade control models.	Highlights the importance of sensor accuracy and blending ratios.

4.3. Computational and Modeling Approaches

To address the complexity of blending and stockpile management, several studies apply computational and data-driven modeling approaches, as summarized in Table 6. These studies utilize cellular automata, digital twins, genetic algorithms, and geostatistical methods to simulate stockpile behavior, optimize blending strategies, and reconcile grade control models. Servin et al. [42] and Wu [14] highlight the potential of digital twin frameworks and automated blending methods to improve predictive accuracy and operational efficiency.

The research conducted by Servin et al. [42] implements digital twins to enable continuous material tracking across crushers, conveyors, and stockpiles. This allows scenario analysis and proactive decision-making before material reaches the mill. In a similar line, Bowler et al. [40] integrate ultrasonic sensors with machine learning to control mixing at the reclaim feeder stage, where blended material is retrieved from the stockpile for SAG mill feed.

Other contributions, such as Benndorf et al. [35] and Wambeke et al. [41], focus on feed prediction and reconciliation using geostatistics and sensor fusion. These techniques are commonly implemented during ore dispatch and stockpile reclaim, helping to align planned and actual feed conditions. Additionally, recent IEEE conference proceedings offer valuable insights into data-driven AI models and control strategies for grinding circuits and particle size prediction [43,44,45], complementing journal-based research with applied engineering solutions.

Altogether, these computational frameworks enhance blending adaptability and precision, especially when integrated with real-time feedback from SCADA systems and online analyzers.

Table 6. Key computational and modeling studies addressing stockpile behavior and blending optimization.

No.	Author	Year	Objective/Goal	Methodology	Instrumentation	Results	Conclusion
1	Z. Ye, M. Hilden, M. Yahyaei [38]	2022	To develop a 3D cellular automaton model for simulating ore pile formation and segregation.	3D CA model for ore pile formation, size segregation.	No specific instrumentation.	Model simulates ore pile formation and segregation; validated with laboratory-scale data.	Limitations include inability to represent conical piles and neglect of compaction effects.
2	Z. Ye, M.M. Hilden, M. Yahyaei [39]	2023	To develop a fast 3D cellular automata-based stockpile model for process control.	3D CA model for real-time stockpile monitoring and size segregation.	Ultrasonic height sensors, machine vision-based particle-size analysis.	Model simulates stockpile behavior accurately, validated with industrial data.	Enhances process control, but assumptions about free-flowing ore and compaction limit its effectiveness.
3	Martin Servin, Folke Vesterlund, Erik Wallin [42]	2021	To explore digital twins with particle simulation for material tracking.	Digital twins, distributed particle simulation for tracking material in mine-to-mill processes.	Sensors, simulation models, AGX Dynamics, MWD data, vehicle telematics.	Particle-based digital twin improves material tracking and process efficiency.	Limitations in supporting size segregation and capturing stockpile surface shape.
4	J. Wu [14]	2022	To develop an open-pit automation ore blending method.	Constraint model for ore-matching, improving blending and reducing grade deviation.	Genetic algorithm, geostatistics for grade prediction, automated blending decisions.	Model improves ore quality and production rates.	Enhances efficiency and reduces costs in open-pit mining.
5	Jörg Benndorf [35]	2016	To improve mine production control with sensor data and geostatistical modeling.	Framework for reducing uncertainties and optimizing real-time decisions.	Sensor-based characterization, geostatistical modeling, Kalman filter.	15%–40% improvement in production control and reduced uncertainty.	Reduces losses from discrepancies in production targets.
6	William Reid [46]	2021	To study uncertainty in advanced ore mining using Maptek’s Evolution system.	Evolutionary computation and neural networks for quantifying uncertainty in extraction.	Evolutionary computation, geological uncertainty, staging approaches.	Demonstrates how staging methods impact NPV and profit.	Improved profitability through geological uncertainty modeling.
7	T Wambeke, J Benndorf [41]	2018	To study the influence of blending ratio and sensor precision on grade control models.	Simulation-based geostatistical approach for grade control model reconciliation.	Sensors for measuring blended material streams, varying measurement errors.	RMSE reductions of 20%–60% in grade control models.	Highlights the importance of sensor accuracy and blending ratios.

Table 6. Key computational and modeling studies addressing stockpile behavior and blending optimization.

No.	Author	Year	Objective/Goal	Methodology	Instrumentation	Results	Conclusion
1	Z. Ye, M. Hilden, M. Yahyaei [38]	2022	To develop a 3D cellular automaton model for simulating ore pile formation and segregation.	3D CA model for ore pile formation, size segregation.	No specific instrumentation.	Model simulates ore pile formation and segregation; validated with laboratory-scale data.	Limitations include inability to represent conical piles and neglect of compaction effects.
2	Z. Ye, M.M. Hilden, M. Yahyaei [39]	2023	To develop a fast 3D cellular automata-based stockpile model for process control.	3D CA model for real-time stockpile monitoring and size segregation.	Ultrasonic height sensors, machine vision-based particle-size analysis.	Model simulates stockpile behavior accurately, validated with industrial data.	Enhances process control, but assumptions about free-flowing ore and compaction limit its effectiveness.
3	Martin Servin, Folke Vesterlund, Erik Wallin [42]	2021	To explore digital twins with particle simulation for material tracking.	Digital twins, distributed particle simulation for tracking material in mine-to-mill processes.	Sensors, simulation models, AGX Dynamics, MWD data, vehicle telematics.	Particle-based digital twin improves material tracking and process efficiency.	Limitations in supporting size segregation and capturing stockpile surface shape.
4	J. Wu [14]	2022	To develop an open-pit automation ore blending method.	Constraint model for ore-matching, improving blending and reducing grade deviation.	Genetic algorithm, geostatistics for grade prediction, automated blending decisions.	Model improves ore quality and production rates.	Enhances efficiency and reduces costs in open-pit mining.
5	Jörg Benndorf [35]	2016	To improve mine production control with sensor data and geostatistical modeling.	Framework for reducing uncertainties and optimizing real-time decisions.	Sensor-based characterization, geostatistical modeling, Kalman filter.	15%–40% improvement in production control and reduced uncertainty.	Reduces losses from discrepancies in production targets.
6	William Reid [46]	2021	To study uncertainty in advanced ore mining using Maptek’s Evolution system.	Evolutionary computation and neural networks for quantifying uncertainty in extraction.	Evolutionary computation, geological uncertainty, staging approaches.	Demonstrates how staging methods impact NPV and profit.	Improved profitability through geological uncertainty modeling.
7	T Wambeke, J Benndorf [41]	2018	To study the influence of blending ratio and sensor precision on grade control models.	Simulation-based geostatistical approach for grade control model reconciliation.	Sensors for measuring blended material streams, varying measurement errors.	RMSE reductions of 20%–60% in grade control models.	Highlights the importance of sensor accuracy and blending ratios.

4.4. Operational Constraints and Sustainability Challenges

Operational and sustainability challenges in blending are explored in Table 7. These studies investigate nonlinear scheduling models, particle size segregation in waste rock, ore fragmentation, and the impacts of automation on workload and safety. Campos et al. [47] and Qiu and Pabst [48] highlight how blending effectiveness depends on operational constraints and environmental factors, underscoring the need for integrated control and monitoring strategies.

The author in [47] proposes a nonlinear model to manage short-term feed variability, while [49] highlights how ore fragmentation influences blending effectiveness, reinforcing the need to coordinate upstream and downstream decisions. Ref. [48] points to the long-term sustainability risks of poor material handling, including particle size segregation in waste piles.

Automation technologies can help address some of these challenges. Programmable logic controllers (PLCs) manage local control actions, while supervisory control and data acquisition (SCADA) systems centralize monitoring. Real-time sensors, such as level and flow transmitters, provide continuous feedback, and vision systems like PSD cameras enable automated quality checks. Refs. [50,51] show how these technologies improve blending consistency, equipment use, and safety.

Predictive control, where model outputs are used to adjust operations in real time, offers further potential. Yet, its adoption remains limited, primarily due to system integration complexity and the need for transparent, operator-friendly interfaces.

4.5. Discussion

From Table 1, Table 2 and Table 3, we identify key machine learning trends and methodologies aligned with the research questions outlined in Section 3.1. These studies reflect the growing use of ML models in ore blending and mineral processing to support real-time prediction and control of variables such as throughput, grinding efficiency, and particle size distribution. Algorithms like regression models, neural networks, gradient boosting, and hybrid optimization methods are applied to enhance blending performance under dynamic conditions.

Table 7. Key studies on operational constraints and sustainability challenges in blending and stockpile management.

No.	Author	Year	Objective/Goal	Methodology	Instrumentation	Results	Conclusion
1	Pedro Henrique Alves Campos [47]	2024	To optimize mine scheduling and metallurgical recovery with nonlinear blending models and simulated annealing.	Nonlinear models, simulated annealing.	Simulated annealing, blending models.	Improves recovery and efficiency.	Penalty factor lacks physical meaning, and linear models reduce effectiveness.
2	Peiyong Qiu, Thomas Pabst [48]	2023	To analyze particle size segregation in a waste rock pile.	Image analysis for segregation.	Drone, 87 photos, waste rock pile.	Segregation increases from top to bottom, with lateral heterogeneity.	Impacts hydraulic conductivity and geotechnical properties.
3	Ayyub Nikkhah [49]	2022	To optimize ore fragmentation from blasting at Sarcheshmeh copper mine.	Blasting parameter optimization.	Statistical analysis, fragmentation efficiency.	Tailored blasting improves performance and reduces costs.	Enhances mining efficiency, further research needed.
4	Loreto Codoceo-Contreras [51]	2024	To explore automation impacts in mining through NLP review of 94 documents.	NLP review, thematic analysis.	94 documents, TF-IDF, anomaly detection.	Identifies themes like workload and communication.	Recommends research on community impacts and cybersecurity.

Table 7. Key studies on operational constraints and sustainability challenges in blending and stockpile management.

No.	Author	Year	Objective/Goal	Methodology	Instrumentation	Results	Conclusion
1	Pedro Henrique Alves Campos [47]	2024	To optimize mine scheduling and metallurgical recovery with nonlinear blending models and simulated annealing.	Nonlinear models, simulated annealing.	Simulated annealing, blending models.	Improves recovery and efficiency.	Penalty factor lacks physical meaning, and linear models reduce effectiveness.
2	Peiyong Qiu, Thomas Pabst [48]	2023	To analyze particle size segregation in a waste rock pile.	Image analysis for segregation.	Drone, 87 photos, waste rock pile.	Segregation increases from top to bottom, with lateral heterogeneity.	Impacts hydraulic conductivity and geotechnical properties.
3	Ayyub Nikkhah [49]	2022	To optimize ore fragmentation from blasting at Sarcheshmeh copper mine.	Blasting parameter optimization.	Statistical analysis, fragmentation efficiency.	Tailored blasting improves performance and reduces costs.	Enhances mining efficiency, further research needed.
4	Loreto Codoceo-Contreras [51]	2024	To explore automation impacts in mining through NLP review of 94 documents.	NLP review, thematic analysis.	94 documents, TF-IDF, anomaly detection.	Identifies themes like workload and communication.	Recommends research on community impacts and cybersecurity.

Machine learning and artificial intelligence (AI) continue to expand their role in modern mining operations [52], offering tools for predictive modeling, process optimization, and sustainability improvement. These applications highlight the shift from reactive to data-driven decision-making in mining.

Several studies adopt ensemble and deep learning techniques to model nonlinear process behavior and improve feed quality. The use of hybrid methods, combining models like CatBoost with evolutionary optimization, or deep learning architectures such as LSTM networks, has shown potential for handling time-series data and capturing long-term dependencies. These approaches are particularly useful in grinding circuits and dynamic blending scenarios.

The models reviewed in Table 1 represent a wide range of approaches to solving key operational problems in mineral processing, particularly in predicting particle size distribution (PSD), optimizing ore blending, and forecasting throughput. Across these studies, performance indicators such as high R² values and reduced RMSE confirm the capability of machine learning techniques to extract meaningful patterns from complex operational data.

In PSD prediction, models like Artificial Neural Networks (ANNs) and Long Short-Term Memory (LSTM) networks have achieved high precision under controlled conditions [12,17]. These models effectively capture non-linear dependencies between variables such as grinding time, feed characteristics, and size output. However, the datasets used in these studies are often well structured, with consistent sampling and limited noise, conditions not always mirrored in plant environments. In industrial settings, factors like ore variability, sensor drift, and asynchronous measurements introduce noise that can compromise model robustness.

For blending optimization, algorithms such as Artificial Bee Colony and Goal Programming offer structured frameworks to support decision-making under constraints [14,15]. These approaches focus not just on prediction but on selecting input combinations that meet multiple objectives, such as grade control, recovery, and cost minimization.

Among more general-purpose learning methods, Random Forest strikes a practical balance between predictive accuracy and usability in real environments. Its tolerance for missing or noisy data, combined with minimal preprocessing requirements, makes it particularly attractive for deployment in supervisory applications. In [19], Random Forest and related models were successfully applied to SAG mill data to improve throughput prediction and energy efficiency, demonstrating the model’s adaptability under variable conditions.

Despite these promising results, the deployment of machine learning models at scale is often limited by inconsistent data practices. Studies such as [12,18,19] differ widely in their use of variables, measurement units, preprocessing steps, and data quality control. These discrepancies hinder not only model transfer between operations but also comparative benchmarking and reproducibility. Even models with strong local performance may fail to generalize if retrained on incompatible datasets. Addressing this issue will require industry-wide adoption of standardized data acquisition protocols, consistent variable naming conventions, and clear preprocessing guidelines. These steps are essential to enable scalable, trustworthy, and transferable predictive tools for blending control.

The values shown in Figure 3 represent the percentage distribution of machine learning methods applied across the reviewed studies in the context of ore blending within mining operations. This visualization highlights the relative popularity of various techniques such as Artificial Neural Networks (ANNs), Random Forest (RF), Gradient Boosting Machines (GBMs), and Support Vector Machines (SVMs). All of these approaches are used to address regression problems, which involve predicting continuous numerical values based on input features [53]. The term regression models in this context, therefore, refers broadly to any algorithm capable of learning such mappings, including both traditional methods (e.g., linear regression and SVM) and more complex approaches like ensemble learning (e.g., RF and GBM) or deep learning architectures (e.g., ANN and LSTM). This category of models is selected depending on the nature of the problem, data dimensionality, input types, and complexity.

While these methods are all within the regression family, RF is specifically emphasized in this figure due to its notably frequent use in the reviewed literature. Random Forest is a robust ensemble-based model that handles high-dimensional, nonlinear data effectively and is widely adopted in industrial applications for its interpretability and performance. Among the identified techniques, ANN and RF are the most frequently employed, indicating a strong preference for models capable of capturing the complex, nonlinear relationships inherent in blending processes. The increasing use of deep learning models, particularly Long Short-Term Memory (LSTM) networks, further reflects a growing interest in modeling temporal patterns and dynamic system behavior in this domain.

These methods have been applied to a wide range of blending-related tasks, including particle size distribution (PSD) prediction, grade control, moisture content estimation, and throughput forecasting. Despite the promising results reported in many studies, challenges remain regarding data quality, feature selection, and the generalization of models across different ore types and processing environments. The prevalence of certain methods also points to potential research gaps, particularly in the exploration of hybrid or ensemble approaches and the integration of domain knowledge into model architectures.

Figure 4 illustrates the evolving adoption of machine learning and optimization techniques for blending optimization over recent years. The chart shows that Artificial Neural Networks (ANNs) and Random Forest (RF) continue to be the most frequently applied methods, consistently appearing across multiple studies. In parallel, more recent techniques, such as Gradient Boosting Machines (GBMs), CatBoost, and deep learning architectures like Long Short-Term Memory (LSTM) networks, are gaining prominence. Additionally, metaheuristic algorithms including Artificial Bee Colony (ABC) and the Non-dominated Sorting Genetic Algorithm II (NSGA-II) have been increasingly applied to address the complex nature of blending optimization tasks. The Artificial Bee Colony (ABC) algorithm is a population-based metaheuristic inspired by the intelligent foraging behavior of honey bee swarms. It divides the bee population into employed bees, onlookers, and scouts, each playing a role in exploring and exploiting the solution space. ABC is particularly effective for solving nonlinear, high-dimensional, and multi-modal optimization problems due to its simplicity, adaptability, and global search capabilities. The Non-dominated Sorting Genetic Algorithm II (NSGA-II) is an evolutionary algorithm widely used for multi-objective optimization. It maintains a diverse set of solutions through a fast non-dominated sorting process and employs a crowding distance mechanism to ensure solution diversity. NSGA-II is especially suited for scenarios involving trade-offs between competing objectives, such as maximizing product quality while minimizing processing costs. CatBoost is a gradient boosting algorithm specifically designed to handle categorical features efficiently without extensive preprocessing. It leverages ordered boosting and advanced regularization techniques to reduce overfitting, offering high performance and robustness in industrial prediction tasks. LSTM (Long Short-Term Memory) networks are a specialized type of recurrent neural network (RNN) capable of learning long-term dependencies in sequential data. Their architecture includes memory cells and gating mechanisms that regulate the flow of information, making them particularly advantageous for capturing the temporal dynamics present in blending and other time-dependent processes.

Together, these trends reflect a growing reliance on real-time sensor data and sophisticated analytics, underscoring the increasing integration of machine learning and advanced optimization methods into operational control systems and future industrial implementations.

5. Conclusions

This review highlights the critical role of blending in improving downstream performance in mineral processing operations such as milling and flotation. The effectiveness of blending is influenced by key variables including ore type, grade, particle size distribution (PSD), moisture content, and flow rates. It is important to clarify that these variables emerge from a subset of the reviewed studies rather than representing the entire dataset, reflecting common factors emphasized in the literature. A multidisciplinary combination of advanced modeling, simulation, sensor integration, and predictive analytics is essential to achieve more consistent and optimized blending outcomes.

Simulation models, integer programming, and geo-metallurgical approaches have demonstrated success in stabilizing ore grades and minimizing variability. However, geo-metallurgical approaches were not deeply covered in the current literature corpus and thus remain an important area for future research to explore more comprehensively. Meanwhile, sensor technologies such as Near-Infrared (NIR) spectroscopy, hyperspectral imaging, and X-ray fluorescence (XRF) enable real-time monitoring of ore properties, providing actionable feedback for dynamic control of blending operations. Machine learning (ML) methods, particularly Artificial Neural Networks (ANNs), Random Forest (RF), and Decision Trees (DTs), have shown significant promise for predicting throughput, PSD, and blending performance using both historical and real-time data. These models reduce reliance on manual heuristics, enhance decision-making, and improve resource efficiency. Hybrid approaches that combine deep learning with traditional ML methods are expected to deliver even more accurate and robust predictions in future implementations.

However, as observed across multiple studies, the lack of data standardization remains a significant barrier to wider adoption. Inconsistencies in data collection frequency, variable labeling, and measurement techniques hinder the comparability and transferability of predictive models. Establishing standardized data acquisition protocols is essential for ensuring model reliability, facilitating cross-site deployment, and enabling more reproducible results.

Looking ahead, future research should focus on refining ML models, integrating them with PI System infrastructure and real-time monitoring tools, and advancing fault detection capabilities. Prioritizing data standardization and transparency will be key to ensuring scalable and sustainable blending optimization across diverse mining contexts. By combining predictive technologies with robust automation, mining operations can achieve improved ore grade stability, reduced variability, and enhanced overall process efficiency.