Development of a Digital Twin Model for Stirred Milling Process by Integrating Machine Learning Models and Discrete Element Method Simulations
PI: Olumide Ogunmodimu (EMS)
Additional Team Members: Nelson Yaw Dzade, Department of Energy and Mineral Engineering; Arnold Barbara, College of Earth & Mineral Sciences
The students will be funded through EME teaching assistantship and support from the department.
The mining industry is required to operate in an increasingly challenging business environment with continuously increasing demand for raw materials and deeper mining. With the complexity of the geological structure of the ore deposits and decreasing ore grade, extensive high-quality mineralogical information is essential to integrate into the process plant’s operating performance. The field of manufacturing processes, specifically mineral processing, has witnessed significant advancements with the advent of digital twin technologies. This project aims to contribute to the evolution of digital twin applications by integrating a combined approach of machine learning models, including Support Vector Machines (SVM), Convolutional and Graph Neural Networks (CCNN, GNN), Physics-Informed Neural Networks (PINN), and Discrete Element Method (DEM) simulations. Digital twin adoption in fine milling will help develop a continuum framework from DEM simulation data. The continuum processes surrogate model will play a crucial role in realizing the overarching goal of efficient grinding by enabling accurate modeling, simulation of various operational conditions, prediction of performance metrics, design optimization, and real-time monitoring. These capabilities will empower engineers to continuously improve grinding processes, optimize equipment performance, and achieve efficient grinding processes. The overarching goal is to enhance the efficiency of grinding processes through a comprehensive understanding of the forces, stress, velocity, and position of particles during stirred milling. Leveraging digital twins to optimize granular flow parameters such as velocity, positions, and forces of particles leads to improved equipment performance, ore fragmentation, safety, environmental compliance, productivity, and cost savings, ultimately enhancing the overall efficiency and sustainability of mineral processing.
Project Overview: Considering the global commitment to achieving net-zero emissions, there arises a critical necessity for the extraction and processing of minerals to meet the current demand. This is related to producing entities for renewable energy systems, such as solar panels, energy storage, and transport materials such as lithium and copper. However, a significant proportion of available mineral deposits are of low grade and contain insufficient quantities of desired minerals. Consequently, the recovery of such low rade minerals necessitates a fine grinding process. This process is used to liberate minerals to fine levels (1–100m). The equipment widely used for this process is a stirred mill. Stirred mills are equipment that grinds minerals ores by attrition through shear stress from grinding media to the ore particles. A stirred mill consists of a chamber with a media agitator that can have spiral screws, pins, or discs, and the media can be either fluidized or agitated (Lichter & Davey, 2006). When the media is fluidized, it forms a cavity near the shaft and is displaced near the wall. The role of the agitator is to stir the media, which causes particle breakage. Notwithstanding, the fine grinding process is inherently expensive and demands substantial energy inputs. Remarkably, it has been documented that a mere 10% of the energy input into the grinding process contributes to the actual comminution of minerals, with the remaining 90% dissipated as heat and sound (Wills, 1992). This underscores the need to enhance the efficiency of the grinding process, which requires a comprehensive understanding of the intricate mechanisms governing the grinding process. To address this, it becomes paramount to develop models across various scales that offer precise predictions of mill performance under diverse grinding conditions. The use of analytical models is limited in understanding the dynamics associated with grinding conditions in relation to mineral extraction and comminution. Therefore, numerical methods have been used in the literature to study this process, and there have been some improvements. However, the mechanism of grinding and its associated parameters are yet to be fully understood. The use of machine learning (ML) and discrete element method (DEM), which embodies an intrinsic non-linear model, is recommended to be suitable for this study.
Project Goal and Research Approach: This research considers a novel approach that integrates ML with DEM. ML algorithms can predict optimal milling parameters such as feed rates, rotation speeds, and tool paths based on material properties, mill characteristics, and desired outcomes, leading to improved efficiency and quality in milling operations (Narayanan et.al., 2022). Whereas DEM simulations accurately replicate the behavior of individual particles, their interactions, and the impact of milling media on particle size reduction, enabling us to optimize mill design, operational parameters, and material properties for efficient milling. ML algorithms trained on DEM simulation data can predict milling performance metrics such as particle size distribution, energy consumption, and wear rates under different operating conditions, facilitating process optimization and decision-making in milling operations. These systems continuously monitor key parameters such as particle size, mill vibration, and power consumption, automatically adjusting milling parameters to maintain optimal performance and efficiency. The goal is to integrate machine learning and DEM simulations to develop a digital twin model for milling processes. Data from DEM simulations will be employed to create a continuum framework that aids in developing a surrogate model capable of monitoring fine milling, thus allowing for flexibility in varying milling parameters and observing the real-time effects on the materials being milled. By synergistically employing DEM and ML, we aim to scrutinize mill operational conditions and design parameters to optimize grinding efficiency. The culmination of this investigation will yield a digital twin of the grinding process, affording a real-time and accurate insight into the intricacies of the grinding operation. The concept of the digital twin refers to a virtual representation of the physical milling process, equipment (stirred mill), and operations (Zhuang et. al., 2021). That is the virtual representation of the stirred mill (milling equipment), materials and process, integration of data from various sources such as sensors, mill specifications (design parameters), historical data, and simulation of material feed rates, rotation speeds, particle interaction, and mills behavior to conduct real-time monitoring and analysis of the grinding mechanism and hence develop a predictive model for performance optimization.
Research significance: A digital twin model in the context of milling processes refers to a virtual representation of physically stirred mills, ores, and operations as described above It simulates the performance and efficiency of the stirred mill, including factors such as ores feed rates, rotation speeds, and particle interactions, allowing stakeholders (researchers or engineers) to monitor, analyze, and optimize milling operations in real-time. The digital twin model will play a pivotal role in predicting and optimizing the optimal speed, ores feed rates, and minimum energy required for ores to be milled at fine and ultra-fine levels. In a milling process, the speed of the grinding mill significantly influences the efficiency of comminution. By integrating machine learning models and Discrete Element Method (DEM) simulations into the digital twin, it becomes possible to analyze the complex dynamics of particle interactions during grinding in real time.
By determining the optimal grinding speed, the digital twin assists in maximizing the efficiency of the milling process, leading to improved production rates and energy utilization. This can be achieved by:
• Data Integration: The digital twin incorporates real-time and historical data from the milling process, including particle sizes, forces, and velocities.
• Machine Learning: Utilizing machine learning algorithms, the model learns from this data to predict how changes in grinding speed affect the output particle size distribution.
• DEM Simulations: The digital twin will leverage DEM simulations to simulate the behavior of particles under varying speeds, providing insights into the optimal grinding speed for achieving desired particle sizes.
The digital twin model extends its utility by enabling the determination of the minimum volume fraction of solids required for optimal grinding. The volume fraction of solids in a milling process influences the collision frequency and, consequently, the grinding efficiency. This can be done by:
• Real-time Monitoring: The digital twin continually monitors the volume fraction of solids during the grinding process.
• Machine Learning and Optimization: Machine learning models analyze the relationship between volume fraction and grinding efficiency. The model can then identify the minimum volume fraction that optimizes the comminution process.
By identifying the minimum volume fraction required for optimal grinding, the digital twin aids in resource management, ensuring that only necessary amounts of ore are processed without excess. The digital twin model actively contributes to the reduction of specific energy consumption during the ore grinding process. Specific energy consumption refers to the amount of energy required to grind a unit mass of ore, and minimizing this consumption is crucial for sustainability and cost-effectiveness. This can be done by:
• Energy Monitoring: The digital twin monitors energy consumption patterns during grinding, considering factors such as speed, volume fraction, and other operational parameters.
• Predictive Analytics: Machine learning algorithms analyze historical data to predict how changes in operational parameters impact specific energy consumption.
By optimizing operational parameters based on real-time data and predictive insights, the digital twin helps minimize specific energy consumption, leading to cost savings and environmental sustainability.
Project software
ANSYS Rocky: This will be used for Discrete Element Method (DEM) simulations. DEM is particularly useful for simulating the behavior of granular materials, such as particles, grains, powders, etc., and their interactions. ANSYS Rocky is a software package specifically designed for DEM simulations, providing tools for analyzing particle behavior under various conditions. This software is available on Roar Collab.
Python for ML algorithms: Python is a popular choice for implementing machine learning algorithms due to its rich ecosystem of libraries such as TensorFlow, PyTorch, scikit-learn, etc. One can use Python to preprocess data, train machine learning models, evaluate their performance, and deploy them if needed. It offers flexibility and scalability, making it suitable for various machine learning tasks. Hence, we are using it for this project.
Corheo (Grains and Fluid): Corheo will be used to test the continuum framework we developed from our trained model, and we currently have the trial version of this software installed on the ICDS cluster. If successful, we would purchase the full version and add it to our strategic modelling tools for industrial scale project.
MATLAB Simulink for digital twin model development: MATLAB Simulink is a powerful tool for modeling, simulating, and analyzing dynamic systems. Developing a digital twin model involves creating a virtual representation of a physical system to mimic its behavior. Simulink provides a graphical interface for designing models using block diagrams, making it intuitive for engineers and researchers to create complex systems and simulate their behavior. By combining these tools, we will leverage the strengths of each to address different aspects of this project effectively.
Research Team Composition, Expertise, and Roles: The research group consists of Professors Ogunmodimu, Dzade, and Arnold as investigators. They are interested in studying the mechanical activation of minerals through fine grinding and developing a digital twin model for the milling process. Professor Arnold Barabra is a senior colleague who specializes in minerals processing and will be instrumental in providing experimental data that will be used to validate our computational models before we employ the computational data to train the ML model that will be employed to develop a surrogate model for real-time analysis of the milling system. My (Olumide) role, alongside Nelson Dzade, involves working closely with graduate students on the project to develop and refine computational simulations for the fine-grinding process. Our collective efforts will aim to extract relevant data for the advancement of machine learning algorithms, which includes tasks such as data curation and analyzing the results to derive meaningful insights. Our interdisciplinary expertise is directed towards enhancing the efficiency and effectiveness of fine grinding technology, with potential applications in mineral processing and cement production.
Budget: The requested funding amounts to $52,972. This budget allocation will cover stipends (2 x $16,656) and fringe benefits (2 x $1,829.75) for two graduate assistants in grade 14 over two semesters, along with $16,000 for the Corheo software. Also included in the budget are 80 cores of high-speed memory at $2,000 and 60 hours of ICDS RISE Team at $4,800. The entire budget is designated from July 2025 to June 2026.