Smart Surrogate Models for Design and Operation of Efficient and Sustainable Comminution Units

点击翻译按钮获取中文摘要

In comminution, computer simulations are commonly used to test unit designs to ensure reliable operation under a variety of operating conditions. Due to complex process dynamics inside the grinding mills, which are further complicated by unavoidable variations in ore feed properties, it is a challenging task to develop a model that can capture unique characteristics of a particular unit. The spatial and temporal resolution requirements for multiphase flows typical for grinding mills make numerical simulations computationally expensive, which are often beyond the reach of many in the mining industry. A paradigm shift in simulation technology is needed to make multi-physics simulations of grinding mills more practical for design, optimization, and control purposes. This can be achieved through the development and application of hybrid physics-based and data-driven surrogate models that can strike a balance between model accuracy or complexity on the one hand, and practical operational considerations on the other. Advances in data analytics and machine learning have enabled the possibility of construction of data-fitted surrogate models that can duplicate the behavior of the numerical results used for their construction. Smart surrogate modeling takes advantage of pattern recognition capabilities of machine learning to build powerful tools to predict the behavior of a system with a far less computational cost. In this research program, an interdisciplinary approach combining machine learning, physics-based simulations, and real-time data analytics will be used to develop a robust surrogate model of a grinding mill. The main goal of this research program is to develop and validate a complete digital replica of a mill, which will be executed in real-time and provide unprecedented visibility into the process under variable operating conditions. A data-driven smart surrogate model will be developed to mimic results from numerical simulations with high accuracy and faster execution time. This unique engineering-based data analytics approach will utilize artificial neural networks in conjunction with supervised fuzzy cluster analysis to identify the most influential parameters for the training process and identify the optimum partitioning of the data for training, calibration, and validation. The proposed research will significantly enhance our capability to find optimal parameter vectors, both design and operational, and asses the global behavior of grinding mills over the entire design space. This will lead to improved understanding of dependencies between input and output parameters, which will in turn help optimize grinding operations, decrease energy and water consumption, and therefore produce significant economic savings. Finally, this program will contribute to the education and training of HQP in the fields of mining and minerals processing, statistical data analysis and visualization, machine learning, and advanced process modeling and control.