Utilising a deep neural network as a surrogate model to approximate phenomenological models of a comminution circuit for faster simulations

Abstract

Comminution circuits can be modelled using phenomenological models of individual unit operations to represent the operation performance. Process optimisation and block model variability analysis require millions of simulations to fully explore process efficiency through mine scheduling which is costly and time consuming. Surrogate modelling is a technique used in process engineering approach of approximating the behaviour of the underlying model by using a model which is computationally more feasible. Neural networks are a form of machine learning which can approximate complicated function mappings of inputs to outputs and are computationally parallelisable.In this study, a neural network is used as a surrogate model to approximate a copper porphyry mine comminution circuit for faster simulations. The neural network was trained to predict throughput using data generated from the phenomenological models of the comminution circuit. The optimal neural network hyperparameters were determined using an evolutionary algorithm to minimise overfitting. The neural network predicted simulation results 3363 times quicker than phenomenological models with errors of 0.37%, 0.55% and 0.45% across three different test sets. The neural network only required a stratified training sample of 1 in 1000 data to interpolate the rest of the data.