From computer vision to minerals processing: Using a convolutional neural network for parameter estimation of, first-order Froth Flotation Models E.J.Y., Koh, E. Amini, and G.J. McLachlan

Inferring individual component flotation rates from recovery-time data of ore complexes is important in optimising and designing minerals processing plants. However, existing two-component first-order flotation models are computationally intensive to solve using conventional non-linear least squares regression. In this paper, a Convolutional Neural Network (CNN) is used to approximate the inverse function which predicts the two-component first-order model parameters using recovery-time data. This paper compares the accuracy and computation speed of the proposed CNN against traditional non-linear least squares regression on 1 million generated data based on a known two-component firstorder kinetic model. It is shown that using a CNN is possible to approximate the inverse function for the non-invertible two-component model with recovery-time as an input matrix for convolution operations. The CNN method is 450 times faster to compute than the Trust-Region (TR) non-linear least squares regression algorithm. However, the CNN model is less accurate compared to the TR algorithm overall. Keywords: Flotation kinetics, convolutional neural network, inverse problem, modelling and simulations