Convolutional neural networks to characterise particle suspensions from ultrasonic backscatter

Ultrasonic backscatter has been used extensively across many applications to characterise suspended particles. It is of particular interest in nuclear decommissioning, as it allows online characterisation without the need to sample, or even contact the suspension in some cases. Industrial processes often utilise dynamic changes to suspended particle concentrations and particle size distributions (PSDs), and as such, characterisation of both simultaneously would be advantageous. At present, there is limited scope within existing analytical methods to achieve this, where the concentration or PSD of the target system must be known to calculate the other. Machine learning (ML) is a method that when trained on representative data, can use non-linear multi-variable minimisations to estimate both concentration and PSD simultaneously and, as such, this study aims to demonstrate that an artificial neural network (ANN) and convolutional neural network (CNN) can accomplish this. A training library of nine spherical glass bead suspension systems, comprising of variable median particle size and coefficient of variation, across six concentrations was compiled using a commercial backscatter instrument at 2 and 4 MHz. The hyperparameters of an ANN and CNN were optimised on these acoustic profiles, before being used to predict median particle size, coefficient of variation, and concentration from acoustic profiles at 2 and 4 MHz of two “unknown” suspensions. While neither the ANN or CNN predictions proved to be successful for estimating the coefficient of variation, moderate agreement between predicted and true values were found for median particle size and concentration from the ANN, while the CNN achieved good agreement for median particle size and very good agreement when predicting particle concentration. Consequently, this study was able to successfully determine that a CNN could simultaneously estimate a median particle size and concentration using ultrasonic backscatter data gathered on an “unknown” suspension. • MHz ultrasonic (US) backscatter characterisation of bidisperse suspensions. • ANN and CNN used in ML evaluation of training sets from different size mixtures. • Particle size distribution (PSD), coefficient of variation and concentration analysed. • CNN predicted PSDs and concentrations from “unknown” mixtures accurately. • Additional features are required for better prediction of Coefficient of Variation.