(671e) Rapid Identification of Chemical Reactor Models Using Artificial Neural Networks Classifiers

Artificial neural networks (ANN) coupled with a differential evolution algorithm (DEA) have been used to design experiments and rapidly recognise the resulting reaction kinetics models. In this work, we apply this ANN-DEA framework to the Taylor-Vortex reactor, a type of reactor that can exhibit the hydrodynamic behaviour of a plug flow reactor (PFR), continuously stirred tank reactor (CSTR) or CSTRs-in-series.

Large parameter sets were sampled and used in the reactor models at fixed experimental conditions to generate in-silico data required to train the ANN classifier. Note that a purely physics-based modelling approach is more computationally burdensome. In this hybrid approach, the in-silico data comprised model predicted reactor outlet concentrations corrupted with noise generated following a Gaussian distribution with zero mean and constant relative variance. Binary classification between PFR and CSTR and multi-class classification, which expands to CSTRs-in-series and captures intermediate hydrodynamics, were considered. An ANN composed of two hidden layers of 32 nodes each and a sigmoid function output layer was employed for binary classification with the ANN performance monitored using the classification accuracy as metric.

Results showed the ANN achieved 100% in correctly classifying new data without noise for binary classification of PFR and CSTR and multiclass classification, investigating quinary classification involving PFR, CSTR, 2, 5 and 10 CSTRs-in-series. Although the accuracy decreases with noise, it increases with input features (i.e., number of prediction sets per data sample), achieving 100% accuracy for binary classification with 4 prediction sets per data sample at 10% noise level. The ANN classification performance also improved at new experimental conditions designed by the DEA. The developed ANN-DEA can be used to optimally design new experiments, and rapidly classify the resulting experimental data into the reactor model classes.