Output details

A mixed-integer programming approach for optimal configuration of artificial neural networks

Artificial Neural Network (ANN) models are ubiquitous in all the domains of chemical engineering with more than 200 reported applications including in the areas of distillation, reaction, crystallization and vapour-liquid equilibrium. ANN can model highly complex and nonlinear systems. However one key bottleneck has been the ad-hoc and hit-and-trial approach for obtaining the appropriate number of nodes and layers in the network so as to avoid under-training and over-training. This paper for the first time presents an optimisation model to overcome this limitation, automates the procedure and reduces user intervention; we used it for refinery-wide optimisation in a separate paper.