A chemical reactor selection expert system created by training an artificial neural network

Abstract

This work investigated the feasibility of using a feed-forward neural network for knowledge acquisition and storage, and subsequent use as a chemical reactor selection expert system. Feed-forward neural networks have the capability of learning heuristics from given examples.

Levenberg-Marquardt method was used to train the network by minimising the sum of squares of residuals. The output of each node was calculated by the logistic activation (sigmoid) function on the weighted sum of inputs to that node. It is shown therein that the number of hidden layers and the number of nodes in the hidden layers are critical, and increase in the number of hidden layers does not always improve the performance of the simulator network. It is possible in certain cases like this one to attribute meanings to the nodes in the hidden layer.

Redundancy in the outputs was considered by having separate output nodes for selecting batch and continuous operations, and for stirred-tank and tubular reactors. The network performance did not significantly change on excluding one of the outputs, although it was not possible to arrive at the converged solution equally easily when four outputs were considered.

This work demonstrated that a selection expert system can be created in a feed-forward neural network. In other words, neural networks can be used for knowledge acquisition and storage for selection expert systems, suitable for convenient retrieval and inferencing. Inspite of covering a wide range (several orders of magnitude) of inputs, the performance was found to be very good.