Reinforcing the phenomenological consistency in artificial neural network modeling of multiphase reactors

Artificial neural networks (ANN) aided with dimensional analysis have been successfully applied in multiphase reactors modeling when considerable amount of experimental data (or database) is available. An important problem that stemmed from this approach was the ambiguity to select the fittest combination of dimensionless numbers to be used as ANN inputs to predict a variable of interest. A genetic algorithm (GA) based methodology was proposed to optimize the combination of inputs by taking into account the phenomenological consistency (PC) of the resulting ANN models along with their fitting capabilities. PC is a measure of the capability of an ANN model to simulate outputs with specified gradient conditions with respect to the process variables. These conditions are imposed based on a priori knowledge of the systemʹs behavior. PC used to be evaluated in the vicinity of a particular point in the database space. The novelty of the approach was the extension of the PC test around all the points available in the training data set. This technique may be regarded as a robust method to prevent data overfitting when the function to be learned by ANN is characterized by a monotonic behavior with respect to some of the process variables. The new approach was illustrated using as a case study the correlation of two-phase pressure drop in randomly packed beds with countercurrent flow.