Neural networks in multiphase reactors data mining: feature selection, prior knowledge, and model DeSIGN
Artificial neural networks (ANN) gained recently enormous popularity in many engineering fields not only for their appealing so called “learning ability” but also due to their larger applicability and superior performance with respect to classical approaches. Without supposing a particular equational form, they are able to mimic complex nonlinear relationships that may exist between an input feature vector x and a dependent (output) variable y. In the context of multiphase reactors the potential of neural networks is high as the modeling by resolution of first principle equations to forecast sought key hydrodynamics and transfer characteristics, is intractable in the gas-liquid-solid systems. As nothing comes for free, the general-purpose applicability of neural networks in regression and classification is paid with revealing some subsidiary difficulties that can make them inappropriate for use in certain modeling problems. Some of these problems are general to any empirical modeling technique, as the feature selection step, in which one have to decide which subset xs Í x should constitute the inputs (regressors) of the model. Other weaknesses specific to the neural networks are the overfitting, the model design ambiguity (architecture and parameters identification) and the lack of interpretability of resulting models.
This work addressed mainly three issues in the application of neural networks: i) feature selection ii) prior knowledge matching within the models (to answer in some extent the overfiting and interpretability issues) and iii) the model design. Feature selection was conduced with genetic algorithms (yet another companion from artificial intelligence area), which allowed identification of good combinations of dimensionless inputs to use in a regression ANNs, or with sequential methods in the classification context. The type of prior knowledge we wanted the resulting ANN models to match is the monotonicity and/or concavity in regression or class connectivity and different misclassification costs in classification. Even the purpose of the study was rather methodological, some resulting ANN models may be considered contributions per se. These models, direct proofs of concept for the underlying methodologies, are useful to predict liquid hold-up and pressure drop in counter current packed beds, and flow regime type in trickle beds.