Generalization performances of two different neural networks are compared with a conventional software for prediction of hollow fiber permeances and the corresponding separation factors. Two experimental data sets were used as the training data for separation of carbon dioxide from methane. Both Radial Basis Function (RBF) and Multi-layer Perceptrone (MLP) networks provide superior performances compared to the conventional software (Table Curve or TC). It is also shown that the RBF networks provide better predictions than MLP networks because of their powerful noise filtering capabilities. For RBF networks, both appropriate choices of isotropic spread and the corresponding optimal level of regularization are crucial for proper reconstruction of the true underlying hyper-surface from a set of noisy data set. The in-house algorithms are used for training both MLP and RBF networks. It is clearly illustrated that the computation of optimal isotropic spread is crucial for proper performance of the RBF network. The use of Leave One Out Cross-validation (LOOCV) criterion was also essential for appropriate estimation of optimum level of regularization.

Keywords