The present study records an attempt to train artificial neural networks (ANNs) to develop an objective flow pattern indicator for air-water flows on the basis of the vast amount of data available in the literature. The technique should be capable of predicting the flow patterns as well as the intermediate bands of transition from known input variables such as the superficial velocity of the two phases, the pipe diameter, and its inclination. For this, the study tried three different types of ANN. The most commonly used feed-forward back-propagation (FFBP) technique accurately yields the flow patterns but fails when the transition regions are incorporated. The radial basis function network gives better predictions than those by FFBP but fails under certain flow conditions. The probabilistic neural network (PNN), based on Bayes-Parzen classification theory, gives accurate predictions of flow patterns for different pipe diameters and inclinations. It has been validated with both experimental and theoretical models available in the literature. It has next been used to generate generalized flow pattern maps for different pipe diameters and orientations. These maps reveal certain interesting features regarding the existence of flow patterns under different input conditions. PNN predicts the disappearance of the smooth stratified flow with slight inclinations from the horizontal orientation and the existence of churn flow only near the vertical orientation. (c) 2006 American Institute of Chemical Engineers.