There is a great deal of instances in multiphase reactor engineering where the system's state is labeled using categorical indices to depict class memberships. Remarkable examples encompass the labeling of flow patterns and regime transitions in multiphase porous media flows, the macro and microscale fluidization states in fluid-solid or fluid-fluid-solid reactors, the bubble wake morphologies, the bed initial contraction/expansion in fluidization, etc. The design of general-purpose flow pattern recognizers for multiphase reactors requires implementation of feature selection algorithms and classifier design methods. Feature selection algorithms enable extracting the most informative feature sets to be included in the data-driven inference engine (or classifier). In this work, supervised classifiers were built by modeling the class-conditional probability functions of class memberships. The discriminant functions yielding the best class separation were searched using the following statistical and neural-network-based classifiers: Gaussian (quadratic discrimination rule), linear (normal based), nearest mean class, nearest neighbor, k-nearest neighbor, binary decision tree, radial basis functions and multilayer perceptron neural networks. These approaches were benchmarked using, as an example, the three-class trickle-bed flow-regimes problem (Low interaction regime [LIR], high interaction regime [HIR], transition regime [TR]) for which a comprehensive knowledge-referenced database is available (5061 flow regime observation records). For a given model complexity, the multilayer perceptron (MLP) neural network outperformed the other approaches in terms of accuracy. Enhanced phenomenological consistency and immunity against overfitting of the MLP model were achieved by embedding specific domain prior knowledge. The model captured with success the hierarchical linking nature of the categories during monotonic increase (respectively, decrease) of fluid throughputs (respectively, gas density) from LIR to HIR through TR. Misclassification between the non-adjacent LIR and HIR classes was kept to a minimum and the model yielded cross-validated error of 16% for the 5061 database records. (C) 2004 Elsevier Ltd. All rights reserved.