Virtual flow metering (VFM) is an attractive and cost-effective solution to meet the rising multiphase flow monitoring demands in the petroleum industry. It can also augment and backup physical multiphase flow metering. In this study, a heterogeneous ensemble of neural networks and regression trees is proposed to develop a VFM model utilizing bootstrapping and parameter perturbation to generate diversity among learners. The ensemble is pruned using simulated annealing optimization to further ensure accuracy and reduce ensemble complexity. The proposed VFM model is validated using five years well-Test data from eight production wells. Results show improved performance over homogeneous ensemble techniques. Average errors achieved are 1.5, 6.5, and 4.7 for gas, oil, and, water flow rate estimations. The developed VFM provides accurate flow rate estimations across a wide range of gas volume fractions and water cuts and is anticipated to be a step forward towards the vision of completely integrated operations. © 2017 IEEE.