Eleven state-of-the-science regional air quality (AQ) models, exercised by 20 independent groups in Europe and North America, have been assembled for the Air Quality Model Evaluation International Initiative (AQMEII). The modelled ground-level ozone mixing ratios are collectively examined from the ensemble perspective and evaluated against observations from both continents. We aim at creating optimized ensembles in order to capture the data variability while keeping the error low. It is shown that the most commonly used ensemble approach, namely the average over all available members, can be outperformed by subsets of members optimally selected in terms of bias, error, and correlation, independent of the skill of the individual members. A clustering methodology is applied to discriminate among members and to build a skilful ensemble based on model association and data clustering.