The computational cost of realistic air traffic simulations is a barrier for a comprehensive assessment of new ATM concepts and solutions, which, in practice, restricts the simulations to a limited number of scenarios, often insufficient to obtain conclusive results. So, a goal for a comprehensive exploration of the simulation space should be finding its most informative instances. This can be done by means of active learning metamodelling, which can be used to translate a complex simulation model into a metamodel, allowing a more efficient exploration of the simulation input-output space. This work presents two metamodels developed within the SESAR ER4 SIMBAD project for one of the state-of-the-art ATM simulation tools, R-NEST. The metamodels were trained using the active learning technique through the metamodelling framework developed by the SESAR ER4 NOSTROMO project. The training process with this tool is also described in the paper.