Atmospheric pollutants pose significant health risks in Europe, where many cities exceed WHO and EU air quality standards. To support informed air quality policies, simulations of air pollution levels are required. They are generally provided by Chemical Transport Models, but these can be computationally expensive at high spatial resolution. In this context, deep learning methods have emerged as alternatives to full simulations. Using low and high resolution EMEP simulations, we investigate how different training strategies affect the robustness of a Residual Channel Attention Network for downscaling hourly particulate matter, nitrogen dioxide and ozone ground concentration fields from a 0.1°×0.05° to a 0.02°×0.01° resolution over Europe. The geographic extent and temporal sampling of the training data appear to be key determinants of the performance of the downscaling process: training the network on days that are evenly distributed throughout the year generally yields a better performance than training it on the same number of days but restricted to the winter season, and, similarly, the choice of a larger spatial domain during training increases the model performance in comparison with a smaller spatial domain. The optimised model achieves better performance in reconstructing fine scale spatial gradients than a bilinear interpolation of low resolution simulations. It generalises to multiple emission reduction scenarios when the link between the perturbed emissions and the target pollutant is strong and direct, while weaker and more indirect responses are less reliably reconstructed. An optimal compromise between temporal resolution and downscaling accuracy is also identified, corresponding to approximately one day of temporal averaging. These results highlight the reliability of neural network-based downscaling techniques, to support their integration into operational air quality assessment and scenario planning tools as complements to full simulations.