By far the biggest cost in air filtration arises from the energy consumption during operation as a result of the pressure drop across the filter. Hence, there is a need for optimizing filter media in order to achieve the highest possible separation efficiency at the lowest possible pressure loss through the filter, while also maximizing filter lifetime due to loading-capacity. This work aims to develop an efficient CFD-DEM coupled model for air filtration applications based on calibrated DEM model parameters, suitable for industrial virtual filter media development. Two approaches are used for this purpose, differing only in the way aerodynamic fluid/particle force interactions are captured. The first, so-called fully-resolved simulation approach uses the immersed boundary method. While this approach is computationally expensive and not suitable for virtual media development in industry, it can be used for highly-accurate micro-scale simulations encompassing simulation domains containing a single or a few fibers. On the other hand, the unresolved Euler-Lagrange approach, in which the flow fields around single free or deposited particles are no longer resolved, is computationally efficient (thereby being suitable for the analysis of fibrous filter media at relevant scales) but it requires additional sub-models such as drag models in order to properly capture aerodynamic fluid/particle interactions. In this context, the present study systematically investigates various drag models, and the influence of other parameters (such as smoothing and mesh resolution) on filtration predictions on the micro-scale. This also includes benchmark comparison with corresponding full-resolved simulations. The most suitable unresolved modeling approach will be subsequently employed to simulate particle filtration for a representative segment of a fibrous filter media. These results will be compared to experimental data thereby providing further validation of the selected modeling approach. Preliminary results for a micro-scale set-up, consisting of a single fiber collector, show that unresolved simulations are approximately 250 times faster than the fully-resolved immersed boundary approach, while maintaining up to 97% accuracy in terms of pressure drop and good qualitative agreement of particle dynamics prior and after their initial deposition.