High resolution groundwater flow modeling, necessary to evaluate effects on a local scale, has traditionally been restricted to small regions given the computational limitations of the CPU memory to handle large numerical MODFLOW-grids. Although CPU-memory size doubles every two years (‘Moore’s law’) the restriction still holds from a hardware point of view. This restriction has traditionally forced a model builder to always choose between (1) building a model for a large area with a coarse grid resolution or (2) building a model for a small area with a fine grid resolution. For some time it appeared that finite element models could fill the gap by refining the grid only where hydrological gradients were anticipated. However, unanticipated stress may also occur in parts of the model area where the grid is not yet refined resulting in a possible undesired underestimation of these effects. Theoretically the modeler could choose to design a finite element network with a high resolution everywhere, but then it becomes more economic to use finite differences. This is why Deltares has based its innovative modeling techniques on MODFLOW considering it is largely seen world-wide as the standard finite difference source code. Still, modelers ideally need an approach that allows: (1) flexibility to generate high resolution model grids everywhere when needed, (2) flexibility to use or start with a coarser model grid, (3) reasonable runtimes / high performance computing and (4) conceptual consistency over time for any part of the area within their administrative boundary. Deltares has invested in understanding all of these requirements and has developed the iMOD software package to advance the methods and approach used by modelers and regulators.

The development of the iMOD approach took off in The Netherlands in 2005 when Deltares and a group of 17 stakeholders decided to jointly build a numerical groundwater model for their common area of interest (Berendrecht *et al.*, 2007, Vermeulen, 2013). The groundwater model encompasses the entire north of the Netherlands at a resolution of 25 x 25 m\({}^{2 }\)and was constructed together via an internet accessible user-interface. This makes it possible for the modelers to easily access the model data, intermediate results and participate in the model construction. The iMOD approach
allows gathering the available input data to be stored at its finest available resolution; these data don’t have to be clipped to any pre-defined area of interest or pre-processed to any model grid resolution.

*The iMOD approach: one input data set:*

Resolutions of parameters can differ and the distribution of the resolution of one parameter can also be heterogeneous. In addition, the spatial extents of the input parameters don’t have to be the same. iMOD will perform up- and down scaling (Vermeulen, 2006) whenever the resolution of the simulation is lower or higher than that of the available data. This approach allows the modeler to interactively generate models of any sub-domain within the area covered by the data set. When priorities change in time (e.g. due to changing political agenda’s) the modeler can simply move to that new area of interest and apply any desired grid resolution. In addition the modeler can edit the existing data set and / or add new data types to the data set. Utilizing the internal up- and down-scaling techniques ensures that sub-domain models remain consistent with the bigger regional model or that the regional model can locally be updated with the details added in the sub-domain model.

Suppose the modeler needs to simulate groundwater flow for the total area covered by the data set, but the theoretical size of the model is far too big to fit in any CPU-memory. iMOD facilitates generating sub models for parts of the whole area of interest with a user-defined resolution depending on how large the available CPU-memory is and how long the modeler permits her/himself to wait for the model calculations to last. To generate a high resolution result for the whole model domain a number of partly overlapping but adjacent sub models are invoked and the result of the non-overlapping parts of the models are assembled to generate the whole picture. The modeler should of course be cautious that the overlap is large enough to avoid edge effects, but this overlap is easily adjustable in iMOD. A big advantage of this approach is that running a number of small models instead of running one large model (if it would fit in memory, which it often will not) takes much less computation time; computation time (T) depends on the number of model cells (n) exponentially: T = f(n\({}^{1,5-2,0}\)). The approach also allows the utilization of parallel computing, but this is not obligatory. Using this approach means that the modeling workflow is very flexible and not limited anymore by hardware when utilizing iMOD.