Large-Scale High-Resolution Groundwater Modelling using Grid Computing

Developments in numerical groundwater modelling have shown that models become more and more ambitious with increasing computer capacity. To meet the need for accurate instruments to support decision-making, both scale and resolution of models have grown enormously during last years.

Within the recent project 'Development of a Methodology for Interactive Planning for Water Management' (MIPWA) the challenge was taken to develop a high-resolution numerical groundwater model for the whole north of the Netherlands. This MODFLOW model encompasses an area of more than 24.000 km(2), has seven quasi-3D model layers and a resolution of 25 x 25 m(2). With this enormous groundwater model - which is unique in its size - 13 years of daily groundwater fluctuations had to be simulated.

Both running and calibrating such a large model require innovations in model building, model processing and data handling. Data-compression techniques were required to store all input and output data. To run the model, both upscaling and model-decomposition techniques were develoved. For a transient run (over 4500 time steps) on the highest resolution, the model was decomposed into 473 overlapping submodels. Transient boundary conditions of the submodels were taken from a lower-resolution model. Each submodel could be run individually, so the process was perfectly suited for parallel processing. Therefore, we developed a computational grid using the 200 computers available in our office. The moment employees logged off, their computer came available for the grid. Obviously the weekends appeared to be the most productive days!

The grid was also crucial for model calibration. We used the Representer method for calibrating model parameters in a stationary mode. The Representer method requires a forward run and an adjoint run each iteration to calculate the so-called representer of each observation. In total more than 8000 groundwater observation locations were available and hence more than 8000 runs had to be carried out. Each representer run was distributed over the grid using PVM (Parallel Virtual Machine).

Grid computing revealed itself as the only way to complete the whole project within reasonable time. Total CPU time of model calibration and running (ca. 50 runs during model-construction process) was estimated at more than 20 years. Using grid computing, the calculation time was reduced to several months.

In addition to model calibration and model running, grid computing is also helpful in data-assimilation applications. In a preliminary study, Ensemble Kalman Filtering techniques were applied for nowcasting and forecasting of groundwater fluctuations using assimilated groundwater. Model states were estimated by calculating 200 ensembles distributed over the grid. Subsequently, 10-day forecasts of groundwater levels were calculated by processing 50 ensembles of the Ensemble Prediction System (EPS) calculated by the European Centre for Medium-Range Weather Forecasts. As the intention is to produce forecasts on daily basis, a computational grid is necessary to run all ensembles within one day.