Meta modeling: using machine learning to assess the model uncertainty of a high-resolution groundwater model

Our study investigates the dynamics of the Gross Glienicker Lake, a groundwater fed lake in the Berlin-Brandenburg region of Germany. This lake (similarly to many others in the region) is experiencing a significant water decline mainly driven by the climate, loosing more than 2 meters of its water levels since the 1970s. To understand the hydrogeological system better, and to identify potential mitigation measures we applied a coupled groundwater-surface water model using HydroGeoSphere (HGS). This 3-D model simulates the hydrological processes of the catchment with high spatial and temporal resolution, incorporating all available geological and hydrological data from the area.

The model was mainly created to evaluate the impacts of different future climate projections on the water levels. We have investigated 3 different RCP scenarios using 43 different climate projection simulations. We have employed machine learning tools to fill in any future data gaps, for example future levels of a river boundary condition and future groundwater extraction rates given population growth trends. To access the uncertainties originating from the HGS model, we have used a meta modeling framework. Meta modeling uses a machine learning based surrogate model (an LSTM in this case), to emulate the input-output numerical relationship of the HGS model in a computationally efficient way. Once trained, the meta model can emulate an HGS model run accurately in a couple of seconds. We fed the meta model with thousands of perturbed climate inputs, showing that the model output is robust even under extreme climatic conditions.

Our results showed that the lake is highly sensitive to precipitation variability, therefore future projections diverge significantly given the scenarios. Except for the wet scenario, all predictions show further water level decrease and they also reveal a strong shift in the seasonal dynamics.