Home

My research

In my PhD, I work on the identification of model structural error in hillslope hydrological models. Specifically, I am using inverse modeling tools to estimate locations in a hillslope where soil water leaks vertically into the deeper bedrock (as opposed to being transported laterally through the soil). For this, I'm using artificial data: i.e. I created the 'observations' of discharge, pressure head and transient groundwater, using a model setup with spatially heterogeneous leakage parameters (Fig.1), thus mimicking the field situation. The artificial observations are shown in Fig. 2.

In the inverse run, I use a model that is structurally inadequate for the data (i.e. the leakage parameter assumes spatially homogeneity). Calibrating that model gives me parameter sets that are optimal, but atthe same time wrong: they are associated with autocorrelation in the model-observation misfits (Fig. 3). This is because the model states are wrong but the parameters compensate optimally for that (and are therefore wrong themselves). This compensation effect can be really annoying if you are trying to test your understanding of how a hillslope function hydrologically, because it masks the origin of errors.

Fortunately, there are other inverse methods available that allow for the simultaneous estimation of parameters and states. I am currently using SODA, which consists of a parameter optimization method (SCEM-UA) on top of an ensemble Kalman filter. When the hillslope model is run within this framework, errors introduced by model structural inadequacy are better accounted for. Perhaps most importantly, the ensemble Kalman filter also yields a time series of how the model states were updated. The patterns present in these updates are crucial in tracking the errors due to model structure; In other words, these updates can be used to identify locations where soil water leaks to the deeper bedrock. This can be seen in Fig. 4 where the overestimation of pressure head (green zones) coincide with the leakage hotspots that I put in the forward model (but not the inverse model).

Fig. 1

Fig. 2

Fig. 3

Fig. 4