Supplementary material to "Data-driven modeling of hydraulic head time series: results and lessons learned from the 2022 groundwater modeling challenge"

Team DA CollectiveThe "DA Collective" team used an ensemble-of-transfer-function-models approach.The Pastas python module (Collenteur et al., 2019) was used as the underlying transfer-function model engine and was wrapped within an ensemble-based data assimilation framework.The ensemble-based data assimilation captured transfer-function input parameter uncertainty as well as temporally-varying forcing uncertainty.The forcing uncertainty was captured in Julian day multiplier parameters that were then transferred to the forecast period so that any biases learned during the training period could be propagated to the forecast period.We also explicitly considered temporally correlated noise in the observed groundwater levels used for training.The iterative ensemble form of the Gauss-Levenburg-Marquardt solution was used to assimilate the observed groundwater levels.An ensemble of 100 realizations was used across 5 iterations for data assimilation; we note that an adaptive rejection filter was used each iteration to remove realizations with statistical outlier phi values.Observed groundwater levels from 1 Jan 2008 to 31 Dec 2012 were held back for verification at each site.More details related to the approach can be found in the workflow python script.