Observational data are inherently subject to considerable uncertainties and errors. Whether measuring groundwater levels, solute concentrations, or other hydrogeological variables, the accuracy and precision of recorded values cannot be entirely assured. These measurement uncertainties propagate through groundwater flow and transport models, and if inadequately accounted for during calibration, their important influence is consequently absent from model forecasts.

Theoretically, parametric uncertainties can be derived from the underlying governing equations using Gaussian error propagation principles. However, except for straightforward applications with low structural complexity, this approach is often computationally intractable. In practice, ensemble-based methods (requiring numerous additional runs of a complex model) or linear post-calibration uncertainty analyses are more commonly employed. While ensemble approaches can be computationally very tedious, linear methods are constrained by the assumption of strictly linear model behaviour, potentially yielding physically unrealistic parameter ranges.

This contribution will provide a short overview on selected classic and novel uncertainty analysis methods discussing their respective advantages and disadvantages. Furthermore, it will outline the methodological concept and planned implementation of the recently started DFG-funded project PathTrAce-GwT (2025 – 2028), which will address the existing methodological gaps through hybrid machine learning approaches.