Can a simple lumped parameter model simulate complex transit time distributions? Benchmarking experiments in a virtual watershed.

The catchment transit-time distribution (TTD) is the time-varying, probabilistic distribution of water travel times through a watershed. The TTD is increasingly recognized as a useful descriptor of a catchment's flow and transport processes. However, TTDs are temporally complex and cannot be observed directly at watershed scale. Estimates of TTDs depend on available environmental tracers (such as stable water isotopes) and an assumed model whose parameters can be inverted from tracer data. All tracers have limitations though, such as (typically) short periods of observation or non-conservative behavior. As a result, models that faithfully simulate tracer observations may nonetheless yield TTD estimates with significant errors at certain times and water ages, conditioned on the tracer data available and the model structure. Recent advances have shown that time-varying catchment TTDs can be parsimoniously modeled by the lumped parameter rank StorAge Selection (rSAS) model, in which an rSAS function relates the distribution of water ages in outflows to the composition of age-ranked water in storage. Like other TTD models, rSAS is calibrated and evaluated against environmental tracer data, and the relative influence of tracer-dependent and model-dependent error on its TTD estimates is poorly understood. The purpose of this study is to benchmark the ability of different rSAS formulations to simulate TTDs in a complex, synthetic watershed where the lumped model can be calibrated and directly compared to a virtually "true" TTD. This experimental design allows for isolation of model-dependent error from tracer-dependent error. The integrated hydrologic model ParFlow with SLIM-FAST particle tracking code is used to simulate the watershed and its true TTD. To add field intelligence, the ParFlow model is populated with over forty years of hydrometric and physiographic data from the WE-38 subwatershed of the USDA's Mahantango Creek experimental catchment in PA, USA. The results are intended to give practical insight into tradeoffs between rSAS model structure and skill, and define a new performance benchmark to which other transit time models can be compared.