Revising Common Approaches for Calibration: Insights From a 1‐D Tracer‐Aided Hydrological Model With High‐Dimensional Parameters and Objectives

Abstract The dimensionality of parameters and objectives has been increasing due to the accelerating development of models and monitoring network, which brings potential challenges for calibration. In this study, two common philosophies for multi‐objective optimisation in hydrology (the use of aggregated scalar criterion or vector functions) were revisited with different sampling strategies: (a) random sampling, (b) DiffeRential Evolution Adaptive Metropolis (DREAM as an example of an aggregated scalar function), and (c) Non‐Dominated Sorting Genetic Algorithm II (NSGA‐II as Pareto‐based multi‐objective optimisation). By testing the ability of algorithms to simultaneously capture soil moisture and soil water isotopes at three depths under four vegetation covers, we found random sampling performed poorly in matching observations due to its inability to explore high‐dimensional parameter space. DREAM, in contrast, could provide efficient parameter convergence with informal likelihood functions, but the choice of formal likelihood function is difficult due to the lack of knowledge about model residuals, leading to poor performance. NSGA‐II is effective and efficient after aggregating objectives to ≤4, but failed when calibrating against all 24 objectives. Overall, both philosophies and all three approaches are challenged by increasing dimensionality, and it generally requires a degree of trial‐and‐error before achieving a successful calibration. This suggests the potential to explore a more flexible way to describe model residuals (e.g., by defining limits of acceptability). Alternatively, improvements could be made by using an ensemble of models to represent the system (instead of “best” model) given the average of a calibrated ensemble usually performed better than any individual model.