Despite the growing use of Aquatic Ecosystem Models (AEMs) for lake modelling, there is currently no widely applicable framework for their configuration, calibration, and evaluation. To date, calibration is generally based on direct data comparison of observed vs. modelled state variables using standard statistical techniques, however, this approach may not give a complete picture of the model’s ability to capture system-scale behaviour that is not prevalent in the state observations, but which may be important for resource management. The aim of this study is to compare the performance of ‘naïve’ calibration and a ‘system-inspired’ calibration, a new approach that augments the standard state-based calibration with a range of system-inspired metrics (e.g. thermocline depth, metalimnetic oxygen minima), in an effort to increase the coherence between the simulated and natural ecosystems. This was achieved by applying a coupled physical-biogeochemical model to a focal site to simulate temperature and dissolved oxygen. The model was calibrated according to the new system-inspired modelling convention, using formal calibration techniques. There was a clear improvement in the simulation using parameters optimised on the additional metrics, which helped to focus calibration on aspects of the system relevant to reservoir management, such as the metalimnetic oxygen minima. Extending the use of system-inspired metrics for the calibration of models of nutrient cycling, algal blooms, and greenhouse gas emissions has the potential to greatly improve the prediction of complex ecosystem dynamics.

Groundwater management involves a complex decision-making process, often with the need to balance the trade-off between meeting society’s demand for water and environmental protection. Therefore effective management of groundwater resources often involves some form of multi-objective optimization (MOO). Many existing software tools offer simulation model-enabled optimization, including evolutionary algorithms, for solving MOO problems. However, such analyses involve a huge amount of numerical process-based model runs, which require significant computational effort, depending on the nonlinearity and dimensionality of the problem, in order to seek the optimal trade-off function known as the Pareto front. Surrogate modeling, through techniques such as Gaussian Process Regression (GPR), is an emerging approach to significantly reduce the number of these model evaluations thereby speeding up the optimization process. Yet, surrogate model predictive uncertainty remains a profound challenge for MOO, as the current Pareto dominance criteria presumes that model responses are deterministic. Such presumption could mislead surrogate-assisted optimization, which may result in either little computational savings from excessive retraining, or lead to suboptimal and/or infeasible solutions. In this work, we present probabilistic Pareto dominance criteria that considers the uncertainty of GPR emulation during MOO, producing a ”cloudy’” Pareto front which provides an efficient decision space sampling mechanism for retraining the GPR. We then developed a novel acquisition strategy to manage the solution repository from this cloud and generate an ensemble of infill points for retraining. We demonstrate the capabilities of the algorithm through benchmark test functions and a typical density-dependent coastal groundwater management problem.