Targeted calibration to adjust stability biases in non-differentiable complex system models

Numerical models of complex systems like the Earth system are expensive to run and involve many uncertain and typically hand-tuned parameters. In the context of anthropogenic climate change, there is particular concern that specific tipping elements, like the Atlantic Meridional Overturning Circulation, might be overly stable in models due to imperfect parameter choices. However, estimates of the critical forcing thresholds are highly uncertain because the parameter spaces can practically not be explored. Here, we introduce a method for efficient, systematic, and objective calibration of process-based models. Our method drives the system toward parameter configurations where it loses or gains stability, and scales much more efficiently than a brute force approach. We successfully apply the method to a simple bistable model and a conceptual but physically plausible model of the global ocean circulation, demonstrating that our method can help find hidden tipping points, and can calibrate complex models under user-defined constraints.