Robust and expert-agnostic digital twin calibration via ensemble learning and Bayesian optimization

Abstract

Digital twins have emerged as a critical tool in tackling climate change. Considering the data scarcity of complex systems, a promising approach to developing digital twins involves combining physics-based models with data assimilation. However, model calibration remains challenging due to uncertainties in both the physical models and observational data, and the reliance on domain knowledge. In this study, we develop an ensemble learning-based approach that aggregates sub-models with diversified calibration configurations. The proposed method streamlines calibration without expert-driven parameter screening and improves the digital twin's extrapolation capability, enabling more robust predictive applications. We demonstrate the effectiveness of our approach by calibrating the energy model of an office building, significantly reducing the extrapolation error and the associated risks. To the best of our knowledge, this is the first study to facilitate the calibration of physics-based models using ensemble learning, especially in the parameter space.