Data-driven and physics-based modeling approaches and their integration in building digital twins: A systematic review

Abstract

Interest in digital twin technology has grown significantly within the building sector as part of the broader digital transformation in the architecture, engineering, and construction industry. A building digital twin is a virtual replica that captures a building’s static and dynamic behavior through data, information, and models. Digital twin models can be developed using data-driven or physics-based approaches, each with distinct advantages and limitations. Data-driven models can learn complex behaviors from data and scale well, but they require large datasets and often lack interpretability. In contrast, physics-based models offer interpretability and generalizability through fundamental principles but can be computationally demanding. Consequently, building digital twins can benefit greatly from integrating both approaches through hybrid modeling. However, the literature lacks a comprehensive analysis of integration strategies within building digital twins. This study addresses that gap by reviewing advances in data-driven and physics-based modeling and analyzing various integration levels. The results show that most studies rely on siloed models, using either approach independently without leveraging their complementary strengths. Some adopted sequential integration, where one model informs the other but lacks real-time or iterative feedback. A few achieved coupled integration, involving active data exchange and collaboration between models. Only three studies explored fusion integration, where both approaches are fully unified into a single model. Based on this review, a method is proposed for selecting the appropriate level of integration, considering factors such as data availability, interpretability, generalizability, and domain knowledge. Finally, key research gaps and future directions are identified to guide further work.