Characterising the Swiss window stock and retrofit potential using machine-learning

Abstract Windows strongly influence both operational energy demand and embodied emissions in the construction sector, yet they remain insufficiently represented in national building material stock models. Accurate representation of window materials, energy performance, and renovation state is essential for assessing both operational and embodied climate impacts, and for informing future refurbishment and re-use strategies. This study combines detailed information from energy performance certificates with machine-learning classification. Window frame materials, glazing types, renovation states, and window to wall ratios are derived from energy certificate data for more than 40,000 buildings. Results show an increasing share of plastic frames since the 1990s and increasing window to wall ratio until 2010. We then train supervised machine-learning classifiers on these 40,000 buildings to predict glazing type and frame material for all 1.51 million Swiss residential buildings from building age, size, and location. While double glazing can be identified with a precision of 0.8 and triple glazing with an F1-score of 0.67, frame-material prediction performs lower due to overlapping characteristics between material classes. We preserve stochastic uncertainty when applying the trained classifiers using probabilistic classification and Monte Carlo analysis. The resulting estimates describe the window stock at an aggregate, probabilistic level. We use this stock estimate to assess the retrofit potential of the Swiss window stock to showcase the application of our framework. Double glazing dominates the residential building stock in all cantons (56.2–67.7%). By contrast, potential annual net $$\hbox {CO}_2$$ CO 2 savings per $$\hbox {m}^{2}$$ m 2 window area for a double-to-triple glazing retrofit show three-fold differences across cantons.