Machine-learned kinetic Façade: Construction and artificial intelligence enabled predictive control for visual comfort

Abstract

The authors present the first on-site investigation of artificial intelligence (AI)-integrated three-dimensionally movable kinetic façade (KF). Despite continued architectural interest on the KF to improve indoor visual comfort, its in-situ operational strategy has been little addressed. To examine our primary hypothesis that the adaptive KF controlled by AI models improves indoor daylight probability (DGP) in real time, we developed an electromagnetic hexagonal KF mechanism, and three machine-learning (ML) regressors (eXtreme Gradient Boosting (XGB), Random Forest (RFR), Decision Tree) were implemented on a Raspberry Pi board to control the KF (width = 1.73 m, height = 1.1 m). 20,000 data from Radiance were used for model construction, and illuminance sensors were installed for on-site validation in a private office mockup. The façade shape was optimally morphed every 90s, using differential evolution. In the verification, XGB showed the greatest accuracy (R2 = 91.2%) with decent prediction time efficiency (μ = 0.58 s), but the RFR accuracy (R2 = 79.8%) slightly outperformed XGB in the field.