Predictive collective control of large-scale kinetic facades via graph neural networks and reinforcement learning: From simulation to physical validation

Abstract Real-time coordination of large-scale kinetic facades is fundamentally challenged by the need to model strongly coupled module dynamics. This study presents the first on-site validation of a predictive collective control framework that integrates a graph neural network (GNN)–based surrogate model with reinforcement learning (RL). Using a GNN surrogate trained on 9229 high-fidelity simulation samples, a proximal policy optimization agent infers opening angles for adaptive solar control across thousands of interconnected units. Summer-solstice stadium simulations show that, compared with an independently controlled multilayer perceptron–RL baseline, the proposed GNN–RL framework reduces spectator heat load by 11.05% at peak solar conditions (14:00) while increasing field solar exposure by 20.63% at low-load periods (16:00), reflecting a time-dependent Pareto trade-off between spectator comfort and turf daylighting. Crucially, the learned policy is deployed directly to physical hardware—without policy transfer—so a 1:30 mock-up reproduces emergent behaviors from global graph interactions. On-site experiments achieve a 10.3% reduction in spectator solar heat gain and a 25.4% decrease in actuation effort, demonstrating stable and energy-efficient real-time automation.