A Digital Twin Framework for Autonomous Kinetic Facades: Bio-inspired SMA Actuation with Reinforcement Learning Control for Tropical Climates
Abstract
Building envelope automation is limited by heavy motor-driven shading and reactive rule-based controllers. This paper presents an autonomous Digital Twin framework for kinetic facades, integrating bio-inspired Shape-Memory Alloy [SMA] actuation with a Reinforcement Learning [RL] control policy. Conceived as a Robotic Skin, Nitinol SMA wires act as artificial muscles mimicking plant stomatal turgor, enabling silent, high energy-density, joint-free actuation. A Python 3.10 Digital Twin (Pandas, NumPy, Scikit-learn) simulates a south-facing 10 m² facade [SHGC 0.60] in Da Nang, Vietnam, a tropical coastal city with high insolation and salt-air corrosion. Observed hourly weather data for Da Nang International Airport [WMO 488550, TMYx 2011-2025, 8760 hours] is split into 6552 training and 2208 test hours. Control is modeled as an MDP [State: GHI, T_out, T_in; Action: {Open, Partially Shaded, Closed}; Reward: -Energy + Comfort], solved by Q-Learning with ε-greedy decay [1.0→0.05] over 10,000 iterations across n=10 seeds. On held-out data, RL achieves 8.67 ± 0.95 kWh cooling load versus 37.34 kWh static and 21.71 kWh PID, reductions of 76.8% and 60.1%, with 74.1% peak load and 86.5% solar gain cuts. An IoT feedback loop and anti-corrosive encapsulation support coastal resilience.
