Zero-Shot Digital Twins Achieved with Graph Neural Networks.

Traditional Digital Twins often require extensive retraining or fine-tuning when the physical geometry or boundary conditions of a system change, limiting their adaptability. This research presents a groundbreaking framework for "Zero-Shot Digital Twins" designed to overcome this rigidity. It seamlessly integrates real-time visual perception with a physics-informed reasoning engine built upon a Thermodynamics-Informed Graph Neural Network (GNN). This GNN architecture is grounded in a metriplectic thermodynamic formalism, ensuring local energy conservation and non-negative entropy production through graph message passing. The framework also incorporates an auxiliary GNN to infer unobservable fields from sparse visual data, mitigating simulation startup transients. To bridge the gap between simulation and reality, a continuous closed-loop data assimilation mechanism tracks macroscopic deformations and fluid boundaries in real-time using deep segmentation networks and optical flow, dynamically correcting the simulation and preventing numerical drift. The approach demonstrates extreme generalization across diverse physical regimes, such as viscoelastic beams and viscous fluid sloshing, instantiating accurate simulations on novel, unseen geometries in real-time without specific retraining.