The emergence of Digital Twins (DTs) has redefined the paradigms of monitoring and control in industrial engineering. However, achieving a seamless synchronization between the physical system and its digital counterpart remains a formidable challenge, particularly in highly non-linear and chaotic phenomena such as fluid sloshing. Traditional Computational Fluid Dynamics (CFD) methods are often too computationally expensive for real-time interaction, while standard data-driven models frequently fail to satisfy fundamental physical constraints, leading to unphysical predictions or numerical drift over time. In this work, we propose a Digital Twin framework specifically designed for fluid sloshing in containers. The core of the system is a Thermodynamics-informed Graph Neural Network (TIGNN) [1, 2] that operates under the GENERIC (General Equation for Non-Equilibrium Reversible-Irreversible Coupling) formalism [3]. This structure ensures that the learned dynamics strictly adhere to the first and second laws of thermodynamics, preserving energy and ensuring correct dissipation. To bridge the gap between the physical reality and the simulation, we integrate a real-time computer vision pipeline based on semantic segmentation. Our results demonstrate that this bidirectional coupling allows the Digital Twin to accurately reconstruct non-observable variables, such as internal velocity and internal energy fields, while maintaining global synchronization. By combining the inductive bias of graph-based architectures with the physical consistency of thermodynamics-informed learning, the proposed framework offers a scalable, efficient, and self-correcting solution for the real-time simulation of complex fluid-structure interactions.