Scalable physics-embedded graph neural networks for million-node thermal-fluid digital twins of converter transformers

Real-time artificial intelligence (AI) digital twins for heavy power equipment require high-fidelity forward simulation for anomaly detection, yet traditional solvers suffer from prohibitive latencies. To bridge this gap, we present a physics-embedded deep learning framework achieving an approximate 970-fold acceleration over conventional finite element methods, reducing forward inference latency to 1.54 s on a single NVIDIA RTX 4090 Graphics Processing Unit (GPU). Traditional solvers for strongly coupled conjugate heat transfer (CHT) and thermo-fluidic partial differential equations (PDEs) face severe complexity bottlenecks on unstructured meshes, while soft-constraint operator learning architectures suffer from interfacial physical distortions. To resolve these limitations, we propose the Physics-Embedded Multi-Feature Graph Attention Network (MFGAT), a reduced-order neural operator solver tailored for million-node thermo-fluidic systems. MFGAT introduces a physics-embedded segregated iterative reasoning mechanism that bypasses explicit velocity field fitting, embedding nonlinear temperature-properties causal feedback as an autoregressive structural constraint to mitigate numerical stiffness in low-Mach-number natural convection flows. Furthermore, a geometry-aware dynamic graph sampling strategy preserving Euclidean isotropy is designed to resolve out-of-memory (OOM) bottlenecks during large-scale message passing. Evaluated on a 915,000-node industrial converter transformer mesh, MFGAT reduces global Mean Absolute Error (MAE) by approximately 50% compared to data-driven baselines, with over an 80% error reduction in local high-gradient regions. Spatial holdout tests demonstrate reliable local extrapolation under geometric variations without global variable shortcuts, providing an efficient, physics-bounded forward inference pathway.