The operational optimization of Small Modular Reactors critically depends on accurate real-time monitoring of Once-Through Steam Generators. Complex thermo-fluid phenomena within OTSGs present significant modeling challenges, making high-fidelity simulations computationally prohibitive; meanwhile, conventional data-driven models suffer from out-of-distribution generalization failure. To bridge this gap, this study proposes a novel Domain-Adaptive Physics-Informed Transformer model enhanced with a systematic transfer learning framework. The model integrates three key innovations: multi-scale spatial discretization (z-binning) focusing on the critical boiling zone, a dual-head architecture for simultaneous prediction of field distributions and outlet temperatures, and physics-based regularizers enforcing thermodynamic consistency, specifically a total-variation smoothness penalty and a monotonicity constraint grounded in the Second Law of Thermodynamics. The proposed model demonstrates consistent and significant superiority over strong baseline models such as Long Short-Term Memory, Autoencoder with Attention, and Feedforward Neural Network, achieving near-perfect accuracy (R2 = 0.998–0.999) for outlet temperatures, phase void fraction, axial temperature, and velocity magnitude, while maintaining remarkable robustness under 2%–10% additive sensor noise. After pre-training on source-domain simulation data, the model is fine-tuned using 10%–90% of target-domain data and attains strong performance with only 50%–60% of target data, demonstrating exceptional data efficiency. Ablation experiments confirm the individual contribution of each architectural component, and evaluation on unseen target-domain data validates robust cross-domain generalization. This work provides a practical, deployment-ready framework demonstrating that physics-aware deep learning, coupled with strategic transfer learning, can yield highly adaptable and reliable surrogate models for nuclear engineering applications.