Thermal-hydraulic coupling system, a physical process where flow and heat transfer interact mutually, is ubiquitous in nature and industrial applications. Although experimental measurements and solution of governing equations can be used to quantify physical fields such as velocity and temperature, acquiring dense and high-fidelity data is still challenging and costly. Here, we develop a physics-informed deep learning framework, the Thermal-Hydraulic Coupling Solution (THCS), that simultaneously exploits the information available from sparse data and governing equations (mass, momentum, and energy). THCS is composed of a physics-informed neural network, a thermal property mapping module, and a hydraulic parameterization modeling module. To demonstrate its capability, we choose the supercritical fluid turbulent convection, characterized by sharp changes in thermal properties when fluids cross the pseudocritical temperature. The ablation experiments demonstrate that THCS shows favorable generalizability and robustness, and the proposed multi-head structures improve both the convergence stability and prediction accuracy of models. Finally, we apply THCS to practical experiment scenarios, demonstrating its potential to directly quantify physical fields from sparse data.