<p>After tobacco smoking, radon exposure is the second most common cause of lung cancer. To accurately measure ²²⁰Rn and its progeny doses, a high-precision thoron chamber is required to calibrate the radiation detectors. However, traditional empirical regulation and analysis methods struggle to satisfy the accuracy requirements for state-parameter control. This study proposes a hybrid data-physics-driven approach to establish an efficient prediction method for ²²⁰Rn and its progeny concentrations, enabling the precise regulation of state parameters in a thoron chamber. First, a high-fidelity computational fluid dynamics model of the thoron chamber was developed and experimentally validated to generate a reliable database for neural-network training. Innovatively, the diffusion equations of ²²⁰Rn and its progeny, along with fluid mass conservation equations, were embedded as physical constraints into the neural-network architecture. The resulting neural-network model achieved rapid prediction of ²²⁰Rn/progeny concentrations and flow-field parameters. The predicted concentration distribution patterns and flow-field characteristics showed strong consistency with previous research, demonstrating prediction deviations within 1.8% for concentrations and 3.27% for flow-field parameters. Notably, the prediction time was reduced by four orders of magnitude compared to that of traditional computational fluid dynamics methods. This breakthrough has significant theoretical importance and practical value for advancing the metrological technology of ²²⁰Rn and its progeny.</p>