Abstract: To address the problems of redundant temperature measuring points, thermal response lag, and insufficient prediction stability under cross-condition scenarios in thermal error prediction of direct-drive dual-axis rotary tables, a thermal error prediction method for the functional reference point based on a physics-informed neural network was proposed. Thermal drift data and temperature response data under different operating conditions were obtained using a ballbar and a multi-channel temperature acquisition system. Representative temperature measuring points were selected through time-lag cross-correlation analysis and clustering analysis. A thermal error prediction model was then constructed by incorporating initial-condition constraints, thermal inertia-dynamic constraints, and quasi-steady-state constraints. The results showed that the proposed model achieved good prediction performance in the main thermal drift directions. The RMSE values of thermal error prediction in the Y and Z directions were 0.15 μm and 0.10 μm, respectively. The proposed model outperformed multiple linear regression, support vector regression, and multilayer perceptron models. Sparse-sample experiments showed that the model still maintained stable prediction performance when the training sample ratio was reduced to 5%. The proposed method provides a reference for thermal error modeling and compensation of direct-drive dual-axis rotary tables.