Although solar radiation is a crucial parameter in designing solar power devices and studying land surface processes, long-term and densely distributed observations of surface solar radiation are usually not available. This paper describes the development of a 50-year dataset of daily surface solar radiation at 716 China Meteorological Administration (CMA) stations. First, a physical model, without any local calibration, is applied to estimate the daily radiation at all 716 CMA routine stations. Then, an ANN-based (Artificial Neural Network) model is applied to extend radiation estimates to earlier periods at each of all 96 CMA radiation stations. The ANN-based model is trained with recent reliable radiation data and thus its estimate is more reliable than the physical model. Therefore, the ANN-based model is used to correct the physical model dynamically at a monthly scale. The correction generally improves the accuracy of the radiation dataset estimated by the physical model: the mean bias error (MBE) averaged over all the 96 radiation stations during 1994–2002 is reduced from 0.68 to 0.11 MJ m2 and the root mean square error (RMSE) from 2.01 to 1.80 MJ m2. The new radiation dataset shows superior performance over previous estimates by locally calibrated Ångström-Prescott models. Based on the new radiation dataset, the annual mean daily solar radiation over China is 14.3 MJ m2. The maximal seasonal mean daily solar radiation occurs in the Tibetan Plateau during summer with a value of 27.1 MJ m2, whereas the minimal seasonal mean daily solar radiation occurs in the Sichuan Basin during winter with a value of 4.7 MJ m2.