Thermal aspects of rolling process play an important role on the product quality and performance of rolling equipments. Inappropriate thermal crown causes to shape and geometry defect such as center buckle, wavy edge or ridge buckle. On the other hand, it can increase the rolling force and the energy consumption. Sheet and roll temperature can affect on production, so mainly high temperature causes the fire crack on the roll surface and wear.The aim of this study is to investigate the effects of the temperature on the quality of sheets. For this purpose analytical models are combined with neural network to predict the work roll temperature for application in the on-line control programs. In addition, the metal structure as a display of the thermal history is analyzed to follow the origins of defects in the production process because non-uniform cooling can affect on structures and mechanical properties of the sheet.In this study, an analytical model based on finite difference method is used under transient condition to calculate the temperature and thermal crown profile of the work roll. The model has the ability of accepting variable boundary conditions in circumferential and axial direction for different cooling configuratio such as using different types and numbers of nozzles and headers in different directions. The results of simulation are compared and verified with an actual rolling program result for which the temperature of a work roll was measured at Mobarakeh Steel Complex.The results of this model are used to train and back-up neural network so that it can cover output domain. After preprocessing, these data set are used to train, verification and test of the neural network. Regarding the large number of influential parameters in the work roll temperature, input correlation of network is assessed and parameters with high correlation coefficients are removed. Physical approach to input and output of neural network can help us to reduce the network size and error significantly.Considering the nature of data structure, dynamic and static networks are trained and their results are compared with each other. The results show that the static networks for any roll layer converge to smaller error than the others. These networks can predict current temperature field, by considering the temperature field in the 35 points of roll as the thermal roll history. The network inputs are sheet dimensions, initial sheet temperature, gap time between two passes and reduction in stands and outputs are the work roll temperature field. k ey Words: Thermal profile, hot rolling, work roll, on-line, neural network