This paper focuses on 2 different models, the multiple regression method and the artificial neural network (ANN), for predicting surface roughness (R-a). Experiments were conducted to measure surface roughness in the cylindrical grinding of AISI 52100 bearing steel which had been conventionally heat-treated and deep cryogenically treated (-145 degrees C). In order to compare the effects of holding time at the deep cryogenic temperatures, 5 different holding times (12, 24, 36, 48 and 60 h) were employed to obtain the optimum R-a. The cylindrical grinding test results showed that optimum R-a values were obtained on specimens cryogenically treated for 36 h. In addition, the prediction results showed that the ANN was superior to the multiple regression method in terms of prediction capability. Moreover, due to a higher determination coefficient (R-2) and lower root-mean-square error (RMSE) and mean error percentage (MEP), the ANN was notably successful in predicting the R-a.