Predicting the contact mechanical response for various types of surfaces is and has long been a subject, where many researchers have made valuable contributions. This is because the surface topography has a tremendous impact on the tribological performance of many applications. The contact mechanics problem can be solved in many ways, with less accurate but fast asperity-based models on one end to highly accurate but not as fast rigorous numerical methods on the other. A mathematical model as fast as an asperity-based, yet as accurate as a rigorous numerical method is, of course, preferred. Artificial neural network (ANN)-based models are fast and can be trained to interpret how in- and output of processes are correlated. Herein, 1,536 surface topographies are generated with different properties, corresponding to three height probability density and two power spectrum functions, for which, the areal roughness parameters are calculated. A numerical contact mechanics approach was employed to obtain the response for each of the 1,536 surface topographies, and this was done using four different values of the hardness per surface and for a range of loads. From the results, 14 in situ areal roughness parameters and six contact mechanical parameters were calculated. The load, the hardness, and the areal roughness parameters for the original surfaces were assembled as input to a training set, and the in situ areal roughness parameters and the contact mechanical parameters were used as output. A suitable architecture for the ANN was developed and the training set was used to optimize its parameters. The prediction accuracy of the ANN was validated on a test set containing specimens not seen during training. The result is a quickly executing ANN, that given a surface topography represented by areal roughness parameters, can predict the contact mechanical response with reasonable accuracy. The most important contact mechanical parameters, that is, the real area of contact, the average interfacial separation, and the contact stiffness can in fact be predicted with high accuracy. As the model is only trained on six different combinations of height probability density and power spectrum functions, one can say that an output should only be trusted if the input surface can be represented with one of these.