The prediction of drill string torque and drag is critical to determine actual mechanical drill string loads. In addition accurate information about expected hookloads allows determining abnormal wellbore conditions. Standard analytical calculations of torque and drag require a friction factor. This pseudo friction factor can be found in an iterative way based on measured hookload. Translating surface measurements to downhole conditions, i.e. actual weight on bit based on the string tension profile, would allow a much more precise analysis of the bit performance. Simulating torque and drag over time using analytical methods requires a large number of input variables, which typically are not readily available in combination with a large number of unknowns, contribute to inaccurate results. As part of this work an alternative approach using neural networks was tested. These networks are trained on ream and wash sequences where the wellbore can be classified as clean. The result is a hookload and torque prediction for a clean wellbore over time. Deviations between prediction and actual measurement can be interpreted as an abnormal wellbore condition. It can be concluded that hookload and torque predictions using neural networks model show good results. A key element is the selection of adequate training data sets and the development of features to characterize a wellbore geometry. Simulations are efficient and can be performed in real-time in a rig environment.