Principal feasability studies using neuro-numerics for prediction of flow fields

In this paper initial studies of the application of a hybrid model using artificial neural networks and conventional numerical methods to predict - as an example - twodimensional, isothermal, steady flow fields is presented. Main topics of the work were to show the principal possibility of using ANN in fluid mechanics and, additionally, to realize the potential of incorporation a priori knowledge of physical phenomena into the training procedure. For training, as well as for rating the prediction, flow fields, consisting of velocity, temperature and pressure, were generated by numerical simulation. Major result was that prediction of the flow and especially the existance of vortices in the bodies outflow depending on the Reynolds number can be realized with a much lesser time consumption than necessary for numerical calculation. Furthermore, a priori physical knowledge could be included in the learning process with an obvious improvement of the predicting ability of the hybrid model.