The model isn't actually learning to mimic linear interpolation. Source and target points are sampled randomly, and linear interpolation is used to randomly choose a midpoint, then the model is trained to predict that the velocity at the midpoint points towards the target. But during training this selection of source, target and midpoint is randomly repeated thousands of times, so for the same spatial location the model will see many examples all going in different directions. Eventually this randomness will average out and the model will learn to predict this average, which happens to be the "true" velocity field at that point. Since this velocity field is usually non linear, the model will learn a behavior that is non linear as well.