We use deep learning techniques to model computational fluid dynamics (CFD) simulations of wind flow over a complex topography. Our motivation is to "speed up" the optimisation of CFD-based simulations (such as the 3D wind farm layout optimisation problem) by developing surrogate models capable of predicting the output of a simulation at any given point in 3D space, given output from a set of training simulations that have already been run. Our promising results using TensorFlow show that deep neural networks can be learned to model CFD outputs with an error of as low as 2.5 meters per second.