The modelling of turbulent airflow with CFD is a computationally expensive task, yet vital for assessing architectural and urban ventilation concepts. Simplified airflow prediction methods exist, however at the cost of lacking accuracy and/or precision. This paper therefore proposes a hybrid simulation and deep learning approach. We utilize images of low resolution CFD domains as input to a U-Net neural network. The generated output is a high resolution upsampled image. As training data and application case, we use turbulent indoor flow with forced convection. Results are promising, as the generated flow fields can recreate higher level of details from the low resolution inputs as when compared to bicubic interpolation. However, the approach leaves room for improvement especially with respect to generated image sharpness.