Micro-meteorological aspects of city comfort, land use management and air quality monitoring are rapidly growing areas of applications where environmental turbulence-resolving or large-eddy simulation (LES) models play the central role. Complex details of the urban surface morphology however remain unresolved or poorly resolved in the state-of-the-art LES due to severe limitations from computer facilities. The LES code LESNIC is applied in this study to simulate turbulent flow interaction with elements of the urban surface morphology. The study investigates a possibility to utilize a trained three layers’ artificial neural network (ANN) to parameterize the flow-to-surface interactions in the coarse LES where surface features are unresolved. It is concluded that the ANN can be a robust predictor for scalar concentrations and components of the surface stress tensor in the urban sub-layer with unresolved scalar sources and surface morphology. It has been noted however that dynamic training of the ANN may require more computational resources than adequate refinement of the LES resolution.