Traditionally, model-based approaches are widely used for leak detection in Water Distribution Networks (WDNs). However, these methods are computationally expensive for larger networks and are highly dependent on the model accuracy. In this work, we propose a hybrid modelling framework for leak detection, wherein, a surrogate model using a variational autoencoder (VAE) neural network is trained with a calibrated hydraulic model. The statistical measures on the performance of VAEs are then used for leak detection. The efficacy of the proposed framework is tested on a theoretical WDN and a real-world WDN using data from steady-state hydraulic simulations and sensor measurements.