Summary This paper presents a neural network architecture for prediction of production performance under different operating conditions by integration of domain insight and simulated production response data. The neural network topology in the developed approach is derived from interwell communication and connectivity between a producer and its surrounding supporting injection wells. Instead of a fully connected neural network that represents a global (field-scale) model that allows any injector to be connected to a given producer, and hence too many unrealistic and irrelevant connections, a local view is taken in building the proxy model. In this case, each producer is assumed to be supported by very few surrounding injection wells and is likely to have weak or no communication with distant wells. However, interwell connectivity in complex large-scale reservoirs is not just a function of distance and rather difficult to determine. Therefore, multiple randomly sized regions around each producer are considered to include different numbers of injectors in each local network for any given producer. The variability in the neighborhood size reflects the prior uncertainty about the potential connectivity between a producer and its nearby injection wells at different distances. This approach results in many local neural networks (several local networks per each producer) that can be aggregated into a single large neural network model with a predefined topological structure to represent possible connections. Training with simulated data is then used to estimate the weights in the resulting neural network architecture. Once the training process is completed, for each producer, the local model with the best prediction performance on the test data is selected and used to construct the final topology of the neural network model for the entire field. The method is applied to predict interwell connectivity and oil production in a large-scale mature field that undergoes waterflooding. The results demonstrate that even a simple domain insight, such as distance-based elimination of wells in a large field, can significantly reduce the amount of training data need and lead to noticeable improvement in the prediction performance of the resulting neural network model.