Mutual coupling model is crucial in designing multiple-input-multiple-output (MIMO) antennas since mutual coupling will degrade overall MIMO performance from distorted radiation patterns and reduced antenna efficiencies. Typically, full-wave simulations have to be employed, which is often timeconsuming. Here, the artificial neural network (ANN) method is developed to reduce the modelling time. Compared to previous ANN methods that directly use model parameters as input, a novel physical preprocessing approach is proposed to incorporate antenna correlation information before the network training. As a proof of concept, the mutual coupling model of a nonuniform strongly-coupled array is realized. Furthermore, we use the trained networks for capacity estimation and power allocation with the water-filling algorithm, showing favourable model performance. The proposed physically preprocessed ANN model significantly outperforms traditional analytical solutions and direct modelling networks in terms of prediction accuracy, dataset construction costs, and network convergence, which could facilitate the fast optimization and design of advanced antenna arrays for MIMO communications.