<p>Computationally intensive hydrodynamic simulations are commonly used to evaluate the performance of wave energy converter (WEC) arrays, but their cost limits large-scale design exploration. This study investigates surrogate modelling approaches for predicting total farm-level absorbed power of large WEC arrays using simulation datasets generated by a frequency-domain potential-flow hydrodynamic model under wave climate conditions corresponding to offshore Perth and Sydney. Surrogate models trained directly on planar device coordinates are evaluated alongside models that augment these inputs with geometric features describing spatial relationships between devices. Results for arrays comprising 49 and 100 devices show that augmenting coordinate-based inputs with geometric features consistently improves predictive accuracy and explained variance relative to coordinate-only inputs. Additional experiments demonstrate that these models retain meaningful predictive performance even when trained on substantially reduced training datasets. Overall, the results indicate that explicitly encoding spatial relationships within the input representation provides a more effective basis for surrogate modelling of large wave energy farms. The proposed approach improves predictive accuracy and demonstrates improved data efficiency, enabling more computationally efficient analysis and early stage design exploration of large WEC arrays.</p>