The analysis and optimisation of large ground source energy systems involving the use of thermo-active piles requires the ability to predict the thermal performance of these geothermal structures using limited computational resources, as many different configurations need to be tested. Current design methods are either based on empirical expressions, the accuracy of which is necessarily limited, or on thermo-hydraulic modelling which is computationally expensive and hence of difficult integration with optimisation procedures. In this paper, a surrogate model of a single thermo-active pile is established by running multiple thermo-hydraulic finite element analyses using different combinations of thermal ground properties, pipe arrangement, fluid temperature, pile length and pile diameter determined using a Latin hypercube sampling approach. The database of results is then used to train an artificial neural network (ANN), which is shown to produce accurate predictions of the thermal performance of a thermo-active pile given its characteristics and those of the surrounding ground. Given the low computational cost of surrogate models, this approach enables the design optimisation of large systems with greater confidence than previously possible using empirical relationships and a fraction of the resources required by thermo-hydraulic finite element models.