In many engineering applications, operational structures are instrumented with sensors that produce large quantities of data. This observed data is commonly utilised by machine learning tools to infer information about the structures. Physics-based and finite element models of these structures are also used to model structural responses; often parameters within these models are updated using the incorporation of observed data. However, it is still the physics-based model that is used to generate response predictions. This talk considers the case of instrumented concrete sleepers on a railway bridge. The application of a novel machine learning technique, based on multi-output Gaussian processes, is then presented; it uses a new type of methodology for blending data and physics: data-centric engineering. The technique utilises Gaussian process priors for both the observed data and the physics-based model, connected by the geometric relationship between the two physical quantities they measure. The method can be interpreted as the inference of railway sleeper response from data, with the physics-based model guiding the inference in unmeasured areas. The benefits of combining the observed data and physics in this way include the estimation of material properties and a tool for damage detection.