This paper describes the application of neural network approaches to the discovery of new materials exhibiting thermoelectric properties. Thermoelectricity is the ability of a material to convert energy from heat to electricity. At present, only few materials are known to have this property to a degree which is interesting for use in industrial applications like, for example, large-scale energy harvesting [3, 8]. We employ a standard neural network architecture with supervised learning on a training dataset representing materials and later predict the properties on a disjoint test set. At this proof of concept stage, both sets are synthetically generated with plausible values of the features. A substantial increase in performance is seen when utilising available physical knowledge in the machine learning model. The results show that this approach is feasible and ready for future tests with experimental laboratory data.