Most industries are now aware of the far-reaching potential benefits of computational-materials science. Including modelling in industrial research can help “prune” the space of resource-intensive R&D experimenting and testing on the basis of appropriate materials screening information. Identifying existing materials with a desired set of properties is most likely to be helped by browsing all the existing data on candidate materials. Crucially, the set of desired properties may have been already computed for a vast, but not exhaustive list of material compounds. 

Inferring the same property for compounds composed of a much wider combinatorial space of single ingredient elements becomes, in this scenario highly desirable: and crucially it can help identify, at virtually zero cost, suitable new candidate materials for the application at hand. Any such screening of candidate materials is, however, only good if the inferred estimate of the new material’s performance is provided with a tightly controlled error bar. Appropriate inference tools should in other words ideally allow industrial R&D users to interrogate the data and be advised, in each case, on how confidently the performance of the new material can be predicted on the basis of the whole available information (estimated in millions of calculation data entries in the NOMAD repository).