RESUMEN
In the past three years, machine learning (ML) in combination with density functional theory (DFT) has enabled computational screening of compounds with the goal of accelerated materials discovery. Unfortunately, DFT+ML has, until now, either relied on knowledge of the atomic positions at DFT energy minima, which are a priori unknown, or been limited to chemical spaces of modest size. Here we report a strategy that we term learning-in-templates (LiT), wherein we first define a series of space group and stoichiometry templates corresponding to hypothesized compounds and, orthogonally, we allow any list of atoms to take on any template. The LiT approach is deployed in combination with previously established position-dependent representations and performs best with the representations that rely least on the atomic positions. Since the positions of the atoms in templates are known and do not change, LiT enables us to infer the properties of interest directly; additionally, LiT allows working with increased chemical spaces, since the same elements can take on a large number of templates. Only by using LiT were we able to span 5 × 106 double-perovskite compounds and achieve an acceleration factor of 700 compared to brute-force DFT, allowing us to predict never-before-screened compounds. Our findings motivated us to synthesize a new BaCu yTa(1- y)S3 perovskite, which we show using an electron probe microanalyzer has a 5:3 molar ratio of Cu to Ta and, using powder X-ray diffraction (XRD) analysis combined with a DFT-based XRD simulation and fitting, indicate a new phase having an I4/ m space group.