How to separate two Ts in a pod: Classifying T- and T′-type Ruddlesden-Popper cuprates by machine learning

The first-order Ruddlesden-Popper (RP) phases A 2 BX 4 adopt three structure types that differ in coordination geometry around the B site: T-type (octahedral) and T′-type (square planar), which are most common, and T∗-type (square pyramidal), which is rare. Especially for RP cuprates A 2 CuO 4–δ , it is not intuitively obvious which structure is preferred depending on the combination of cations occupying the A site. Machine learning models were developed that can separate the T- and T′-type structures among these cuprates with an accuracy of >90 %, provided that the T∗-type does not form and the phases can be synthesized. Based on these models, structures were predicted for solid solutions ( A ′, A ″, A ‴) 2 CuO 4–δ containing a complex mixture of A cations ( A ′, A ″, A ‴ = Sr, La, Gd, Ho, In, Bi). The predictions were tested by targeting various members of these solid solutions through high-temperature reactions followed by slow cooling. Three samples contained pure RP phases which were confirmed to adopt the predicted structures: T-type for Sr 0.4 La 1.5 Ho 0.1 CuO 3.8 , and T′-type for Gd 1.7 Ho 0.2 Bi 0.1 CuO 4 and La 0.4 Gd 1.2 Ho 0.4 CuO 4 . Five other samples were mixtures that contained RP phases whose structures (when not T∗-type) were correctly identified by a slightly better performing model based on extra randomized trees classifier. Machine learning models to distinguish T- and T′-type structures of cuprates A 2 CuO 4–δ were experimentally validated by synthesis of new phases. • Machine learning models were built to classify structures of cuprates A 2 CuO 4–δ . • T- and T′-type structures can be distinguished with high accuracy. • ( A ′, A ″, A ‴) 2 CuO 4–δ phases with complex mixtures of A cations were predicted. • Predictions were validated by synthesis and structure determination.