Machine-Learning Accelerated Synthesis of Nitride Materials—Prediction of Synthesis Pathways

Nitride materials are underexplored, specifically magnetic nitrides. The Inorganic Crystal Structure Database (ICSD) lists about 210000 crystal structures of experimentally realized materials. A large fraction of these structures contain oxygen, many of them are oxide materials. However, fewer structures contain nitrogen and only very few nitrides, about 1100 structures, are known. Nitrides are materials with nitrogen atoms in a formal oxidation state of 3- that act as anions. Almost every transition metal and rare-earth element is found in the class of nitrides, however, only very few publications about magnetic nitrides are found in literature. Synthetic difficulties contribute to the limited exploration of this materials space.[1]<br/>Synthesizing new nitride phases and probing their magnetic behavior is part of our research groups effort to study the results magnetic superexchange coupling mediated by nitrogen atoms. In this class of materials, the magnetic properties arise from N^3- ions interacting with open <i>d</i>-shell transition metals or rare-earth ions. This work is distinct from most of the nitride research that focusses on semiconducting and metallic nitrides, such as GaN and TiN, for optoelectronic, piezoelectric, and lighting applications. A known magnetic nitride is Ca₃CrN₃, that is host to a quasi-1D antiferromagnetic spin chain with nitrogen mediated superexchange coupling and large exchange of <i>J</i> = 185 K [2] in comparison to the oxygen based spin chain β-TeVO₄ with <i>J</i> = 20 K.[3] Quasi-1D antiferromagnetic spin chains can exhibit exotic properties arising from quantum fluctuations.[4] Employing strong nitrogen-mediated superexchange can be a pathway to shift the emergence of desired quantum phases to higher temperatures.<br/>Machine-learning models were previously used to predict new ternary nitride phases [1], however, no guidance on possible synthesis pathways was given. Often multiple pathways could be feasible, like high temperature solid state and nitrogen gas flow reactions, liquid state reactions, high pressure reactions, and more exotic pathways that involve an arc-generated nitrogen plasma, for example. This can make it difficult to select a synthesis route for a given composition. In this work we employ a machine-learning model to predict feasible synthesis pathways for not experimentally realized nitrides. We retrieved a dataset of ~5000 experimentally realized compounds containing nitrogen from the ICSD database. Each entry of our dataset includes the chemical composition, crystal structure information (.cif files), and a publication (.pdf files) with synthesis information. 245 compounds were identified as nitrides and synthesis information was extracted. The trained model can predict if a given nitride structure can be synthesized using a simple solid-state nitrogen gas flow type reaction or if harder-to-realize high pressure conditions or the generation of a nitrogen plasma are necessary, for example. 421 nitride compounds were retrieved from the Materials Project that are classified by DFT as stable and as lowest on their convex hull. The machine learning model was used to identify 220 compounds that can be synthesized in a simple solid-state nitrogen gas flow reaction. Synthesis experiments on the identified phases are undergoing.<br/>[1] Sun, Bartel, Arca, Bauers, Matthews, Orvananos, Chen, Toney, Schelhas, Tumas, Tate, Zakutayev, Lany, Holder, Ceder, <i>Nat. Mater.</i> <b>18</b> (2019) 732–739.<br/>[2] Vennos, Badding, DiSalvo <i>Inorg. Chem.</i> <b>29</b> (1990) 4059–4062.<br/>[3] Weickert, Harrison, Scott, Jaime, Leitmae, Heinmaa, Stern, Janson, Berger, Rosner, Tsirlin, <i>Phys. Rev. B</i> <b>94</b> (2016) 064403.<br/>[4] Luther, Paschel <i>Phys. Rev. B</i> <b>12</b> (1975) 3908.