Using Machine-Learning and Deep-Learning to Discover Relationships in Intercalation Compounds

Intercalation refers to the insertion of guest species into the interlayer space of a host material. This can create drastically different properties in the resulting compound and therefore, intercalation can be used to tune the properties of a material. Layered materials such as graphite are common host structures because of their large van der Waals gap between layers; this allows for the insertion of guest species, resulting in graphite intercalated compounds (GICs). As guest species insert, the interlayer space may expand or contract, the guest may donate electrons (n-doped) or accept electrons (p-doped) from the host--resulting in an increased conductivity of the GIC, and may insert with different staging--where a stage I intercalation compound has molecules inserted in between each host layer while a stage II is every other layer, and so on. Thusly, intercalation is a powerful tool, as it provides a unique avenue to tune and engineer the properties of a material.<br/><br/>However, because of the vast combinations of guest and host compounds as well as the variations in staging, understanding the relationships between guest, host, and GIC properties are incredibly complex and challenging. However, machine learning (ML) and deep learning (DL) provide a route to elucidate the complex relationships between the properties of guest species, graphite host, and intercalated compounds. Here, we successfully employ these techniques to elucidate structure-property relationships in GICs. This work has application toward improving efforts in engineering materials without needing to test every possible combination; thereby, materials can be strategically selected to synthesize for specific applications, including medical, optoelectronic, and energy storage (<i>Inorg. Chem. Front</i>., 2016, 3, 452; <i>Adv. Mater</i>., 2019, 1808213).<br/><br/>Example: Strategies for Synthesis of GICs<br/>By training on input parameters (donor/acceptor of the intercalant, stage, in-plane conductivity, and interlayer spacing of the GIC), we predicted thermodynamic values for the enthalpy of formation, ΔH_f ± 5%^, with 40% accuracy and entropy, ΔS ± 5%, with 80% accuracy. These results allow for the calculation of Gibbs Free Energy, ΔG_f, for the formation of a GIC. Furthermore, because of the dependance of on temperature, we can determine the conditions which a GIC will form—including temperature, pressure, and applied voltage. This is significant for the broader materials community, since these ML- and DL-discovered relationships provide a new tool to guide the parameters for the synthesis of GICs. Furthermore, these relationships create the possibility of predicting and discovering new GICs.<br/><br/>In summary, with this work, we successfully used ML to elucidate structure-property relationships between intercalants, graphite host, and GICs, including thermodynamic values, which could be relevent to efficiently synthesizing GICs. With DL, we utilized the individual properties of guest species, graphite, and GICs to predict composite characteristics of GICs. More broadly, this research can lead to facilitating more sustainable efforts in material design for intercalation compound