Elton Ogoshi de Melo1,Gustavo Dalpian1
UFABC1
Elton Ogoshi de Melo1,Gustavo Dalpian1
UFABC1
Traditional chemical heuristics have been very important for discovering new inorganic compounds in the past, but as they are usually based on patterns observed on a limited or specific set of materials, they are not generalizable as once thought. Today, as the material scientist can access the data of hundreds of thousands of inorganic materials present in databases, it has become frequent that many Material Science projects have become data-driven, i.e., they use large amounts of materials data and statistical methods, such as Machine Learning, to uncover new patterns and explore the phase space of compounds. In the ML context, recommender systems are widely used in various scientific and nonscientific fields. For instance, in the e-commerce area, these systems recommend users new products by analyzing their past purchased products or by collaborative filtering: a system where a heterogeneous network/graph composed of users and items and their interactions is used as input to recommend new products to users. By using this same principle framework and OQMD's database, we were able to build an analogous graph with the elements in the periodic table together with the set of all described crystal structures' sites as two classes of entities. The relationships between the two classes are given by the occupancy of a site by an element and it is weighted by the compound's thermodynamic stability. The heterogeneous graph is used to build an embedding space with a vector representation for each entity (elements and sites). These vector representations were then used as input to model a recommender system. In addition to finding new materials by uncovering new element-site occupancies, we were also able to analyze the embedding space and find interesting patterns on chemical similarities between elements and the local geometry of sites.