How to Search for Stable Inorganic Compounds More Efficiently

The computational search for new stable inorganic compounds remains highly expensive, due to the combinatorically explosive number of hypothetical compounds to consider. To guide the search towards the most likely stable compounds, several recommendation engines have been developed, varying in their strategy from data mining to machine learning. We conduct a systematic comparison of the performance of previously developed recommendation engines in recovering stable hypothetical compounds in the Open Quantum Materials Database (OQMD), and develop workflows to execute these methods in a highly efficient manner. For example, we find that crystal graph convolution neural networks outperform methods based on substitution of chemically similar elements into existing compounds; that employing a feedback loop (where method parameters are periodically retrained during execution) greatly improves the performance of recommendation engines in identifying already-calculated stable Heusler compounds; and that design of training set is crucial for the performance of neural networks. We also examine the status of stable compound searches in OQMD and find that, while thousands of non-experimentally-known stable compounds have already been identified, there are evidently at least thousands more that remain to be found, and hence the quest for materials discovery is far from over.