Predicting the Synthesizability of All Crystalline Inorganic Materials with Semi-Supervised Machine Learning and Learned Atomic Embeddings

The autonomous computational design of new materials is being actively pursued as a new paradigm of materials science research. However, the current lack of predictive capabilities for directly identifying synthesizable inorganic crystalline materials is a significant roadblock for achieving autonomous materials discovery. Computationally, inorganic material synthesizability has been assessed by employing high throughput density-functional theory (DFT) to calculate the formation energy of a material or through the enforcement of a charge-balancing criteria. However, using DFT and charge-balancing as proxy methods for assessing synthesizability can only account for 50% and 28% of all previously synthesized inorganic crystalline materials, respectively.<br/> <br/>In this work, we develop a deep-learning classification model (<i>SynthNN</i>) to directly predict the synthesizability of inorganic chemical formulas without requiring any structural information. We accomplish this goal by training <i>SynthNN</i> on a database of chemical formulas consisting of previously synthesized crystalline inorganic materials that has been augmented with artificially generated unsynthesized materials. Chemical formulas are represented by learning an optimal set of descriptors for predicting synthesizability directly from the corpus of synthesized materials, allowing our approach better capture the complex array of factors that influence synthesizability. <i>SynthNN </i>offers numerous advantages over previous methods for identifying synthesizable materials. Whereas expert synthetic chemists typically specialize in a specific chemical domain of a few hundred materials, this approach generates predictions that are informed by the entire spectrum of previously synthesized materials. Additionally, since this method trains directly on the corpus of synthesized materials (rather than employing proxy metrics such as energy above the hull or charge-balancing), this approach also eliminates questions of how well these metrics can describe synthesizability. Finally, this method is computationally efficient enough to enable screening through billions of candidate materials.<br/> <br/>We benchmark the performance of <i>SynthNN </i>in a head-to-head material discovery comparison against 20 expert chemists and material scientists. <i>SynthNN</i> outperformed all experts, achieved 3.6X higher precision and completed the task five orders of magnitude faster than the average human expert. Remarkably, without any prior chemical knowledge, our experiments indicate that <i>SynthNN</i> learns the chemical principles of charge-balancing, chemical families and ionicity, and utilizes these principles to inform synthesizability predictions. Since <i>SynthNN </i>can be seamlessly integrated with Materials Screening or Inverse Design workflows, the development of <i>SynthNN</i> serves to greatly improve the success rate of computational material discovery efforts by ensuring that the candidate materials discovered through these efforts are synthetically accessible.<br/> <br/>This work was performed under the auspices of the U.S. Department of Energy by Lawrence Livermore National Laboratory under Contract DE-AC52-07NA27344.