Motivation and Problem Statement

Predicting whether a certain crystalline form of a compound can be synthesized or naturally formed is not an easy task! In fact, a general and accurate predictive model for synthesizability across various crystal structure types and chemical compositions is still absent. But this absence is not accidental! The synthesizability of a crystalline material is governed by many parameters, such as thermodynamic stability, the kinetics of the synthesis process, and synthesis routes (synthesis history), making a one-on-one map of synthesizability to simple physical parameters infeasible. Yet, predicting whether a novel crystalline material is synthesizable or not is critical for accelerating materials discovery because the structure-property relationship often necessitates the synthesis of a specific crystal form for a compound.

In this paper, we have developed a deep learning model to tackle the problem of synthesizability prediction! We were inspired by the promise of convolutional neural networks in pattern recognition and image classification. We thought that if we represent crystalline materials by digitized images, we can readily utilize the powerful convolutional neural networks for learning the hidden patterns that govern their synthesizability. 

Technical Approach 

Our approach is straightforward. We represent crystalline materials by three-dimensional (3D) voxel-wise images. Voxels are color-coded by the chemical attributes of the elements occupying them so that each crystalline material is digitized into a uniquely distinguishable image. We trained both supervised and unsupervised convolutional neural networks to learn the visual features embedded in the 3D digitized images. The learned features can now be used to predict the synthesizability of never-seen-before hypothetical crystalline materials. 

We collected the class of synthesizable crystalline materials from available crystallographic databases where synthesized crystalline materials are reported. However, we needed a second class of unsynthesizable crystalline materials, which we refer to as crystal anomalies, i.e., the crystalline forms of compounds that are really hard to synthesize. Generating crystal anomalies is not straightforward because many of the unobserved or never-synthesized crystalline forms of a compound could be easily unexplored rather than unsynthesizable. We, therefore, generated a data set of crystal anomalies according to the following approach: We identified the unobserved crystal structures for the most studied compounds in the published literature by utilizing natural language processing tools. These unobserved crystal structures are most likely unsynthesizable given that the compound has been well-explored and studied. 

Key Outcomes and  Usefulness

The primary outcome of our work is a predictive model for synthesizability prediction across a wide range of crystalline materials, from elemental ionic and covalent crystals to complex molecular crystals, which can aid in the synthesis of future crystalline materials in many different applications. We have illustrated the usefulness of our model by predicting the synthesizability of hypothetical crystals for battery electrode and thermoelectric applications. These two sets of crystalline materials have a different distribution of samples (chemical compositions and crystal structures distribution) compared to the samples in the training data and can show how well the prediction of our model can be generalized to the out-of-range samples. One remarkable observation is that our studies on both the electrode and thermoelectric materials indicate that unsupervised learning of synthesizability features has a stronger generalization compared to supervised learning, which further encourages the use of unsupervised learning for different tasks, given that proper unsupervised learning methods are used.