Solid State Li-Ion Conductor Discovery Enabled by Graph Based Machine Learning

Finding a solid state Li-ion conductor with high ionic conductivity is critical for the development of all-solid-state batteries. Graph based machine learning with structrual information as input has been applied to predict many materials properties with high accuracy. However, Li ionic conductivity prediction with graph based machine learning is less seen in the literature. The dataset for Li ionic conductivity usually contains no structural information, thus graph based machine learning algorithms which requires structrual information cannot be applied.<br/><br/>Finding a solid state lithium(Li)-ion conductor with high ionic conductivity is critical for the development of all-solid-state batteries. A promising route to accelerate the materials discovery process is to build machine learning models that predict the Li ionic conductivity. However, graph based machine learning models, which reach high accuracy predictions for many materials properties, are less used for Li ionic conductivity prediction due to the lack of structural information in the Li ionic conductivity dataset.<br/>We propose to overcome the lack of structural information by constructing representative structures from input stoichiometry of the solid state Li-ion conductors. The stoichiometry of the ionic conductor is first matched with the most similar structure in the materials project database according to the Element mover distance and Pettifor scale. Element swapping, atom insertion and deletion are performed subsequently to match the stoichiometry of the representative structure to that of the ionic conductor. With the above structure generation method, we reconstructed an open-sourced Li ionic-conductivity dataset, resulting in 916 data points with representative structures. We used the dataset to train a Crystal Graph Convolutional Neural Network and achieved mean absolute error of 0.8 in log₁₀(mS/cm) scale on test set. Using the trained machine learning model, we accurately predicted the ionic conductivity of amorphous Li-ion conductors that are reported in recent literatures and are outside the training set.