Predicting materials properties without crystal structure: deep representation learning from stoichiometry

Abstract

Machine learning has the potential to accelerate materials discovery by accurately predicting materials properties at a low computational cost. However, the model inputs remain a key stumbling block. Current methods typically use descriptors constructed from knowledge of either the full crystal structure — therefore only applicable to materials with already characterised structures — or structure-agnostic fixed-length representations hand-engineered from the stoichiometry. We develop a machine learning approach that takes only the stoichiometry as input and automatically learns appropriate and systematically improvable descriptors from data. Our key insight is to treat the stoichiometric formula as a dense weighted graph between elements. Compared to the state of the art for structure-agnostic methods, our approach achieves lower errors with less data.

Compositionally restricted attention-based network for materials property predictions

“…This allows for consistent comparison to past works [5][6][7]9 . Figure 1 shows the performance metrics from training and testing the models on all the benchmark materials properties outlined above.…”

“…The random forest (RF) model utilizes a Magpiefeaturized CBFV to represent chemistry. This is included as a performance baseline to match similar works 5,9,36 .…”

“…Model weights are updated using the look-ahead 62 and Lamb optimizer 63 with a learning rate that is cycled between 1 × 10 −4 and 6 × 10 −3 every 4 epochs to achieve consistent model convergence. A robust MAE is used as the loss criterion for model performance 9 . The default parameters generalize well when predicting most of the benchmark materials properties.…”

“…Predictions for all materials properties were generated using code from the Roost repository 9 . Minor adaptations were made to the code to allow for automated training and benchmarking.…”

“…Materials informatics (MI) is the resulting field of research that utilizes statistical and machine learning (ML) approaches in combination with high-throughput computation to analyze the wealth of existing materials information and gain unique insights [2][3][4] . As this wealth has increased, practitioners of MI have increasingly turned to deep learning techniques to model and represent inorganic chemistry, resulting in approaches such as ElemNet, IRNet, CGCNN, SchNet, and Roost [5][6][7][8][9] . In specific cases including CGCNN and SchNet, the compounds are represented using their chemical and structural information 7,8,[10][11][12][13][14][15] .…”