Materials Informatics of Molecular Crystals Compared by Molecular and Crystal Representations: The Case of Band Gap Prediction

Abstract

In materials informatics, the representation of material structures is fundamentally important to obtain better prediction results. Molecular crystals can be represented by both molecular and crystal representations, but there has been no examination to determine which representation is the most effective for the materials informatics of molecular crystals. In this work, different representations for molecular crystals were compared in an exemplified task of band gap prediction. We demonstrated that the predictive ability using molecular graph outperformed those of molecular fingerprints and crystal graphs. This result motivated the screening of molecular big data from PubChem, and the inference suggested candidate molecules of organic semiconductors for photovoltaics and luminescence. The novelty of this work relies on the representation comparison of molecular crystals and the finding that molecular graph works better even though the property prediction of crystalline materials. This finding will enable to machine-learning-aided screening and design of functional molecular crystals.

Exploration of elastic moduli of molecular crystals via database screening by pretrained neural network potential

“…Materials informatics offers advantages in materials screening, as demonstrated by machine learning-assisted research. [17][18][19] Neural network potentials (NNPs) can approximate the relationship between material structures and potential energies with an accuracy close to that of DFT-D, provided sufficient training data is available. There have been reported various geometric graph neural networks (GNNs), including CGCNN, SchNet, and MEGNet.…”