Graph neural networks classify molecular geometry and design novel order parameters of crystal and liquid

S Ishiai and K Endo and K Yasuoka, JOURNAL OF CHEMICAL PHYSICS, 159, 064103 (2023).

DOI: 10.1063/5.0156203

Molecular dynamics simulation produces three-dimensional data on molecular structures. The classification of molecular structure is an important task. Conventionally, various order parameters are used to classify different structures of liquid and crystal. Recently, machine learning (ML) methods have been proposed based on order parameters to find optimal choices or use them as input features of neural networks. Conventional ML methods still require manual operation, such as calculating the conventional order parameters and manipulating data to impose rotational/translational invariance. Conversely, deep learning models that satisfy invariance are useful because they can automatically learn and classify three-dimensional structural features. However, in addition to the difficulty of making the learned features explainable, deep learning models require information on large structures for highly accurate classification, making it difficult to use the obtained parameters for structural analysis. In this work, we apply two types of graph neural network models, the graph convolutional network (GCN) and the tensor embedded atom network (TeaNet), to classify the structures of Lennard-Jones (LJ) systems and water systems. Both models satisfy invariance, while GCN uses only length information between nodes. TeaNet uses length and orientation information between nodes and edges, allowing it to recognize molecular geometry efficiently. TeaNet achieved a highly accurate classification with an extremely small molecular structure, i.e., when the number of input molecules is 17 for the LJ system and 9 for the water system, the accuracy is 98.9% and 99.8%, respectively. This is an advantage of our method over conventional order parameters and ML methods such as GCN, which require a large molecular structure or the information of wider area neighbors. Furthermore, we verified that TeaNet could build novel order parameters without manual operation. Because TeaNet can recognize extremely small local structures with high accuracy, all structures can be mapped to a low-dimensional parameter space that can explain structural features. TeaNet offers an alternative to conventional order parameters because of its novelty.

Return to Publications page