In a recent study published in Physical Review Letters, a research team led by Prof. JIANG Bin from University of Science and Technology of China (USTC) proposed a recursively embedded atom neural network (REANN) model based on their previous work on creating high-precision machine-learning (ML) potential surface methods.
With the advancement of machine learning technologies, a common method to build potential functions is atom neural networks (ANNs), under which the total energy is the sum of each atomic energy dependent on the local environment. Three-body descriptors have long been considered complete to describe the local environment.
Recent work, however, has found that three-body (or even four-body) descriptors could lead to the local structural degeneracy and thus fail to fully describe the local environment. This problem has posed difficulties in improving the precision of ANNs’ potential surface training.
The REANN model, using a recursively embedded density descriptor, shares the same nature with the less physically intuitive message-passing neural networks (MPNNs). The team proved that iteratively passing messages (namely updating orbital coefficients) to introduce many-body correlations can achieve a complete description of the local environment without explicitly computing high-order features.
Researchers tested dataset of CH4 and bulk water, and results revealed the local completeness and nonlocality of this new model as well as its better accuracy than current ML models.
The study provides a general way to easily improve existing ML potential surface frameworks to include more complicated many-body descriptors without changing their basic structures. The research also realizes more accurate and efficient ML models.
Schematic diagram of the REANN model showing how the density descriptor is recursively embedded (Image by ZHANG Yaolong et al.)
（Written by ZHANG Wenjing, edited by LI Xiaoxi, USTC News Center）