This paper proposes a new interatomic potential energy
neural network,
AisNet, which can efficiently predict atomic energies and forces covering
different molecular and crystalline materials by encoding universal
local environment features, such as elements and atomic positions.
Inspired by the framework of SchNet, AisNet consists of an encoding
module combining autoencoder with embedding, the triplet loss function
and an atomic central symmetry function (ACSF), an interaction module
with a periodic boundary condition (PBC), and a prediction module.
In molecules, the prediction accuracy of AisNet is comparabel with
SchNet on the MD17 dataset, mainly attributed to the effective capture
of chemical functional groups through the interaction module. In selected
metal and ceramic material datasets, the introduction of ACSF improves
the overall accuracy of AisNet by an average of 16.8% for energy and
28.6% for force. Furthermore, a close relationship is found between
the feature ratio (i.e., ACSF and embedding) and the force prediction
errors, exhibiting similar spoon-shaped curves in the datasets of
Cu and HfO2. AisNet produces highly accurate predictions
in single-commponent alloys with little data, suggesting the encoding
process reduces dependence on the number and richness of datasets.
Especially for force prediction, AisNet exceeds SchNet by 19.8% for
Al and even 81.2% higher than DeepMD on a ternary FeCrAl alloy. Capable
of processing multivariate features, our model is likely to be applied
to a wider range of material systems by incorporating more atomic
descriptions.