Prediction of the Potential Energy Surface for polyatomic molecule with the Three-dimensional Special Euclidean Equivariant Transformer Network

Feng An,Longchao Da,Lehao Yang,Shanyu Han,Hua Wei

DOI: https://doi.org/10.26434/chemrxiv-2024-62lrz

2024-02-07

Abstract:The process of fitting potential energy surfaces using machine learning methods typically involves manually constructing feature vectors and transforming molecular graphs into the inputs of networks for the polyatomic molecule structure. In this study, we introduce a novel approach using a three-dimensional special Euclidean equivariant transformer network that can directly learn the potential energy of polyatomic molecules and represent the structure of the molecule in a universal and interpretable way. Our method accurately predicts the potential energy of polyatomic molecules, as determined by coupled cluster theory, for various molecular graphs. Moreover, our framework is interpretable about the molecular physics, as one can extract molecular equivariant positional information regarding the global invariant energy surface. To demonstrate the utility of our approach, we present a detailed description of the training process used to fit the potential energy surfaces of polyatomic molecule CH5, as well as the properties of its resulting potential energy surfaces.

Chemistry

What problem does this paper attempt to address?

This paper primarily addresses how to more effectively utilize machine learning methods to construct the potential energy surface (PES) of multi-atomic molecules. Traditional approaches typically require manual construction of feature vectors and conversion of molecular graphs into network inputs. The study proposes a new method called the Three-Dimensional Special Euclidean Equivariant Transformer Network (SE3-Equivariant Transformer Network), which can intuitively learn the potential energy of multi-atomic molecules and represent molecular structures in a universal and interpretable manner. This method accurately predicts the potential energy of different molecular graphs and has interpretability for molecular physics, being able to extract molecular equivariant position information related to the global invariant energy surface.

The paper demonstrates the fitting of the potential energy surface of the multi-atomic molecule CH5 through the training process of SE3-Equivariant Transformer Network, as well as the properties of the resulting potential energy surface. Compared to traditional methods, this framework considers the invariance and equivariance of rotational, translational, and permutation symmetries, improving the efficiency of predicting molecular potential energy. By using the molecular's Cartesian coordinates as equivariant input and the potential energy as invariant output, the network can learn molecular features considering symmetry.

Experimental results show that the model's error in constructing the potential energy surface of the CH5 molecule is within an acceptable range and consistent with experimental results. Compared to other methods such as FI-NN and PIP-NN, the SE3-Equivariant Transformer Network uses fewer polynomial inputs while still maintaining energy invariance, thereby improving prediction accuracy.