Introducing Virtual Points in Equivariant Networks by Extending Atom Representation for Effective Prediction

Area of Science:

• Computational Chemistry
• Machine Learning in Chemistry
• Quantum Chemistry

Background:

• Current equivariant models represent molecules as fixed atoms (ball-and-stick).
• These models may not fully capture the atomic environment, including electron configurations.
• Limitations exist in representing the complex interactions around atomic nuclei.

Purpose of the Study:

• Introduce Neural Polarization (NP), a novel method to enhance molecular representations.
• Extend equivariant networks to incorporate richer atomic environments.
• Improve prediction performance in molecular modeling tasks.

Main Methods:

• Embed each atom as a pair: an original atom and a virtual atom.
• Update virtual atom positions through parameterization during training.
• Apply NP to existing equivariant models flexibly.

Main Results:

• NP improves prediction performance across various targets, including electron density.
• Experimental results demonstrate enhanced accuracy on benchmark datasets.
• Extended atomic representations lead to better overall molecular task performance.

Conclusions:

• Neural Polarization offers a significant advancement in molecular representation learning.
• The method effectively captures electron density and surrounding atomic environments.
• NP provides a versatile approach to boost the capabilities of existing equivariant models.