Broadening the Scope of Neural Network Potentials through Direct Inclusion of Additional Molecular Attributes

Area of Science:

• Computational chemistry
• Materials science
• Machine learning in science

Background:

• State-of-the-art neural network potentials (NNPs) typically use only atomic numbers and positions.
• This limited input restricts the applicability of NNPs to systems with varying electronic configurations.
• Existing models face input degeneracy issues, hindering accurate predictions for complex chemical environments.

Purpose of the Study:

• To demonstrate the importance of including electronic attributes in NNPs.
• To enhance the predictive accuracy and applicability of neural network potentials.
• To address input degeneracy issues in current NNP models.

Main Methods:

• Minimal architectural modification to TensorNet, an equivariant NNP model.
• Inclusion of additional electronic attributes alongside atomic numbers and positions.
• Experimental validation using custom and public benchmarking datasets.

Main Results:

• The modified TensorNet resolved input degeneracy issues inherent in position- and atomic number-based inputs.
• Enhanced predictive accuracy was observed across diverse chemical systems, including those with varying charge and spin states.
• The improvements were achieved without specialized strategies or physics-based energy terms.

Conclusions:

• Incorporating electronic attributes into NNPs significantly improves their performance and applicability.
• Minimal architectural changes can lead to substantial gains in predictive power for computational chemistry.
• Future research should prioritize NNPs that leverage these richer representations for greater accuracy and broader use.