Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

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

Guillem Simeon, Antonio Mirarchi, Raul P. Pelaez, Raimondas Galvelis, Gianni De Fabritiis

TL;DR

The paper addresses the degeneracy in neural network potentials that arises when only atomic numbers and positions are used by introducing a minimal, zero-cost extension to TensorNet to include molecular attributes such as total charge $Q$ and spin $S$. This state-aware augmentation preserves $O(3)$-equivariance and works for both atomic and molecular inputs, avoiding physics-based energy terms. Across toy datasets, public benchmarks, and challenging systems, the approach resolves degeneracy and yields improved or competitive accuracy, including dramatic gains on charged species and spin-state differences (e.g., QMspin dataset). The findings demonstrate broader applicability of neural network potentials to charged and spin-rich chemical spaces, with practical implications for data efficiency and model versatility.

Abstract

Most state-of-the-art neural network potentials do not account for molecular attributes other than atomic numbers and positions, which limits its range of applicability by design. In this work, we demonstrate the importance of including additional electronic attributes in neural network potential representations with a minimal architectural change to TensorNet, a state-of-the-art equivariant model based on Cartesian rank-2 tensor representations. By performing experiments on both custom-made and public benchmarking datasets, we show that this modification resolves the input degeneracy issues stemming from the use of atomic numbers and positions alone, while enhancing the model's predictive accuracy across diverse chemical systems with different charge or spin states. This is accomplished without tailored strategies or the inclusion of physics-based energy terms, while maintaining efficiency and accuracy. These findings should furthermore encourage researchers to train and use models incorporating these additional representations.

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

TL;DR

The paper addresses the degeneracy in neural network potentials that arises when only atomic numbers and positions are used by introducing a minimal, zero-cost extension to TensorNet to include molecular attributes such as total charge and spin . This state-aware augmentation preserves -equivariance and works for both atomic and molecular inputs, avoiding physics-based energy terms. Across toy datasets, public benchmarks, and challenging systems, the approach resolves degeneracy and yields improved or competitive accuracy, including dramatic gains on charged species and spin-state differences (e.g., QMspin dataset). The findings demonstrate broader applicability of neural network potentials to charged and spin-rich chemical spaces, with practical implications for data efficiency and model versatility.

Abstract

Most state-of-the-art neural network potentials do not account for molecular attributes other than atomic numbers and positions, which limits its range of applicability by design. In this work, we demonstrate the importance of including additional electronic attributes in neural network potential representations with a minimal architectural change to TensorNet, a state-of-the-art equivariant model based on Cartesian rank-2 tensor representations. By performing experiments on both custom-made and public benchmarking datasets, we show that this modification resolves the input degeneracy issues stemming from the use of atomic numbers and positions alone, while enhancing the model's predictive accuracy across diverse chemical systems with different charge or spin states. This is accomplished without tailored strategies or the inclusion of physics-based energy terms, while maintaining efficiency and accuracy. These findings should furthermore encourage researchers to train and use models incorporating these additional representations.
Paper Structure (10 sections, 7 equations, 1 figure, 10 tables)

This paper contains 10 sections, 7 equations, 1 figure, 10 tables.

Table of Contents

  1. Introduction
  2. Method
  3. Results and discussion
  4. Toy degeneracy problem
  5. SPICE PubChem
  6. Carbon chain, silver clusters and sodium chloride clusters
  7. Solvated protein fragments
  8. QMspin
  9. Conclusion
  10. Hyperparameters

Figures (1)

  • Figure 1: Molecules included in the A and B toy datasets, from which 2,000 data points per molecule are obtained by generating conformers and computing potential energies and atomic forces using GFN2-xTBxtb. Columns illustrate degenerate pairs of molecules for a neural network that uses solely atomic numbers and positions as inputs.