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

Guillem Simeon, Antonio Mirarchi, Raúl P. Peláez, Raimondas Galvelis, Gianni De Fabritiis

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.

Introduction: The Challenge of Input Degeneracy

In the field of computational chemistry, neural network potentials (NNPs) have emerged as powerful tools for predicting the energy and forces of atomic systems at a fraction of the cost of traditional quantum mechanics. However, most state-of-the-art models rely solely on two primary inputs: atomic numbers ($Z$) and positions ($r$).

This reliance creates a fundamental flaw known as input degeneracy. Consider a neutral molecule and its positively charged cation; if they share the same geometry, a standard NNP perceives them as identical inputs, despite their vastly different physical energies and forces. This "blindness" to electronic states often forces researchers to "sanitize" datasets by removing charged species, which severely limits the model's realism and applicability to real-world biochemistry or materials science where ions and varying spin states are common.

Visualizing input degeneracy with neutral and charged molecular pairs.

Methods: A Minimalist Equivariant Modification

To address this, the authors propose a minimalist extension to TensorNet, a model that utilizes rank-2 Cartesian tensors ($3 \times 3$ matrices) to represent atomic environments. The key innovation is a node-wise modification of the interaction layers. Instead of introducing complex, manually-derived physics terms like explicit Coulomb or dispersion equations, the researchers introduce a state-dependent scalar factor:

$$Y'^{(i)} = (1 + \lambda_k \psi_k)(\dots)$$

Where $\psi_k$ represents a molecular attribute (such as total charge $Q$ or spin $S$) and $\lambda_k$ is a learnable weight. This approach is particularly elegant for several reasons:

  1. Equivariance: It preserves $O(3)$ equivariance (rotational and reflection symmetry) by simply rescaling tensor features with a scalar.
  2. Minimalism: It requires no fundamental architectural changes to the core TensorNet framework.
  3. Consistency: For neutral systems in a singlet state ($Q=0, S=0$), the model mathematically reverts to the original, high-performing TensorNet.

Results and Discussion

The researchers validated their approach across a diverse range of benchmarks, from custom-made toy datasets to large-scale public datasets like SPICE PubChem.

Resolving Degeneracy in the Toy Dataset

On a merged dataset containing both neutral and charged versions of the same geometries (Dataset A ∪ B), the baseline TensorNet failed significantly, yielding an error of 4437 meV. In contrast, the modified version achieved an error of just 2.5 meV, effectively solving the degeneracy problem.

Quantitative resolution of input degeneracy using modified TensorNet.

Broad Performance Gains

The benefits of "charge awareness" extended beyond just resolving degeneracy. In the SPICE dataset, including charge information improved accuracy for neutral molecules as well, suggesting that the model develops a more physically meaningful internal representation of chemistry. Furthermore, on the Solvated Protein Fragments test, the modified TensorNet ($0.55 \text{ kcal/mol}$) nearly halved the error of an ensemble of five PhysNet models ($0.95 \text{ kcal/mol}$).

The model also achieved a 10-fold accuracy improvement on the QMspin dataset, overcoming the notorious hurdle of spin state degeneracy. Importantly, these gains were achieved while maintaining TensorNet’s $O(N)$ scaling, making it faster than competing models like SpookyNet in head-to-head benchmarks.

Conclusion

By directly incorporating molecular attributes into the neural network representation, this work demonstrates a path toward potentials that are no longer "blind" to the electronic state of a system. This modification allows for the accurate simulation of charged and spin-polarized systems without sacrificing computational efficiency or requiring complex physics-based energy terms. These findings encourage the broader adoption of additional electronic representations in the next generation of neural network potentials.

Summary updated

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.
BibTeX
@article{simeon2025broadening,
  author = {Guillem Simeon and Antonio Mirarchi and Raúl P. Peláez and Raimondas Galvelis and Gianni De Fabritiis},
  title = {{Broadening the Scope of Neural Network Potentials through Direct Inclusion of Additional Molecular Attributes}},
  journal = {Journal of Chemical Theory and Computation},
  year = {2025},
  doi = {10.1021/acs.jctc.4c01625},
}
Figure