Ab-Initio Potential Energy Surfaces by Pairing GNNs with Neural Wave Functions

Nicholas Gao, Stephan Günnemann

Abstract

Solving the Schrödinger equation is fundamental to understanding quantum mechanical properties, yet analytical solutions are only tractable for the simplest systems. While neural wave functions combined with Variational Monte Carlo (VMC) have recently matched the accuracy of top-tier classical methods, they are computationally hindered by the need to train a separate model for every molecular geometry. We present PESNet (Potential Energy Surface Network), which integrates a Graph Neural Network (GNN) with a neural wave function to solve the Schrödinger equation for multiple geometries simultaneously. This approach achieves up to a 40x speedup in mapping potential energy surfaces while maintaining or exceeding the accuracy of state-of-the-art neural methods.

Introduction

In quantum chemistry, the Potential Energy Surface (PES) represents the ground-state energy of a molecule across various spatial arrangements of its nuclei. Accurate PES mapping is essential for predicting chemical reaction paths and material behaviors. High-accuracy neural methods like FermiNet have shown great promise but are limited by a "one-model-per-geometry" constraint, making the evaluation of hundreds of configurations prohibitively expensive. PESNet breaks this bottleneck by using amortized optimization to learn the wave function's dependence on geometry in a single training pass.

Methodology: The PESNet Architecture

The PESNet architecture is composed of two primary, differentiable components:

  1. MetaGNN: A Graph Neural Network that acts as a meta-learner. It takes nuclear coordinates and charges as input and predicts the optimal parameters ($\omega$) for the wave function model.
  2. WFModel: A neural wave function based on the FermiNet architecture. It evaluates the electronic wave function $\psi(\mathbf{r})$ for a given set of electron positions, utilizing the geometry-specific parameters provided by the MetaGNN.
Schematic of the PESNet architecture showing the MetaGNN parametrizing the WFModel for different molecular structures.

By making the entire pipeline differentiable, the model can be trained end-to-end via VMC, allowing it to "learn" how the wave function evolves as the molecular geometry changes.

Handling Physical Symmetries

A key technical innovation in PESNet is the maintenance of physical symmetries. While molecular energy is invariant to rotation and reflection, the wave function itself must be equivariant. To achieve this, the authors developed an equivariant coordinate system based on Principal Component Analysis (PCA) of the nuclei positions. This ensures that the model generalizes across symmetric configurations, effectively acting as a physics-informed data augmentation strategy that reduces the complexity of the required training domain.

Experimental Results

PESNet was benchmarked against several molecular systems, including hydrogen chains and complex molecules like Nitrogen ($N_2$).

  • Computational Efficiency: PESNet provides a speedup of up to 40 times compared to training separate models for high-resolution PES scans.
  • Accuracy and Generalization: In benchmarks such as the $H_{10}$ hydrogen chain, PESNet matched the accuracy of FermiNet and outperformed other neural methods like PauliNet and DeepErwin.
  • Complex Systems: For the Nitrogen molecule ($N_2$), which involves a complex triple bond, PESNet achieved results within ~0.37 m$E_h$ of the FermiNet baseline while requiring only a fraction of the total training time.
Performance of PESNet on the Nitrogen molecule ($N_2$). The plot illustrates that PESNet (blue) closely tracks the high-accuracy FermiNet (orange) across varying bond lengths and significantly outperforms a baseline without the MetaGNN component (green).

Conclusion

PESNet demonstrates that pairing GNNs with neural wave functions can overcome the primary computational hurdle in high-accuracy ab-initio quantum chemistry. By enabling the simultaneous solution of the Schrödinger equation for multiple geometries, it paves the way for efficient and accurate mapping of potential energy surfaces. Future work will explore scaling to larger systems and handling varying chemical species within a single network.

Summary updated

Abstract

Solving the Schr\"odinger equation is key to many quantum mechanical properties. However, an analytical solution is only tractable for single-electron systems. Recently, neural networks succeeded at modeling wave functions of many-electron systems. Together with the variational Monte-Carlo (VMC) framework, this led to solutions on par with the best known classical methods. Still, these neural methods require tremendous amounts of computational resources as one has to train a separate model for each molecular geometry. In this work, we combine a Graph Neural Network (GNN) with a neural wave function to simultaneously solve the Schr\"odinger equation for multiple geometries via VMC. This enables us to model continuous subsets of the potential energy surface with a single training pass. Compared to existing state-of-the-art networks, our Potential Energy Surface Network PESNet speeds up training for multiple geometries by up to 40 times while matching or surpassing their accuracy. This may open the path to accurate and orders of magnitude cheaper quantum mechanical calculations.
BibTeX
@inproceedings{gao2022ab,
  author = {Nicholas Gao and Stephan Günnemann},
  title = {{Ab-Initio Potential Energy Surfaces by Pairing GNNs with Neural Wave Functions}},
  booktitle = {International Conference on Learning Representations},
  year = {2022},
  url = {https://openreview.net/forum?id=ooKd0EVsjV},
}
Figure