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

Nicholas Gao, Stephan Günnemann

TL;DR

PESNet is a novel deep learning architecture designed to solve the Schrödinger equation for multiple molecular geometries simultaneously. By utilizing a Graph Neural Network (GNN) to predict the parameters of a neural wave function, the authors achieve up to a 40x speedup in mapping potential energy surfaces while maintaining state-of-the-art accuracy.

The Bottleneck in Quantum Chemistry

In quantum chemistry, the "holy grail" is solving the Schrödinger equation to find the ground-state energy of a molecule. Recently, neural wave functions (like FermiNet) have achieved incredible accuracy using Variational Monte Carlo (VMC) methods. However, these models suffer from a massive computational bottleneck: they are "per-geometry" specialists.

If you move a single atom by a fraction of an Angstrom, you must retrain the entire neural network from scratch to find the new energy. This makes mapping a Potential Energy Surface (PES)—the multidimensional map of how a molecule's energy changes with its shape—prohibitively expensive. Researchers needed a way to "train once" and then evaluate any geometry on that surface instantly.

Methods: Introducing PESNet

The authors introduce PESNet (Potential Energy Surface Network), which overcomes this limitation using two main components:

  1. MetaGNN: A Graph Neural Network that acts as a "parameter generator." It takes the coordinates and charges of the atomic nuclei as input and outputs the weights and biases ($\omega$) for the wave function model.
  2. WFModel (Wave Function Model): A neural wave function (based on the FermiNet architecture) that takes electron positions as input and calculates the probability amplitude (the wave function $\psi$).
High-level architecture of PESNet

Crucially, the authors developed a unique equivariant coordinate system using Principal Component Analysis (PCA). While energy is invariant to rotation (rotating a molecule doesn't change its energy), the wave function itself is equivariant (it must rotate with the molecule). By defining a stable, molecule-fixed coordinate frame, PESNet can handle these physical symmetries without losing the ability to represent complex, non-spherical electronic states.

The model is trained ab-initio (from first principles) using VMC. It does not require a pre-existing dataset of energies; instead, it learns by minimizing the energy calculated directly from the Schrödinger equation.

Key Insights and Innovations

  • Amortized Optimization: Instead of optimizing a wave function for one specific geometry, PESNet "amortizes" the cost by learning a function (the MetaGNN) that maps any geometry to its optimal wave function parameters.
  • Continuous Learning: During training, the authors use "random walkers" to slightly perturb the molecular geometries. This allows the model to learn a continuous, smooth potential energy surface rather than just memorizing a few discrete points.
  • Symmetry is Tricky: The paper highlights that standard GNNs (which are usually invariant) are insufficient for wave functions. You cannot represent excited states or certain directional bonds if your model can only "see" distances. Their equivariant PCA-based coordinate system is the "secret sauce" that resolves this.

Key Results

PESNet was benchmarked on several challenging molecular systems, including the $H_4^+$ transition path, the Hydrogen chain ($H_{10}$), the Nitrogen molecule ($N_2$), and Cyclobutadiene. The results demonstrate significant improvements in both efficiency and accuracy:

  • Massive Speedup: For high-resolution scans of potential energy surfaces (such as the 39 geometries of Nitrogen tested), PESNet was up to 40.7 times faster than training individual FermiNet models.
  • State-of-the-Art Accuracy: Despite the massive speedup, PESNet matched or even surpassed the accuracy of per-geometry models. In the $H_4^+$ system, it estimated energies $\approx 0.27$ m$E_h$ lower (more accurate) than previous weight-sharing attempts like DeepErwin.
  • Generalization on Failure Cases: The model successfully captured the "fake minimum" problem in the Hydrogen rectangle ($H_4$). It correctly identified the local maximum at 90°, a scenario where traditional classical methods like CCSD(T) often fail and predict an artificial dip.
Energy surface of the H4 rectangle

Limitations and Future Work

While a significant leap forward, PESNet has a few current limitations:

  • Fixed Nuclei/Electrons: A single PESNet model is currently limited to a specific set of nuclei and a fixed number of electrons. You cannot train one model on Methane and expect it to work on Ethane; the MetaGNN must be retrained for the new "class" of molecule.
  • Initialization: The authors note that PauliNet still offers better initialization (using Hartree-Fock baselines), which helps it converge in fewer iterations. Integrating better physical priors into the MetaGNN's starting state is a clear next step.
  • Architecture Search: The GNN component has not been heavily optimized yet. More advanced GNN architectures could further improve performance and generalization in future iterations.

Summary updated

Abstract

Solving the Schrödinger equation is key to many quantum mechanical properties. However, an analytical solution is only tractable for single-electron systems. Recently, neural networks succeeded at modelling 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 Schrödinger 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