GPR_calculator: An on-the-fly surrogate model to accelerate massive nudged elastic band calculations

Isaac Onyango, Byungkyun Kang, Qiang ZhuComput. Phys. Commun 2025 · DOI

TL;DR

This paper introduces GPR_calculator, a Python/C++ package that uses Gaussian Process Regression (GPR) to create an "on-the-fly" surrogate model for expensive electronic structure calculations. By dynamically training and updating the GPR model during simulations, it significantly accelerates Nudged Elastic Band (NEB) calculations for surface diffusion and reactions, achieving 3-10x speedups while maintaining accuracy.

What Problem Does This Paper Address?

Simulating chemical reactions on surfaces, particularly finding the transition states (saddle points) that dictate reaction rates, is crucial for designing catalysts. The Nudged Elastic Band (NEB) method is a standard tool for this, but it requires hundreds or thousands of energy and force calculations using computationally expensive electronic structure methods like Density Functional Theory (DFT). This makes NEB simulations for complex systems or large numbers of configurations prohibitively time-consuming, often taking days. While surrogate models (like neural networks or GPR) have been used to speed this up, they often require a separate, upfront training phase and lack dynamic adaptation during the simulation, potentially limiting their accuracy for exploratory tasks.

Methods

The core innovation is an "on-the-fly" surrogate model approach. GPR_calculator integrates a Gaussian Process Regression (GPR) model directly into the NEB simulation workflow. Here's how it works:

  • Hybrid Calculator: It uses a base electronic structure calculator (e.g., DFT) as the "ground truth" and a GPR surrogate model as a fast approximation.
  • Dynamic Training: During the NEB simulation, the GPR model predicts energies and forces. If the GPR's uncertainty in its prediction exceeds a user-defined threshold, the expensive base calculator is called to get the accurate value.
  • Model Update: This new, accurate data point is then used to update and retrain the GPR model, improving its accuracy for future predictions.
  • Structural Descriptors: The GPR model uses a "SO(3) descriptor" derived from the power spectrum of spherical harmonic expansion coefficients to represent the local atomic environment. This descriptor is rotationally invariant and captures the chemical environment around each atom.
  • Kernel Functions: Both Radial Basis Function (RBF) and Dot Product kernels are implemented to define the similarity between atomic environments and predict energies and forces.
  • Optimization: The package includes optimizations for computational cost, such as rewriting kernel computations in C++ and using MPI for parallelization.

Datasets

The paper does not explicitly mention training on specific pre-existing datasets. Instead, it focuses on an "on-the-fly" training strategy where the GPR model is trained dynamically during the NEB simulations using data generated by the base electronic structure calculator (DFT). The paper presents benchmark calculations on systems like Au diffusion on Al(100), Pd4 cluster diffusion on MgO(100), and H2S dissociation on Pd(100) and other transition metal surfaces. These serve as evaluation benchmarks for the GPR_calculator.

Key Insights and Ideas

  • Dynamic Adaptation is Key: The "on-the-fly" nature of the GPR training is crucial. Unlike static surrogate models, this dynamic updating allows the model to adapt to the specific reaction pathway being explored, ensuring accuracy where it's needed most without requiring a massive upfront dataset.
  • Uncertainty-Driven Refinement: Using uncertainty quantification from the GPR to decide when to call the expensive DFT calculator is an efficient strategy. It avoids unnecessary DFT calls for regions of the potential energy surface that the GPR can predict reliably.
  • Balancing Accuracy and Cost: The approach offers a tunable trade-off between computational cost and accuracy. Users can adjust the uncertainty thresholds to achieve desired speedups while maintaining acceptable error margins.
  • Serendipitous Discoveries: The inherent "noise" in GPR predictions can sometimes prevent convergence to shallow local minima, potentially helping to discover more physically meaningful pathways, as demonstrated in the H2S dissociation example.

Key Results

  • Significant Speedup: The GPR_calculator achieved 3-10 times acceleration for NEB simulations compared to pure ab initio (DFT) calculations.
  • Accuracy Maintained: The GPR-optimized NEB paths and energy barriers were found to be nearly identical to those obtained from pure DFT calculations, with energy differences typically within 0.05 eV.
  • Reduced DFT Calls: For example, Au diffusion on Al(100) required only 8 DFT calls with GPR, compared to 53 for pure EMT (a less expensive DFT-like method).
  • Practical Application: The package was successfully applied to model surface diffusion (Au on Al(100), Pd4 on MgO(100)) and chemical reactions (H2S dissociation on Pd(100) and other surfaces), demonstrating its versatility.

Limitations and Future Work

  • Convergence Iterations: Surrogate models, including this GPR approach, may sometimes require more iterations to converge compared to direct DFT calculations, especially as the system approaches the minimum energy pathway where high force accuracy is critical.
  • Scalability: While optimizations are implemented, GPR's inherent O(N³) complexity with dataset size (N) can still be a challenge for extremely large datasets.
  • Future Work: The authors plan to focus on improving the efficiency of the GPR model, particularly its scaling with larger datasets, and enhancing prediction accuracy. They also aim to make the framework more extensible by allowing users to easily add new calculators, NEB optimization algorithms, and GPR functionalities (e.g., new descriptors or kernel choices).

Summary updated

BibTeX
@article{onyango2025gpr,
  author = {Isaac Onyango and Byungkyun Kang and Qiang Zhu},
  title = {{GPR\_calculator: An on-the-fly surrogate model to accelerate massive nudged elastic band calculations}},
  journal = {Comput. Phys. Commun},
  year = {2025},
  doi = {10.1016/j.cpc.2025.109781},
}
Figure