Comparing the accuracy of high-dimensional neural network potentials and the systematic molecular fragmentation method: A benchmark study for all-trans alkanes

Michael Gastegger, Clemens Kauffmann, Jörg Behler, Philipp MarquetandThe Journal of Chemical Physics 2016 · DOI · arXiv

Abstract

Many approaches, which have been developed to express the potential energy of large systems, exploit the locality of the atomic interactions. A prominent example is the fragmentation methods in which the quantum chemical calculations are carried out for overlapping small fragments of a given molecule that are then combined in a second step to yield the system's total energy. Here we compare the accuracy of the systematic molecular fragmentation approach with the performance of high-dimensional neural network (HDNN) potentials introduced by Behler and Parrinello. HDNN potentials are similar in spirit to the fragmentation approach in that the total energy is constructed as a sum of environment-dependent atomic energies, which are derived indirectly from electronic structure calculations. As a benchmark set, we use all-trans alkanes containing up to eleven carbon atoms at the coupled cluster level of theory. These molecules have been chosen because they allow to extrapolate reliable reference energies for very long chains, enabling an assessment of the energies obtained by both methods for alkanes including up to 10,000 carbon atoms. We find that both methods predict high-quality energies with the HDNN potentials yielding smaller errors with respect to the coupled cluster reference.

Introduction

In computational chemistry, a fundamental challenge is the trade-off between accuracy and system size. High-level ab initio methods like CCSD(T) (Coupled Cluster with Single, Double, and perturbative Triple excitations) provide "gold standard" accuracy but scale poorly with the number of atoms, limiting their use to small molecules. To address this, researchers leverage chemical locality—the principle that an atom's energy contribution is primarily determined by its immediate local environment.

Two prominent locality-based strategies are Systematic Molecular Fragmentation (SMF) and High-Dimensional Neural Network (HDNN) potentials. SMF physically divides a molecule into overlapping fragments, while HDNNs use machine learning to map local environments to atomic energies. This study provides a rigorous comparison of these methods using all-trans alkanes as a benchmark.

Methodology

The study evaluated two primary techniques against CCSD(T)-F12 reference calculations for alkanes ranging from $C_3$ to $C_{11}$.

Systematic Molecular Fragmentation (SMF): This method breaks the molecule into functional groups (CH₂ and CH₃). The "level" of fragmentation dictates the size of overlapping regions; higher levels increase accuracy but also computational cost.

High-Dimensional Neural Networks (HDNN): Based on the Behler-Parrinello architecture, the total energy is represented as a sum of atomic energies. Local environments are described by Atom-Centered Symmetry Functions (ACSFs), which ensure invariance to rotation and translation.

Schematic of the High-Dimensional Neural Network (HDNN) architecture.

The benchmark set included small alkanes and 474 non-optimized fragments. To test scalability, the authors extrapolated reference energies for massive chains of up to 10,000 carbon atoms, exploiting the linear scaling of error in these regular systems.

Key Results

The comparison revealed that HDNN potentials consistently outperform SMF in accuracy across all tested chain lengths.

  1. Superior Precision: The HDNN achieved a Root Mean Square Error (RMSE) of approximately 0.0006 to 0.0012 kcal/mol per atom, significantly lower than SMF even at high fragmentation levels (e.g., level 7).
  2. Stability at Extreme Scale: As shown in the comparison for long chains, the HDNN predictions remain remarkably stable even when extrapolated to 10,000 carbon atoms. While SMF errors grow predictably with chain length, the HDNN maintains a much tighter grip on the reference energy.
Comparison of energy errors for fragmentation levels and HDNN potentials across long carbon chains.
  1. Efficiency: Once trained, the HDNN can evaluate the energy of massive systems in a fraction of the time required for fragmentation-based ab initio calculations.

Conclusion

This study confirms that HDNNs, trained only on small fragments, can accurately predict the energy of molecules three orders of magnitude larger without losing physical fidelity. While all-trans alkanes represent a geometrically regular system, the success of the HDNN suggests it is a highly effective tool for exploiting chemical locality. Future work will focus on more complex organic molecules and systems where long-range interactions, such as electrostatics and dispersion, play a more significant role.

Summary updated

Abstract

Many approaches, which have been developed to express the potential energy of large systems, exploit the locality of the atomic interactions. A prominent example is the fragmentation methods in which the quantum chemical calculations are carried out for overlapping small fragments of a given molecule that are then combined in a second step to yield the system's total energy. Here we compare the accuracy of the systematic molecular fragmentation approach with the performance of high-dimensional neural network (HDNN) potentials introduced by Behler and Parrinello. HDNN potentials are similar in spirit to the fragmentation approach in that the total energy is constructed as a sum of environment-dependent atomic energies, which are derived indirectly from electronic structure calculations. As a benchmark set, we use all-trans alkanes containing up to eleven carbon atoms at the coupled cluster level of theory. These molecules have been chosen because they allow to extrapolate reliable reference energies for very long chains, enabling an assessment of the energies obtained by both methods for alkanes including up to 10 000 carbon atoms. We find that both methods predict high-quality energies with the HDNN potentials yielding smaller errors with respect to the coupled cluster reference.
BibTeX
@article{gastegger2016comparing,
  author = {Michael Gastegger and Clemens Kauffmann and Jörg Behler and Philipp Marquetand},
  title = {{Comparing the accuracy of high-dimensional neural network potentials and the systematic molecular fragmentation method: A benchmark study for all-trans alkanes}},
  journal = {The Journal of Chemical Physics},
  year = {2016},
  doi = {10.1063/1.4950815},
}
Figure