Linear-scaling generation of potential energy surfaces using a double incremental expansion

Carolin König, Ove ChristiansenThe Journal of Chemical Physics 2016 · DOI · arXiv

Introduction

Generating an accurate Potential Energy Surface (PES) is a fundamental requirement for simulating vibrational spectra and chemical reaction dynamics. However, the computational cost of constructing a PES for large molecules is traditionally prohibitive due to two compounding factors: the "dimensionality curse," where the number of required single-point energy calculations (SPCs) grows exponentially with the number of atoms, and the "super-linear scaling" of the electronic structure methods themselves.

While $n$-mode expansion techniques effectively mitigate the dimensionality problem by truncating high-order vibrational couplings, the cost of each individual SPC still increases dramatically as the system size grows. This paper introduces a "double incremental" framework that addresses both challenges simultaneously, achieving true linear scaling ($O(N)$) in the total computational effort required for PES generation.

The Double Incremental Framework

The proposed approach combines two distinct approximation strategies to decouple the system size from the computational cost:

  1. $n$-mode Expansion (Vibrational): This standard method approximates the full PES as a sum of lower-order terms (1-body, 2-body, and 3-body couplings) between vibrational modes, significantly reducing the total number of points that need to be sampled.
  2. Many-Body Expansion (Electronic): This fragmentation technique calculates the total electronic energy by summing the energies of small molecular fragments and their combinations (dimers, trimers, etc.), rather than treating the entire molecule as a single unit.

The Role of FALCON Coordinates

The synergy between these two expansions is made possible by Flexible Adaptation of Local COordinates of Nuclei (FALCON). Standard normal modes are typically delocalized across the entire molecule, which would force every fragment calculation to account for every mode, negating the benefits of fragmentation. FALCON coordinates, however, are "semi-local." Most coordinates are strictly confined to a single fragment or a small group of neighboring fragments, allowing the vibrational and electronic expansions to work in tandem.

DIF and DIFACT

The authors present two iterations of this approach:

  • DIF (Double Incremental expansion in FALCON): This model uses FALCON coordinates to restrict the number of modes considered in each fragment's PES.
  • DIFACT (DIF with Auxiliary Coordinate Transformation): This is the most advanced variant. It introduces local "auxiliary" coordinates for the modes that bridge different fragments. By transforming these back to a global set, DIFACT ensures that the number of modes involved in any single fragment calculation remains constant, regardless of the total system size. This decoupling is the key to achieving true linear scaling.

Evaluation and Results

The methodology was tested on oligo-phenyl chains (specifically tetra-phenyl and hexa-phenyl), which serve as an ideal testbed for evaluating fragmentation and locality. Electronic structure calculations were performed using HF-3c and RI-BP86/SVP methods.

Computational Efficiency

The primary achievement of the DIFACT model is its linear scaling. The authors demonstrated that the accumulated cost of SPCs grows linearly with the number of fragments. For large, covalently bound systems, this represents a significant reduction in computational overhead, allowing for the use of more accurate electronic structure methods on systems that were previously computationally inaccessible.

Accuracy and Convergence

Despite the approximations, the method maintains high precision. For the tetra-phenyl system:

  • The DIF-3F2M model (3-fragment level, 2-mode coupling) yielded fundamental vibrational frequencies (VSCF) with a maximum deviation of only 0.05 cm⁻¹ compared to the reference.
  • The DIFACT models, which introduce a coordinate transformation step to achieve linear scaling, showed slightly higher but still modest errors, typically within a 5 cm⁻¹ range. This transformation error is notably smaller than the errors typically introduced by the underlying electronic structure methods themselves.

Limitations and Future Work

While the results for linear chains are promising, the authors identify several areas for further development:

  • System Geometry: The current test cases are chain-like. The performance on highly branched or 3D globular structures, such as proteins, remains to be seen, as the number of neighboring fragment interactions would increase.
  • Transformation Errors: While small, the errors in DIFACT are present. The authors suggest that using higher-order mode couplings during the transformation could further refine accuracy.
  • Property Surfaces: To simulate actual infrared (IR) spectra, the double incremental approach must be extended from energy surfaces to dipole moment surfaces.

Conclusion

The double incremental expansion framework, particularly the DIFACT model, represents a major step toward the applicability of vibrational wave function methods to sizable, covalently bound systems. By leveraging the locality of FALCON coordinates, the authors have successfully decoupled the cost of individual energy calculations from the total system size, enabling high-level spectroscopic simulations for a broader range of molecular systems.

Summary updated

Abstract

We present a combination of the incremental expansion of potential energy surfaces (PESs), known as n-mode expansion, with the incremental evaluation of the electronic energy in a many-body approach. The application of semi-local coordinates in this context allows the generation of PESs in a very cost-efficient way. For this, we employ the recently introduced flexible adaptation of local coordinates of nuclei (FALCON) coordinates. By introducing an additional transformation step, concerning only a fraction of the vibrational degrees of freedom, we can achieve linear scaling of the accumulated cost of the single point calculations required in the PES generation. Numerical examples of these double incremental approaches for oligo-phenyl examples show fast convergence with respect to the maximum number of simultaneously treated fragments and only a modest error introduced by the additional transformation step. The approach, presented here, represents a major step towards the applicability of vibrational wave function methods to sizable, covalently bound systems.
BibTeX
@article{konig2016linear,
  author = {Carolin König and Ove Christiansen},
  title = {{Linear-scaling generation of potential energy surfaces using a double incremental expansion}},
  journal = {The Journal of Chemical Physics},
  year = {2016},
  doi = {10.1063/1.4960189},
}
Figure