The energy landscape of silicon systems and its description by force fields, tight binding schemes, density functional methods and Quantum Monte Carlo methods

S. Alireza Ghasemi, Maximilian Amsler, Richard G. Hennig, Shantanu Roy, Stefan Goedecker, C. J. Umrigar, Luigi Genovese, Thomas J. Lenosky, Tetsuya Morishita, Kengo Nishio

Abstract

This study evaluates the accuracy of silicon energy landscapes generated by various computational methods, including density functional theory (DFT), tight-binding schemes, and classical force fields. Using Quantum Monte Carlo (QMC) results as a quasi-exact reference, the researchers found that while DFT provides a robust description of both ground states and transition states, popular force fields often produce "fake" landscapes characterized by excessive ruggedness and non-existent local minima. These findings suggest that force-field-based simulations of silicon dynamics should be interpreted with caution.

Introduction: The Energy Landscape Problem

In materials science, understanding how a system evolves—such as during crystal growth or atomic diffusion—requires mapping its "potential energy surface" or energy landscape. This multidimensional map consists of valleys (stable local minima) and mountain passes (saddle points). Atoms must traverse these saddle points to move from one stable configuration to another.

While high-precision methods like Quantum Monte Carlo (QMC) can solve the Schrödinger equation with great accuracy, they are computationally prohibitive for large systems. Consequently, researchers often rely on faster approximations: Density Functional Theory (DFT) for medium-sized systems and empirical Force Fields (FF) for large-scale simulations. However, because force fields are typically fitted to ground-state data, their ability to accurately predict the transition states that govern dynamics has remained a critical question.

A transition path for the Si8 cluster showing two local minima connected by a saddle point.

Methodology

The researchers, led by S. Alireza Ghasemi, employed a multi-tiered benchmarking approach. They used Diffusion Monte Carlo (DMC) as the "gold standard" reference. This was compared against several DFT functionals (LDA, PBE, and B3LYP) and four widely used silicon force fields: Tersoff, Stillinger-Weber, EDIP, and Lenosky. Additionally, the Lenosky Tight-Binding (LTB) scheme was evaluated as a middle-ground approach. To explore these landscapes efficiently, they utilized the Minima Hopping Method (MHM), which systematically identifies both global minima and the transition states connecting them.

Benchmarking the Si8 Cluster

The study focused heavily on small $Si_8$ clusters to allow for direct comparison with expensive QMC calculations. By identifying various saddle points, the team could measure the accuracy of barrier heights—the energy required to move between states.

A collection of eight distinct saddle points (SP1-SP8) identified in the Si8 energy landscape.

Surprisingly, DFT performed remarkably well for these transition states. While DFT often struggles with "self-interaction errors" in chemical reactions, the researchers found that in silicon systems, the local environment of an atom in a transition state is sufficiently similar to its ground state that these errors largely cancel out. This makes DFT a reliable tool for studying silicon dynamics.

The "Fake" Landscape of Force Fields

In contrast, classical force fields often failed to capture the true topology of the energy landscape. Beyond simple numerical discrepancies in energy levels, force fields frequently predicted "fake" local minima and saddle points that do not exist in the quantum mechanical reality.

The study also quantified the "ruggedness" of these models. For instance, the Tersoff force field was found to create a highly rugged, "glassy" landscape, making it difficult for optimization algorithms to find the true ground state. In contrast, the Lenosky force field provided a much smoother landscape, acting as a "structure seeker." However, even the Lenosky model had flaws, such as incorrectly predicting the lowest-energy excitation in bulk silicon as a "double defect" rather than a single one.

Conclusion

The Lenosky Tight-Binding (LTB) method emerged as a successful compromise, correctly ordering the stability of different isomers where classical force fields failed. However, the overall results serve as a significant warning: silicon simulations based solely on a single force field may be misleading. The researchers recommend that any critical findings involving complex defects or transition states be verified with higher-level DFT calculations to ensure the predicted physical behavior is not merely an artifact of a "fake" energy landscape.

Summary updated

Abstract

The accuracy of the energy landscape of silicon systems obtained from various density functional methods, a tight binding scheme and force fields is studied. Quantum Monte Carlo results serve as quasi exact reference values. In addition to the well known accuracy of DFT methods for geometric ground states and metastable configurations we find that DFT methods give a similar accuracy for transition states and thus a good overall description of the energy landscape. On the other hand, force fields give a very poor description of the landscape that are in most cases too rugged and contain many fake local minima and saddle points or ones that have the wrong height.
BibTeX
@inproceedings{ghasemi2010energy,
  author = {S. Alireza Ghasemi and Maximilian Amsler and Richard G. Hennig and Shantanu Roy and Stefan Goedecker and C. J. Umrigar and Luigi Genovese and Thomas J. Lenosky and Tetsuya Morishita and Kengo Nishio},
  title = {{The energy landscape of silicon systems and its description by force fields, tight binding schemes, density functional methods and Quantum Monte Carlo methods}},
  booktitle = {Physical},
  year = {2010},
  doi = {10.1103/physrevb.81.214107},
}
Figure