Greater Transferability and Accuracy of Norm-conserving Pseudopotentials using Nonlinear Core Corrections

Wan‐Lu Li, Kaixuan Chen, Elliot Rossomme, Martin Head‐Gordon, Teresa Head‐Gordon

TL;DR

This paper demonstrates that adding Nonlinear Core Corrections (NLCC) to standard norm-conserving pseudopotentials significantly improves the accuracy of Density Functional Theory (DFT) calculations, particularly for atomization energies and chemical reactivity. While NLCCs derived from the PBE functional transfer well to hybrid functionals, the authors show that meta-GGA functionals require specific re-optimization to avoid massive errors, providing a new protocol for high-accuracy condensed phase simulations.

What Problem Does This Paper Address?

In computational chemistry, "pseudopotentials" (PPs) are a standard trick to save time. Instead of calculating the behavior of every single electron in an atom, researchers treat the inner "core" electrons and the nucleus as a single static block, focusing only on the "valence" electrons that actually participate in chemical bonding.

However, standard PPs (like the popular GTH variety) often assume a linear relationship between the core and valence electron densities. In reality, the interaction between these two regions is highly nonlinear due to strong electron-electron interactions near the nucleus. This simplification leads to systematic errors, especially in spin-polarized systems or when calculating the energy required to break covalent bonds (atomization energies). The authors aim to bridge the gap between the efficiency of pseudopotentials and the high accuracy of "all-electron" calculations.

Methods

The researchers utilized the Goedecker, Teter, and Hutter (GTH) protocol for norm-conserving pseudopotentials. To fix the linearity issue, they incorporated Nonlinear Core Corrections (NLCC), where the core charge density is represented by a single Gaussian function.

Key aspects of their approach include:

  • Optimization: They used the ATOM module in the CP2K software package to fit NLCC parameters ($c_{core}$ and $r_{core}$) against all-electron reference data.
  • Multi-state Fitting: Unlike traditional PPs fitted only to ground states, these were optimized against ground, excited, and ionized electronic configurations to ensure better "transferability" (the ability of a PP to work in different chemical environments).
  • Functional Testing: They tested how well an NLCC optimized for one functional (PBE) performed when used with others, such as PBE0 (hybrid), $\omega$B97X-V (range-separated hybrid), and B97M-rV (meta-GGA).

Datasets

The study evaluated the PPs across a massive range of chemical properties using established benchmarks:

  • Thermochemistry (TC): G2 and TAE140 datasets (evaluating covalent bond formation and atomization energies).
  • Barrier Heights (BH): HTBH38, NHTBH38, and WCPT27 (hydrogen and non-hydrogen transfer reactions).
  • Isomerization and Potentials (I): ISOMERIZATION20, G21EA (electron affinities), and G21IP (ionization potentials).
  • Non-covalent Interactions (NC): S66 (organic molecules), CT20 (charge-transfer complexes), and NCED (easy dimers).

Key Insights and Ideas

  • Atomization is the "Aha!" Moment: The most dramatic improvements were seen in atomization energies. Standard PPs tend to underestimate covalent bond strengths; NLCC effectively "rectifies" this, bringing errors down from ~10% to less than 1%.
  • The Meta-GGA Trap: A major finding was that PPs are not universally transferable. While a PBE-optimized NLCC works great for hybrid functionals, it fails spectacularly for the meta-GGA functional B97M-rV, yielding errors over 50 kcal/mol.
  • Why Meta-GGAs Struggle: Meta-GGAs are sensitive to kinetic energy density. Current implementations often neglect the core's contribution to this density, leading to "Pseudopotential Inconsistency Errors" (PPIEs).
  • MOLOPT Synergy: The authors show that NLCCs work exceptionally well with medium-sized, molecularly optimized (MOLOPT) basis sets. This is a big win for researchers running Ab Initio Molecular Dynamics (AIMD), as it allows for near-all-electron accuracy at a fraction of the computational cost.

Key Results

  • PBE Accuracy: For the PBE functional, adding NLCC reduced the Mean Absolute Deviation (MAD) for atomization energies from roughly 10–15 kcal/mol to just 1–2 kcal/mol.
  • Successful Transfer: The PBE-derived NLCC transferred seamlessly to PBE0 and $\omega$B97X-V, maintaining high accuracy without needing re-optimization.
  • B97M-rV Fix: By re-optimizing the NLCC parameters specifically for the B97M-rV functional (focusing on lighter elements like B, C, N, O, and F), the authors reduced the massive 50+ kcal/mol error to a manageable ~6 kcal/mol.
  • Non-thermal Properties: For non-covalent interactions and barrier heights, the impact of NLCC was smaller but generally positive or neutral, suggesting that the "chemical environment" in these cases doesn't change enough to trigger the nonlinear errors seen in atomization.

Limitations and Future Work

  • Electron Affinities: Even with NLCC, electron affinities remained sensitive to the size of the basis set, indicating that NLCC isn't a "silver bullet" for all properties.
  • Kinetic Energy Density: The authors note that a proper correction for the kinetic energy density of core orbitals is still missing in many meta-GGA implementations. Developing this could further improve accuracy for complex functionals.
  • Heavy Elements: The re-optimization for meta-GGAs was primarily focused on lighter main-group elements; extending this to heavier elements and transition metals remains a future challenge.

Summary updated

Abstract

We present an investigation into the transferability of pseudopotentials (PPs) with a nonlinear core correction (NLCC) using the Goedecker, Teter, and Hutter (GTH) protocol across a range of pure GGA, meta-GGA and hybrid functionals, and their impact on thermochemical and non-thermochemical properties. The GTH-NLCC PP for the PBE density functional demonstrates remarkable transferability to the PBE0 and $ω$B97X-V exchange-correlation functionals, and relative to no NLCC, improves agreement significantly for thermochemical benchmarks compared to all-electron calculations. On the other hand, the B97M-rV meta-GGA functional performs poorly with the PBE-derived GTH-NLCC PP, which is mitigated by reoptimizing the NLCC parameters for this specific functional. The findings reveal that atomization energies exhibit the greatest improvements from use of the NLCC, which thus provides an important correction needed for covalent interactions relevant to applications involving chemical reactivity. Finally we test the NLCC-GTH PPs when combined with medium-size TZV2P molecularly optimized (MOLOPT) basis sets which are typically utilized in condensed phase simulations, and show that they lead to consistently good results when compared to all-electron calculations for atomization energies, ionization potentials, barrier heights, and non-covalent interactions, but lead to somewhat larger errors for electron affinities.
BibTeX
@inproceedings{li2023greater,
  author = {Wan‐Lu Li and Kaixuan Chen and Elliot Rossomme and Martin Head‐Gordon and Teresa Head‐Gordon},
  title = {{Greater Transferability and Accuracy of Norm-conserving Pseudopotentials using Nonlinear Core Corrections}},
  booktitle = {Chemical},
  year = {2023},
  doi = {10.1039/d3sc03709f},
}
Figure