Performance Improvement of Deorbitalized Exchange-Correlation Functionals

1. Problem Statement

Density Functional Theory (DFT) relies on exchange-correlation (XC) functionals to balance computational cost and accuracy. Meta-Generalized Gradient Approximations (meta-GGAs), such as r2SCAN, offer high accuracy by incorporating the Kohn-Sham kinetic energy density ($\tau_S$), which depends explicitly on orbital information. However, this orbital dependence necessitates the use of Generalized Kohn-Sham (gKS) equations with non-local potentials, leading to:

• High computational cost: Slower convergence and more expensive per-cycle calculations compared to pure Kohn-Sham (KS) local potentials.
• Implementation difficulties: Challenges in developing Projector-Augmented Wave (PAW) potentials and sensitivity to grid density.

Deorbitalization is a strategy to replace the orbital-dependent $\tau_S$ with an orbital-free Kinetic Energy Density (KED) functional dependent only on the electron density ($n$), its gradient ($\nabla n$), and its Laplacian ($\nabla^2 n$). This restores a pure local KS potential, theoretically enabling faster calculations.

The Core Challenge: While deorbitalization simplifies the potential structure, existing methods (specifically those based on the Laplacian) often introduce numerical instabilities. The dependence on the density Laplacian ($\nabla^2 n$) leads to:

• "Roughness" or spiky oscillations in the XC potential due to high-order spatial derivatives.
• Poor Self-Consistent Field (SCF) convergence, requiring many more iterations than the parent meta-GGA, which negates the per-cycle speed gains.
• Violation of physical constraints (e.g., non-negativity of kinetic energy) in certain limits.

2. Methodology

The authors propose a new deorbitalization strategy that prioritizes potential smoothness while maintaining constraint satisfaction.

• Theoretical Framework:
  • They replace the exact $\tau_S$ in the r2SCAN functional with an approximate orbital-free KED ($\tau_{OF}$) of the form $\tau(n, \nabla n, \nabla^2 n)$.
  • The KED is constructed using the Pauli enhancement factor ($\alpha$), which transitions between the slowly-varying electron gas limit and the iso-orbital limit (von Weizsäcker).
• Model Construction:
  • The authors build upon the RPP (r2SCAN Piecewise Polynomial) deorbitalizer, which satisfies important constraints (like the fourth-order gradient expansion) but suffers from potential roughness due to its switching function.
  • They introduce a Smoothed RPP (SRPP) and a variant SRPP2. These replace the non-analytic, piecewise polynomial switching function of RPP with the smoother switching function derived from the Cancio-Redd (CR) model.
  • Key Innovation: The CR switching function provides a smoother transition between the slowly-varying and iso-orbital limits, reducing the abrupt changes that cause potential oscillations.
• Quantification of Noise:
  • To objectively measure the "roughness" of the potentials, the authors define an action measure ($I_{deorb}$) based on the Laplacian term of the kinetic potential. Lower $I_{deorb}$ values indicate smoother potentials.
• Computational Setup:
  • Codes: NWChem (molecules, Gaussian basis sets) and VASP (solids, plane-wave PAW).
  • Test Sets: G3X/99 (223 molecules), T96-R (bond lengths), T82-F (frequencies), and a 55-solid set (lattice constants, cohesive energies, bulk moduli, band gaps).
  • Dynamics: Short Ab Initio Molecular Dynamics (AIMD) simulations of liquid Aluminum.

3. Key Contributions

1. Development of SRPP/SRPP2: Creation of new deorbitalized functionals that retain the constraint compliance of the RPP model but significantly improve the smoothness of the XC potential by adopting the CR switching mechanism.
2. Noise Quantification: Introduction of the I_{deorb metric to correlate the mathematical smoothness of the switching function with the numerical stability of the resulting potential.
3. Comprehensive Benchmarking: A systematic comparison of timing, accuracy, and SCF convergence across molecules, solids, and AIMD, contrasting the new smooth functionals against the parent r2SCAN, the RPP deorbitalizer, and the PCopt deorbitalizer.

4. Results

A. Potential Smoothness and Atomic Tests

• Smoothness: The SRPP and SRPP2 models exhibit significantly lower I_{deorb values (orders of magnitude lower than PC and RPP), indicating much smoother potentials.
• Visual Evidence: Plots of the Pauli potential and exchange potential for H and Si atoms show that SRPP/SRPP2 eliminate the large, unphysical oscillations (spikes) seen in PCopt and RPP, particularly in the transition regions between iso-orbital and slowly-varying densities.
• Constraint Compliance: The new functionals maintain the correct slowly-varying limit behavior (fourth-order gradient expansion) and non-negativity constraints, which PCopt violates.

B. Structural Accuracy

• Solids: SRPP and SRPP2 outperform RPP and PCopt for solid-state properties (lattice constants, cohesive energies, bulk moduli). They achieve Mean Absolute Deviations (MADs) comparable to or better than the parent r2SCAN, with reduced error spreads (fewer outliers).
• Molecules: For molecular properties (heats of formation, bond lengths), the deorbitalized forms (SRPP/RPP) generally show higher errors than the parent r2SCAN and the PCopt deorbitalizer. However, SRPP represents a 40% improvement in heat of formation MAD over RPP.
• Trade-off: The results suggest that strict adherence to constraints (RPP/SRPP) improves solid-state performance but slightly degrades molecular accuracy compared to the empirically fitted PCopt.

C. Timing and Computational Performance

• Static Solids (VASP):
  • Major Speedup: SRPP and SRPP2 are ~2x faster in total compute time than the parent r2SCAN.
  • Mechanism: While the time per SCF cycle for deorbitalized functionals is roughly 3x faster than r2SCAN (due to the local potential), the number of cycles required is higher. However, the reduction in cycle time outweighs the increase in cycle count.
  • Comparison: SRPP2 is the fastest, requiring only ~75% of the time of r2SCAN and ~72% more time than PBE.
• Molecules (NWChem):
  • Deorbitalization offers no speed advantage for molecules. The time per cycle is similar, but the number of SCF cycles increases significantly for deorbitalized forms, negating gains.
  • Basis Set Effect: The smoothness of Gaussian basis sets appears to mask the Laplacian-induced noise, making the performance of deorbitalized forms similar to the parent functional in terms of cycle counts.
• AIMD (Liquid Aluminum):
  • Failure Case: In AIMD simulations, deorbitalized functionals (SRPP/RPP) performed worse than the parent r2SCAN.
  • Cause: The number of SCF cycles required per MD step was prohibitively high (4-5x more than r2SCAN). The density Laplacian dependence introduces instabilities that prevent rapid convergence when starting from the density of the previous MD step.
  • Observation: Convergence issues worsened with smaller timesteps, and "warming up" with PBE did not resolve the instability.

5. Significance and Conclusion

• Validation of Smoothness: The study confirms that smoothing the switching function in deorbitalized functionals (via the CR approach) significantly reduces potential noise and improves SCF convergence for static solid-state calculations.
• Performance Gains: For solid-state materials science (high-throughput screening, equation of state), the SRPP/SRPP2 deorbitalized r2SCAN offers a viable path to meta-GGA accuracy with near-GGA computational costs (up to 2x speedup).
• Limitations: The approach currently struggles with molecular systems (where accuracy is slightly compromised) and AIMD (where SCF convergence becomes unstable). The sensitivity of the Laplacian term to small density changes in dynamic simulations remains a significant hurdle.
• Future Outlook: The authors suggest that the I_{deorb metric can guide the design of future functionals. Further work is needed to stabilize deorbitalized functionals for AIMD and to optimize them for molecular applications, potentially by further refining the switching functions or the "appropriate norms" used in fitting.

In summary, this paper successfully demonstrates that smoothing the potential is a critical design goal for deorbitalized meta-GGAs, yielding substantial performance benefits for static solid-state calculations while highlighting the remaining challenges in dynamic and molecular contexts.