Restoring the Density-Gradient Expansion for Exchange in Solids and Surfaces

This paper addresses a central problem in density functional theory (DFT) for solids: common generalized gradient approximations (GGAs), especially the widely used PBE functional, are “biased toward atomic energies.” That bias shows up as systematic errors in equilibrium properties of densely packed solids and their surfaces—most notably lattice constants and surface energies. The authors’ research question is therefore: can one revise a GGA so that it restores the correct first-principles density-gradient expansion for exchange (the regime relevant to solids), thereby reducing the atomic bias while still maintaining key exact constraints? This matters because equilibrium lattice constants, bulk moduli, phonons, magnetism, and surface energetics are all sensitive to the exchange-correlation (xc) functional; small systematic shifts in lattice constants can propagate into many predicted materials properties.

The significance of the work is both conceptual and practical. Conceptually, the authors explain a “Procrustean dilemma” for GGAs: functionals tuned to improve atomic energetics often violate the gradient expansion that is valid for slowly varying densities, which is the regime relevant for solids and surfaces. The paper argues that at the GGA level one cannot simultaneously satisfy the requirements for accurate atomic exchange energies and for accurate exchange in solids/surfaces. Practically, the authors propose a specific revised GGA—PBEsol (PBE for solids)—that restores the gradient expansion for exchange over the density-gradient range important for solids, and then adjusts correlation in a way that targets jellium surface energetics.

Methodologically, the paper is primarily a functional-construction and validation study rather than a new ab initio dataset generation. The authors start from the standard GGA form for exchange $$E_x^{\mathrm{GGA}}[n]=\int d^{3}r{e}_x^{\mathrm{unif}}(n(\mathbf{r}))F_x(s(\mathbf{r})),$$ where the enhancement factor depends on the reduced density gradient $$s=\frac{|\nabla n|}{2k_{F}n},k_F=(3{\pi }^{2}n{)}^{1/3}.$$ They emphasize that any GGA recovering the uniform electron gas limit has the small- $s$ expansion $$F_x(s)=1+\mu s^2+\dots (s\to 0),$$ with the exact second-order gradient coefficient for exchange in the slowly varying limit given by ${\mu }_{\mathrm{GE}}=10/81\approx 0.1234$.

A key derivation motivates why exchange in solids requires restoring this coefficient. The authors note that accurate exchange energies of neutral atoms require a much larger effective $\mu$, approximately $\mu \approx 2{\mu }_{\mathrm{GE}}$, because the atomic asymptotic expansion includes a term that is generated by the $s^2$ contribution. They argue that PBE and other common GGAs adopt values of $\mu$ near $2{\mu }_{\mathrm{GE}}$ to match atomic exchange, but that this is harmful for condensed matter where valence densities are often slowly varying (typically $s\lesssim 1$ in densely packed solids). Therefore, they choose $\mu ={\mu }_{\mathrm{GE}}$ for PBEsol to restore the gradient expansion for exchange.

Correlation is treated with a complementary argument. For correlation, the gradient expansion coefficient $\beta$ (in the high-density slowly varying limit) is ${\beta }_{\mathrm{GE}}=0.066725$. However, the authors argue that correlation in solids is less governed by the second-order gradient expansion than by the uniform-gas linear response. They discuss a condition for cancellation between exchange and correlation beyond-LSDA contributions to restore LSDA response, expressed as $$\mu =\frac{{\pi }^2\beta }{3}.$$ If one fully restored the gradient expansion for both exchange and correlation, $\beta$ would need to be $0.0375$ given $\mu ={\mu }_{\mathrm{GE}}$. But the authors prioritize a different constraint: matching the jellium xc surface energy $e_{\mathrm{xc}}^{\mathrm{surf}}$, which controls surface energetics and is dominated by moderately varying densities (again with $s\lesssim 1$) inside the classical turning plane.

To implement this, they fit the correlation parameter $\beta$ using jellium surface energetics. They reference a plausible range for $e_{\mathrm{xc}}^{\mathrm{surf}}$ bounded by TPSS meta-GGA and RPA-like values, and select $\beta =0.046$ (with $\mu ={\mu }_{\mathrm{GE}}$ within the PBE exchange form) to best fit TPSS results. This choice intentionally violates the exact cancellation/linear-response condition in favor of improved surface energies.

The validation uses two main test sets. For equilibrium lattice constants, they employ a test set of 18 solids from Staroverov et al. (Ref. [26]), grouped into simple metals (Li, Na, K, Al), semiconductors (C, Si, SiC, Ge, GaAs), ionic solids (NaF, NaCl, LiCl, LiF, MgO), and transition metals (Cu, Rh, Pd, Ag). They compute lattice constants with the Gaussian orbital periodic code and basis sets of comparable or higher quality, including removal of zero-point anharmonic expansion effects as in the reference dataset.

The results are summarized in Table I. For the full set, the mean error (in $\text{A˚}\times 10^{-2}$) is $-5.5$ for LSDA and $+6.6$ for PBE, indicating opposite systematic biases. TPSS yields a mean error of $+5.4$, while PBEsol reduces the mean error to $+1.3$. In terms of mean absolute error, LSDA is $5.6$, PBE is $6.7$, TPSS is $5.4$, and PBEsol is $1.3$ (a dramatic reduction). The authors note that PBEsol improves lattice constants for essentially all categories, with only a marginal exception for SiC in their discussion.

For atomization energies, they use the AE6 molecular set (SiH4, S2, SiO, C3H4, C2H2O2, C4H8) and report errors in eV (Table II). Here, PBEsol is not designed to excel: the mean error is $3.35$ for LSDA, $0.54$ for PBE, $0.18$ for TPSS, and $1.56$ for PBEsol. The mean absolute error is $3.35$ for LSDA, $0.67$ for PBE, $0.26$ for TPSS, and $1.56$ for PBEsol. The authors interpret this as expected: restoring the exchange gradient expansion for solids improves condensed-matter energetics but worsens atomic total energies, and thus atomization energies.

They also provide a quantitative jellium check for surface exchange energies at bulk density $r_s=3$. The errors in surface exchange energy are reported as 27% for LSDA, $-11$% for PBE, and $+2.7$% for PBEsol. This supports the claim that PBEsol restores the exchange gradient expansion in the relevant density-gradient regime and yields accurate jellium surface exchange.

Limitations are acknowledged implicitly through the design trade-off: PBEsol is explicitly not expected to give good atomization energies, and the authors show that it indeed performs worse than PBE and TPSS on AE6. They also note that the lattice-constant test set is not claimed to be fully representative, being chosen for availability of basis functions and anharmonic corrections. More broadly, the paper’s validation is limited to equilibrium lattice constants, surface-related reasoning via jellium, and atomization energies; it does not present a comprehensive survey of all solid-state properties (e.g., phonons, magnetism, ferroelectricity) within the provided excerpt.

Practical implications are clear. The authors recommend PBEsol for equilibrium geometries and related properties of densely packed solids and their surfaces, especially where PBE’s systematic overestimation of lattice constants and underestimation of surface energies are problematic. They also emphasize ease of adoption: any code implementing PBE can be modified to use PBEsol by replacing only two parameters ($\mu$ and $\beta$) in the GGA form. They caution that pseudopotentials compatible with PBEsol may be needed for pseudopotential-based codes.

Overall, the paper’s core contribution is the restoration of the exchange gradient expansion within a PBE-like GGA framework, paired with a correlation parameter fitted to jellium surface xc energetics. This yields a functional that substantially improves lattice constants and surface energetics for solids, while accepting reduced accuracy for molecular atomization energies—an explicit and defensible trade-off grounded in the physics of slowly varying densities.