betaFeedback

WorksPhysics and Astronomy

Meta-Generalized-Gradient Approximation made Magnetic

Abstract

The Jacob's ladder of density functional theory (DFT) proposes the compelling view that by extending the form of successful approximations – being guided by exact conditions and selected (least empirical) norms – upper rungs will do better than the lower, thus allowing to balance accuracy and computational effort. Meta-generalized-gradient-approximations (MGGAs) belong to the last rung of the semi-local approximations before hybridization with non-local wave function theories. Among the MGGAs, the Strongly Constrained and Appropriately Normed Approximation (SCAN) greatly improves upon GGAs from the lower rung. But the over magnetized solutions of SCAN make GGAs more reliable for magnetism. Here, we provide a solution that satisfies the most pressing desiderata for density functional approximations for ferromagnetic, antiferromagnetic and non-collinear states. The approach is available in an implementation in the Crystal electronic structure package.

Introduction. Nearly one hundred years after Heisenberg proposed the first quantum mechanical explanation of ferromagnetism[1] magnetism continues to pose a challenge for ab initio electronic structure theory. An unsettling example is given by the recently developed strongly-constrained and appropriately normed (SCAN) meta-generalized gradient approximation (MGGA) for density functional theory (DFT).[2, 3]

SCAN satisfies a large set of exact constraints, from known properties of the exact exchange-correlation (xc) functional and gives accurate results at a computational cost of about a generalized-gradient-approximation (GGA). Thus, it can be considered a modern milestone from which it is possible to develop more accurate approximations.[4, 5]

But SCAN does not reproduce correctly the stability and properties of prototypical ferromagnets (FM) like Fe.[6, 7, 8, 9, 10] Overestimated tendency toward magnetism and exaggeration of magnetic energies are obtained for transition metal systems. Extension of the analysis to antiferromagnetic (AFM) solids has shown that modifications of MGGAs yielding an improved description of FM metals produce a deterioration in AFM insulators, and vice versa.[10] The status is such that MGGAs yield no advantage over lower rungs of the semi-local approximation.[10, 11]

Here, we argue that the problem has persisted because the most important exact constraint for dealing with magnetism has not guided most of the DFT developments thus far. We resolve this issue by switching from collinear spin-DFT[12, 13] (the current standard methodology) to the fully non-collinear (nc) Vignale-Rasolt-Bencheikh spin-current DFT (SCDFT) – a complete DFT theory for dealing with magnetism within the framework of the two-component Pauli equation.[14, 15] As will be shown, this extended framework can also solve problems for collinear DFT.

In SCDFT, the most distinctive property of the exact xc energy functional is the form invariance under U(1)×SU(2) gauge transformations.[14, 15, 16] These transformations rotate spinors as well as their phase at each point in space (see Eq. (1) below). Failing to satisfy this invariance leads to functional forms that do not necessarily distinguish between physical variations of the spin density and those that arise from a mere change of reference frame. The gauge invariance of the exact xc-energy functional is thus the most important exact constraint for guiding the construction of magnetic approximations.

Working within the general SCDFT, extra building blocks for the construction of extended approximate functional forms can be gained.[16] In this work, we shall see, the collinear isoorbital indicators used in the SCAN (more below), which treat spin-up and spin-down channels separately, yield unbalanced magnetic solutions,[9, 10] and can be replaced by a unified non-collinear quantity normed by the gauge principle. As a result, the major issues observed in the collinear SCAN solutions get resolved, and the revised functional form is seamlessly extended to the more general non-collinear regime.

From DFT to collinear Spin-DFT. DFT can in principle find the particle density and total energy of real interacting ground states by solving (self-consistently) an auxiliary non-interacting problem

{−12∇2+vext+∫n(r′)∣r−r′∣ d3r′+vxc}φi=εiφi,(1)\left\{ -\frac{1}{2}\nabla^2+v_{\mathrm{ext}}+\int\frac{n(\mathbf{r}')}{\lvert\mathbf{r}-\mathbf{r}'\rvert}\,\mathrm{d}^3r'+v_{\mathrm{xc}}\right\}\varphi_i=\varepsilon_i\varphi_i, \tag*{(1)}

which is the celebrated Kohn-Sham equation. This is a Schrödinger-like equation that describes single-particle orbitals corresponding to the actual external potential (the second term on the l.h.s.) added to the Hartree potential (the third term on the l.h.s.) and the xc potential, vxcv_{\mathrm{xc}}. The latter potential is such that the actual particle density is retrieved n=∑iocc∣φi∣2n=\sum_i^{\mathrm{occ}}\lvert\varphi_i\rvert^2.[17] Correspondingly, the energy of the ground state of the real system EE is thus obtained by

E=Ts+∫d3r n vext+EH[n]+Exc[n],(2)E = T_s + \int d^3r\, n\,v_{\mathrm{ext}} + E_{\mathrm{H}}[n] + E_{\mathrm{xc}}[n], \tag*{(2)}

where Ts=12∑iocc∫d3r ∣∇φi∣2T_s = \frac{1}{2}\sum_i^{\mathrm{occ}}\int d^3r\,|\nabla\varphi_i|^2 is the Kohn-Sham kinetic energy, EH[n]=∬d3r d3r′n(r)n(r′)2∣r−r′∣E_{\mathrm{H}}[n] = \iint d^3r\,d^3r'\frac{n(r)n(r')}{2|r-r'|} is the Hartree energy and Exc[n]E_{\mathrm{xc}}[n] is the xc energy functional. For retrieving the exact total energy of the actual state, Exc[n]E_{\mathrm{xc}}[n] accounts for the electron-electron energy beyond the classical evaluation obtained via EH[n]E_{\mathrm{H}}[n] as well as the kinetic energies beyond TsT_s. Also notice that vxc=δExc/δnv_{\mathrm{xc}} = \delta E_{\mathrm{xc}}/\delta n. All that is needed to make the above approach practical is a density functional approximation (DFA) for Exc[n]E_{\mathrm{xc}}[n]. Solutions to nn and EE can thus be found in an iterative fashion.

The first and simplest ExcDFA[n]E_{\mathrm{xc}}^{\mathrm{DFA}}[n] is the local density approximation (LDA)[18]

ExcLDA[n]=∫d3r n ϵxcunif(n),(3)E_{\mathrm{xc}}^{\mathrm{LDA}}[n] = \int d^3r\,n\,\epsilon_{\mathrm{xc}}^{\mathrm{unif}}(n), \tag*{(3)}

where the xc energy density is approximated locally with that of the uniform gas, ϵxcunif(n)\epsilon_{\mathrm{xc}}^{\mathrm{unif}}(n), for which analytical and accurate numerical reference data are available. To capture the variations of more realistic states, a GGA uses information retrieved from the gradients of the density via s=∣∇n∣/[2(3π2)1/3n4/3]s = |\nabla n|/[2(3\pi^2)^{1/3}n^{4/3}]

ExcGGA[n]=∫d3r n ϵxcunif(n) FxcGGA(n,s),(4)E_{\mathrm{xc}}^{\mathrm{GGA}}[n] = \int d^3r\,n\,\epsilon_{\mathrm{xc}}^{\mathrm{unif}}(n)\,F_{\mathrm{xc}}^{\mathrm{GGA}}(n,s), \tag*{(4)}

where FxcGGA(n,s)F_{\mathrm{xc}}^{\mathrm{GGA}}(n,s) is the xc enhancement factor that connects the LDA energy densities to the GGA energy densities. For slowly varying density, the LDA is retrieved (i.e. FxcGGA(n,0)=1F_{\mathrm{xc}}^{\mathrm{GGA}}(n,0) = 1). The most popular choice for the enhancement factor is the one by Perdew, Burke and Ernzerhof (PBE) FxcPBEF_{\mathrm{xc}}^{\mathrm{PBE}}.[19]

MGGAs[2, 20, 21] introduce an extra dependence beyond the electron density by employing

ExcMGGA[n]=∫d3r n ϵxcunif(n) FxcMGGA(n,s,τ),(5)E_{\mathrm{xc}}^{\mathrm{MGGA}}[n] = \int d^3r\,n\,\epsilon_{\mathrm{xc}}^{\mathrm{unif}}(n)\,F_{\mathrm{xc}}^{\mathrm{MGGA}}(n,s,\tau), \tag*{(5)}

where τ=12∑iocc∣∇φi∣2\tau= \frac{1}{2}\sum_i^{\mathrm{occ}}|\nabla\varphi_i|^2 is the (KS) kinetic energy density.[22] Here we focus on those MGGAs that depend on τ\tau via

α:=τ−∣∇n∣28nτunif,(6)\alpha:= \frac{\tau-\frac{|\nabla n|^2}{8n}}{\tau_{\mathrm{unif}}}, \tag*{(6)}

where τunif=310(3π2)2/3n5/3\tau_{\mathrm{unif}} = \frac{3}{10}(3\pi^2)^{2/3}n^{5/3} is the kinetic energy density of the uniform gas at density nn. In particular, the SCAN form is defined by setting FxcSCAN(n,s,α)=hxc1(n,s)+[hxc0(n,s)−hxc1(n,s)]fxc(α)F_{\mathrm{xc}}^{\mathrm{SCAN}}(n,s,\alpha) = h_{\mathrm{xc}}^1(n,s) + [h_{\mathrm{xc}}^0(n,s)-h_{\mathrm{xc}}^1(n,s)]f_{\mathrm{xc}}(\alpha) where fxc(α)f_{\mathrm{xc}}(\alpha) interpolates between the semi-local xc enhancement factors for single-orbital densities (hxc1h_{\mathrm{xc}}^1) and for slowly-varying densities (hxc0h_{\mathrm{xc}}^0). This modeling exploits the fact that α\alpha [Eq. (6)] tends to zero in the iso-orbital limit (i.e., when the density is dominated by a single occupied orbital), tends to one in the uniform-density limit, and becomes ≫1\gg1 in regions of low density between closed shells

For solving ground states involving collinear spin polarization either induced by external Zeeman couplings or spontaneously via magnetic phase transitions, the standard approach is to switch from DFT to Spin-DFT.[12, 13] The method requires to solve a spin-dependent, single-orbital problem

{−12∇2+vext,σ+vHxc,σ}φi,σ=εi,σφi,σ.(7)\left\{-\frac{1}{2}\nabla^2+v_{\mathrm{ext},\sigma}+v_{\mathrm{Hxc},\sigma}\right\}\varphi_{i,\sigma}=\varepsilon_{i,\sigma}\varphi_{i,\sigma}. \tag*{(7)}

Here, EH[n]E_{\mathrm{H}}[n] and corresponding vHv_{\mathrm{H}} are unchanged but the spin-dependent xc-potential, vxc,σ=δExc/δnσv_{\mathrm{xc},\sigma} = \delta E_{\mathrm{xc}}/\delta n_\sigma, is generated from the functional Exc[n↑,n↓]E_{\mathrm{xc}}[n_\uparrow,n_\downarrow], which depends both on n↑n_\uparrow and n↓n_\downarrow and not just on their sum n=∑σnσn = \sum_\sigma n_\sigma.

Extended forms for exchange can be generated from given DFT approximations using the collinear spin decomposition for exchange[23]

ExSDFA[n↑,n↓]=12ExDFA[2n↑]+12ExDFA[2n↓].(8)E_{\mathrm{x}}^{\mathrm{SDFA}}[n_\uparrow,n_\downarrow] = \frac{1}{2}E_{\mathrm{x}}^{\mathrm{DFA}}[2n_\uparrow] + \frac{1}{2}E_{\mathrm{x}}^{\mathrm{DFA}}[2n_\downarrow]. \tag*{(8)}

For correlation, we must use

EcSDFA[n↑,n↓]=∫d3r n ϵcunif(n,ζ) FcSDFA(n,ζ,s,α).(9)E_{\mathrm{c}}^{\mathrm{SDFA}}[n_\uparrow,n_\downarrow] = \int d^3r\,n\,\epsilon_{\mathrm{c}}^{\mathrm{unif}}(n,\zeta)\,F_{\mathrm{c}}^{\mathrm{SDFA}}(n,\zeta,s,\alpha). \tag*{(9)}

Unlike exchange, the correlation functional depends in a non separable fashion on n↑n_\uparrow and n↓n_\downarrow through the spin polarization (density) ζ=∣(n↑−n↓)/(n↑+n↓)∣\zeta= |(n_\uparrow-n_\downarrow)/(n_\uparrow+n_\downarrow)|; α=(τ−∣∇n∣28n)/τunifds\alpha= (\tau-\frac{|\nabla n|^2}{8n})/\tau_{\mathrm{unif}}d_s is the spin-summed iso-orbital indicator with τ=12∑i,σocc∣∇φi∣2\tau= \frac{1}{2}\sum_{i,\sigma}^{\mathrm{occ}}|\nabla\varphi_i|^2 and ds=12[(1+ζ)5/3+(1−ζ)5/3]d_s = \frac{1}{2}[(1+\zeta)^{5/3}+(1-\zeta)^{5/3}]. Also note that the exchange functional of [Eq. (8)], ExDFA[n↑,n↓]E_{\mathrm{x}}^{\mathrm{DFA}}[n_\uparrow,n_\downarrow], can be equivalently reexpressed as ExDFA[n,ζ]E_{\mathrm{x}}^{\mathrm{DFA}}[n,\zeta] in terms of the total density nn and polarization ζ\zeta.

Issues. Notice that α\alpha does not carry any explicit information on ∇ζ\nabla\zeta. Therefore, the above expressions for GGAs and MGGAs come with a striking asymmetry. While the gradients of ζ\zeta are included in the exchange functional indirectly via the spin-scaling relation [i.e., Eq. (8)], the correlation form has no dependence on ∇ζ\nabla\zeta. This asymmetry reflects the lack of invariance of the approximate forms under U(1)×\timesSU(2) gauge transformations. Although this deficiency might not be perceived as a problem as long as we consider scenarios that are not gauge symmetric — such as FMs or AFMs — one should not confuse the symmetry properties of particular solutions with those of the exact functional. To sort the latter out, we must enlarge our modeling framework.

Switching to SCDFT. Let us account not only for a scalar multiplicative potential vextv_{\mathrm{ext}} and a multiplicative magnetic field BaB^a, but also for a (particle-) vector potential AA and a spin-vector potential AaA^a. While AA is useful to represent the vector potential of an external magnetic field, AaA^a is useful to represent the (one-body) spin-orbit couplings in the system.[14, 15]

The KS equations in SCDFT have the form of single-particle Pauli equations[14, 15]

[12(−i∇+1cAKS)2+vKS]Φk=ϵkΦk(10)\left[\frac{1}{2}\left(-i\nabla+\frac{1}{c}\mathbf{A}_{\mathrm{KS}}\right)^2+v_{\mathrm{KS}}\right]\Phi_k=\epsilon_k\Phi_k \tag*{(10)}

where Φk\Phi_k are two-component spinors, AKS=(A+Axc)+σa(Aa+Axca)\mathbf{A}_{\mathrm{KS}}=(\mathbf{A}+\mathbf{A}_{\mathrm{xc}})+\sigma^a(\mathbf{A}^a+\mathbf{A}^a_{\mathrm{xc}}), vKS=(v+vH+vxc)+σa(Ba+Bxca)+12c2[(A+σaAa)2−AKS2]v_{\mathrm{KS}}=(v+v_{\mathrm{H}}+v_{\mathrm{xc}})+\sigma^a(B^a+B^a_{\mathrm{xc}})+\frac{1}{2c^2}\left[(\mathbf{A}+\sigma^a\mathbf{A}^a)^2-\mathbf{A}_{\mathrm{KS}}^2\right], in which σa\sigma^a denotes the component of the vector of Pauli matrices, 1cAxc=δExc/δj\frac{1}{c}\mathbf{A}_{\mathrm{xc}}=\delta E_{\mathrm{xc}}/\delta\mathbf{j} is an Abelian xc-vector potential, 1cAxca=δExc/δJa\frac{1}{c}\mathbf{A}^a_{\mathrm{xc}}=\delta E_{\mathrm{xc}}/\delta\mathbf{J}^a is the aa-th component of a non-Abelian xc-vector potential, Bxca=δExc/δmaB^a_{\mathrm{xc}}=\delta E_{\mathrm{xc}}/\delta m^a is the aa-th component of a xc-magnetic-like potential, vxc=δExc/δnv_{\mathrm{xc}}=\delta E_{\mathrm{xc}}/\delta n is a xc-scalar potential, and vHv_{\mathrm{H}} is the usual Hartree potential. The KS densities are obtained from the (occupied) two-component KS spinors as follows: n=∑koccΦk†Φkn=\sum_k^{\mathrm{occ}}\Phi_k^\dagger\Phi_k, m⃗=∑koccΦk†σ⃗Φk\vec{m}=\sum_k^{\mathrm{occ}}\Phi_k^\dagger\vec{\sigma}\Phi_k, j=12i∑koccΦk†[∇Φk]−[∇Φk†]Φk\mathbf{j}=\frac{1}{2i}\sum_k^{\mathrm{occ}}\Phi_k^\dagger[\nabla\Phi_k]-[\nabla\Phi_k^\dagger]\Phi_k, and J⃗=12i∑koccΦk†σ⃗[∇Φk]−[∇Φk†]σ⃗Φk\vec{\mathbf{J}}=\frac{1}{2i}\sum_k^{\mathrm{occ}}\Phi_k^\dagger\vec{\sigma}[\nabla\Phi_k]-[\nabla\Phi_k^\dagger]\vec{\sigma}\Phi_k.

Because Exc[n,m⃗,j,J⃗]E_{\mathrm{xc}}[n,\vec{m},\mathbf{j},\vec{\mathbf{J}}] must yield the exact interacting densities via the KS densities, it must also be form invariant under (local) U(1)×\timesSU(2) gauge transformations.[14, 15, 16] In detail, such a transformation acts on spinors as follows

Φ′(r)=exp⁡[icχ(r)I+icλa(r)σa]Φ(r)(11)\Phi'(r)=\exp\left[\frac{i}{c}\chi(r)I+\frac{i}{c}\lambda^a(r)\sigma^a\right]\Phi(r) \tag*{(11)}

where II is the 2×22\times2 identity and χ\chi and λa\lambda^a are arbitrary real-valued functions.

We recall from Eq. (5) that MGGAs, such as SCAN, explicitly depend on the KS kinetic energy density τ\tau. For dealing with nc solutions, the kinetic energy density τ\tau must be expressed in terms of the occupied two-component KS spinors τnc=12∑kocc(∂μΦk†)(∂μΦk)\tau_{\mathrm{nc}}=\frac{1}{2}\sum_k^{\mathrm{occ}}(\partial_\mu\Phi_k^\dagger)(\partial_\mu\Phi_k). But, it can be directly verified that τnc\tau_{\mathrm{nc}} is not gauge invariant.[16] Thus, αnc\alpha_{\mathrm{nc}} obtained by replacing τ\tau with τnc\tau_{\mathrm{nc}} in Eq. (6) is not invariant either. The corresponding extension of the SCAN functional by letting α→αnc\alpha\to\alpha_{\mathrm{nc}} would therefore lack the most fundamental property for dealing with magnetism. Fortunately, this problem can be fixed by the replacement τnc→τ~nc\tau_{\mathrm{nc}}\to\widetilde{\tau}_{\mathrm{nc}} with[16, 25]

τ~nc=τnc+maτncan+∇ma⋅∇ma8n−j⋅j2n−Ja⋅Ja2n(12)\widetilde{\tau}_{\mathrm{nc}}=\tau_{\mathrm{nc}}+\frac{m^a\tau_{\mathrm{nc}}^a}{n}+\frac{\nabla m^a\cdot\nabla m^a}{8n}-\frac{\mathbf{j}\cdot\mathbf{j}}{2n}-\frac{\mathbf{J}^a\cdot\mathbf{J}^a}{2n} \tag*{(12)}

where τnca=12∑kocc(∂μΦk†)σa(∂μΦk)\tau_{\mathrm{nc}}^a=\frac{1}{2}\sum_k^{\mathrm{occ}}(\partial_\mu\Phi_k^\dagger)\sigma^a(\partial_\mu\Phi_k) is the spin-kinetic energy density.

Eq. (12) not only includes ∇m⃗\nabla\vec{m} and the (spin-)paramagnetic currents but, for consistency with the exact gauge-invariance requirements, also the coupling between m⃗\vec{m} and τ⃗nc\vec{\tau}_{\mathrm{nc}}.

Eq. (12) includes the well-known particle current corrections that ensure the U(1) invariance of the MGGA, which has been advanced both in practice and in theory.[26, 27, 28, 29, 30, 31, 32, 33, 34, 35] It also includes the spin current corrections that were recently shown to ensure the SU(2) invariance of the MGGA for time-reversal symmetric spin-orbit coupling — the JSCAN.[36] Eq. (12) was also at the center of the derivation of a non-collinear ELF.[25] But it did not find exploitation in energy functional approximations to determine magnetic solutions.

We recall that the problems of overmagnetization by SCAN were imputed, in Refs. 9,10, to the different behaviour of the isoorbital indicators in the exchange of SCAN for the two spin channels. Below, we solve these problems by invoking a single α~\widetilde{\alpha} for both spin channels — normed by the gauge principle. For capturing energies rather than just bond pictures, the following gauge invariant extension of SCAN suggests itself

Exnc[n,m⃗,j,J⃗]=∫d3r n εxunif(n,ζnc)FxSCAN(n,s,α~nc)(13)E_{\mathrm{x}}^{\mathrm{nc}}[n,\vec{m},\mathbf{j},\vec{\mathbf{J}}]=\int d^3r\,n\,\varepsilon_{\mathrm{x}}^{\mathrm{unif}}(n,\zeta_{\mathrm{nc}})F_{\mathrm{x}}^{\mathrm{SCAN}}(n,s,\widetilde{\alpha}_{\mathrm{nc}}) \tag*{(13)}

and

Ecnc[n,m⃗,j,J⃗]=∫d3r n εcunif(n,ζnc)FcSCAN(n,ζnc,s,α~nc)(14)E_{\mathrm{c}}^{\mathrm{nc}}[n,\vec{m},\mathbf{j},\vec{\mathbf{J}}]=\int d^3r\,n\,\varepsilon_{\mathrm{c}}^{\mathrm{unif}}(n,\zeta_{\mathrm{nc}})F_{\mathrm{c}}^{\mathrm{SCAN}}(n,\zeta_{\mathrm{nc}},s,\widetilde{\alpha}_{\mathrm{nc}}) \tag*{(14)}

where ζnc≡∣m⃗∣/n\zeta_{\mathrm{nc}}\equiv|\vec{m}|/n, and

α~nc=2nτ~nc−(∇n)⋅(∇n)4(n+∣m⃗∣)τunif++(n−∣m⃗∣)τunif−(15)\widetilde{\alpha}_{\mathrm{nc}}=\frac{2n\widetilde{\tau}_{\mathrm{nc}}-\frac{(\nabla n)\cdot(\nabla n)}{4}}{(n+|\vec{m}|)\tau_{\mathrm{unif}}^+ +(n-|\vec{m}|)\tau_{\mathrm{unif}}^-} \tag*{(15)}

in which τunif±≡3/10(3π2)2/3(n±∣m⃗∣)5/3\tau_{\mathrm{unif}}^\pm\equiv3/10(3\pi^2)^{2/3}(n\pm|\vec{m}|)^{5/3}, and all uniform gas quantities have been upgraded to an nc form by locally quantizing spin along the axis defined by the spin density.[37, 38] The derivation of Eqs. (15) and (16) is provided in the Supplementary Material (this includes Refs. 39–59).[60, 61]

For self-consistent collinear solutions, a modified SCAN functional (named mSCAN) is straightforwardly obtained solely in terms of nσn^\sigma, εx/cunif(n↑,n↓)\varepsilon_{\mathrm{x/c}}^{\mathrm{unif}}(n^\uparrow,n^\downarrow), and

α~=n↑τ↑+n↓τ↓−(∇n↑)⋅(∇n↓)4n↑τunif↑+n↓τunif↓(16)\widetilde{\alpha}=\frac{n^\uparrow\tau^\uparrow+n^\downarrow\tau^\downarrow-\frac{(\nabla n^\uparrow)\cdot(\nabla n^\downarrow)}{4}}{n^\uparrow\tau_{\mathrm{unif}}^\uparrow+n^\downarrow\tau_{\mathrm{unif}}^\downarrow} \tag*{(16)}

where τunifσ=3/10(6π2)2/3(nσ)5/3\tau_{\mathrm{unif}}^\sigma=3/10(6\pi^2)^{2/3}(n^\sigma)^{5/3}.[62] Finally, the mSCAN reduces to the SCAN on closed-shell solutions (more below).

Notable magnetic SCAN failures — resolved. Let’s scrutinize the novel collinear solutions. We implemented the mSCAN functional in a developer’s version of the CRYSTAL23 program.[63] We study current-less, collinear magnets. Thus, the single-particle equation of SCDFT given in Eq. (10) is considerably simplified, as the vector potentials Axc\mathbf{A}_{\mathrm{xc}} and Axca\mathbf{A}^a_{\mathrm{xc}} vanish, and we must self-consistently solve the simpler Eq. (7) in a basis of Bloch orbitals φi,σ,k\varphi_{i,\sigma,\mathbf{k}} (with wavevector k\mathbf{k}) formed from a linear combination of atom-centered spherical Gaussian-type functions (GTF).[64] In Eq. (7) for mSCAN, in a generalized KS context,[65] the spin-dependent xc potential vxc,σv_{\mathrm{xc},\sigma} includes an additional semilocal term coming from the spin-kinetic energy dependence.[41, 66, 67] Computational details for the general case are provided in the SM.

Before applying mSCAN to prototypical magnets, notice that the x-mSCAN does not fulfill the (strictly collinear) spin decomposition relation of Eq. (8), yet it fulfills the exact linear scaling conditions (collinear or non-collinear). An extended analysis is provided in the SM.

Although SCAN fits the energy of the H atom as an appropriate norm and assigns no correlation to one-electron systems, Hartree and x-SCAN do not cancel for one-electron systems. mSCAN breaks both the constraint of zero correlation energy and zero Hx for one-electron systems. Yet, tests show that the overall treatment of the Hxc energies in mSCAN is nonetheless appropriately normed by fulfilling the gauge invariance — not fulfilled by the SCAN form. For example, for H2+_2^+ (see Figure S1) neither SCAN nor mSCAN matches the exact energies: while SCAN is accurate at short separation, mSCAN is accurate at large intermolecular separation. Taking another one-electron example, the potential energy curve of Li25+_2^{5+} is most accurate (as compared to LDA, GGA and SCAN) by mSCAN at all distances (Fig. S2).

Next, we consider also materials. In Table I we list the spin magnetic moments (calculated analytically in the Gaussian basis set) and the magnetic energies (in meV), calculated as energy differences w.r.t. the self-consistent spin-restricted solution for the FM metals Fe, Co and Ni. All calculations employ the experimental geometries reported in Ref. 10, unless explicitly stated otherwise. Comparisons are provided against the LDA (employing the parametrization of correlation energies by Vosko, Wilk and Nusair, and the usual Slater-Dirac energy functional for exchange SVWN5),[72] the GGA of PBE,[19] and the original spin-unrestricted SCAN.[2] Comparison against experimental data for the spin-magnetic moments are also given.

FeCoNi
LDA2.23 (467)1.57 (396)0.56 (40.8)
PBE2.26 (581)1.59 (504)0.56 (49.0)
SCAN2.62 (1177)1.74 (1175)0.66 (113)
mSCAN2.15 (461)1.58 (447)0.56 (47.3)
EXP1.98–2.08[68, 69, 70]1.52–1.62[68, 69, 70, 71]0.52–0.55[69, 70]

Table I. Spin moments (in μB\mu_\mathrm{B}/atom) and, within parentheses, magnetic energies (in meV) of FM metals as calculated with DFAs and compared against experiments.

In accord with previous reports,[6, 9, 10] we find an agreement with experiments at the level of LDA, which is not significantly changed with the GGA. SCAN displays the previously reported overmagnetization in all FM metals, both in its significant overestimation of the magnetic moments, as well as overestimation of the magnetic energies.[6, 9, 10]

In the case of Fe, for example, the self-consistently predicted magnetic moment of 2.62 μB\mu_\mathrm{B} by SCAN represents an overestimation of around 30 %. Overmagnetization of around 10 % and 15 % w.r.t. experimental values are predicted by SCAN for Co and Ni respectively, while the LDA and GGA both fall within the experimental uncertainty. Magnetic energies of SCAN are consistently more than double the values predicted by the LDA and GGA for the FM metals. The problem is completely solved by switching to our gauge-invariant mSCAN approach, which provides similar magnetic moments (and magnetic energies) as the LDA and GGA on Co and Ni, while the predicted magnetic moment on Fe of 2.15 μB\mu_\mathrm{B} is the closest to the experimental range of 1.98–2.08 μB\mu_\mathrm{B}.

The improvement provided by our approach is even more striking when applied to the non-magnetic metals V and Pd (see Table II), as well as the difficult case of Cr2_2.[73, 74] We perform the calculations starting from a spin-polarized guess and let the spin moments relax during the self-consistent procedure. Like in previous works,[6, 10] SCAN calculations incorrectly predict a net magnetic moment of 0.33 and 0.44 μB\mu_\mathrm{B} on V and Pd respectively. In the case of Pd, the incorrect magnetic behaviour is also predicted by PBE. For Cr2_2, the LDA, PBE and SCAN all predict a spurious magnetization at the equilibrium geometry (ground state properties are reported in Table S1 of the SM). Remarkably, the correct non-magnetic behaviour of V, Pd and Cr2_2 is predicted by the proposed gauge-invariant mSCAN approach. As for itinerant two-dimensional systems, in Graphene, SCAN incorrectly predicts a magnetic insulating phase,[74] while mSCAN correctly provides a non-magnetic, metallic state, with a Dirac point at the zone corners (Table S2).

VPdCr2_2
LDA0.00 (0.00)0.00 (0.00)2.69 (0.30)*
PBE0.00 (0.00)0.20 (0.70)5.32 (0.87)*
SCAN5.57 (6.27)0.44 (17.2)5.62 (10.3)*
mSCAN0.00 (0.00)0.00 (0.00)0.00 (0.00)
EXP0.00[75]0.00[76]0.00[73]

Table II. Same as in Table I but for non-magnetic systems.

Although the failures of SCAN on itinerant magnets are daunting, accurate magnetic moments and gaps have been predicted in the AFM Mott insulators FeO, CoO and NiO, which exhibit more localized magnetism.[10, 77] Values obtained from our mSCAN on these highly-correlated Mott insulators are provided in Table III, including a comparison against the experiment both on magnetic moments and fundamental band gaps (in parentheses). thesis, given in eV). The LDA and GGA incorrectly predict a metallic behaviour for FeO and CoO, while both SCAN and mSCAN provide the qualitatively correct insulating behaviour. All functionals provide similar values for the spin magnetic moments in the Mott insulators, although both moments and band gaps are larger with SCAN. We also consider a topological AFM – the Chern magnet TbMn6_6Sn6_6, in which calculation of spin moments is a reported challenge for SDFT.[78] With mSCAN, we obtain excellent agreement against the experimental values (with large improvement over SCAN and PBE) in Table S3.

FeOCoONiO
LDA3.51 (0.000)2.40 (0.000)1.20 (0.452)
PBE3.57 (0.000)2.48 (0.000)1.37 (0.986)
SCAN3.67 (0.635)2.65 (1.29)1.62 (2.66)
mSCAN3.57 (0.224)2.52 (0.440)1.44 (1.29)
EXP*2.32-4.00[79, 80, 81]1.75-2.98[79, 82, 83, 84]1.45-1.90[79, 85, 86, 87]
EXP(2.40)[88](3.10-4.10)[89](3.7)[90]

Table III. Spin moments (in μB\mu_{\mathrm{B}}/atom) and, within parenthesis, band gaps (in eV) of AFM Mott insulators.

Next, we investigate the performance of mSCAN on structural optimizations in all tested systems (see Table S4). A previous attempt at partially fixing the spin moments of SCAN by varying the c1xc_{1x} parameter has produced significantly worsened results for lattice constants, with errors increased by around 50%.[10] Here, mean absolute errors (MAE) decrease in the order LDA, GGA, mSCAN (SCAN), with values of 0.663, 0.521, 0.295 (0.319), which is again consistent with the principle of the Jacob’s ladder of DFT.

Let us now scrutinize the dissociation of first-row diatomic molecules involving open-shell atoms.[91] Fig. S3 reports the case of H2_2 compared to the accurate result of Ref. 92: the mSCAN and SCAN energies match up until an intermolecular separation of about 3 Å, at which point a spurious magnetization appears in the spin-unrestricted LDA, PBE and SCAN solutions. mSCAN provides a correct vanishing spin magnetization up to about 3.5 Å. In the C2_2 molecule LDA, PBE and SCAN provide spurious magnetization even at the equilibrium geometry, which is resolved by mSCAN, see Fig. S4 and Table S5. mSCAN also provides an improvement over SCAN, with respect to the reference total energies of Ref. 93 on C2_2, both at the equilibrium geometry and at large distances. Fig. S5 reports the dissociation of the N2_2 molecule. mSCAN provides a significant improvement over LDA, PBE and SCAN at distances larger than about 1.5 Å, where strong correlation becomes important (as also corroborated by comparison of multi- and single-reference coupled-cluster treatments).[94]

A word of caution and of encouragement is in order: The novel functional behaves differently from SCAN not only for magnetized states but also for open-shell states, therefore atomization energies should be scrutinized with care. We have also pointed out that band gaps may be worsened. Therefore, superseding the SCAN functional in all respects will require further efforts, which, we believe, can exploit the recently discovered extra-degrees of freedom that are allowed by the MGGAs.[5]

Noncollinear extension. Fully self-consistent SCDFT MGGA calculations for non-collinear magnetic states are reported in a parallel work.[95] Ref. 95 also calls for a gauge-invariant nc-extension of SCAN.

Thus, we finally solve Eq. (10) using the nc forms in Eqs. (13a) and (13b). Technical details are provided in the SM, where we also demonstrate that the functional provides a non-zero xc magnetic torque, which vanishes over the full space. Last but not least, spin-orbit coupling can be accounted for self-consistently. So, our proposal fulfills the most pressing * desiderata* for a nc approach.

We applied it to the Cr3_3 (bond length of 3.7 Bohr), which has been an important model system in previous studies.[25, 95, 96, 97, 98] The magnetization reported in Figure 1 is indicative of a spin-frustrated system. We observe a spin-moment of 1.38 μB\mu_{\mathrm{B}}, which is significantly lower than the SCDFT-MGGA value of 3.15 μB\mu_{\mathrm{B}} reported in Ref. 95, but closer to the LDA value of 1.68 μB\mu_{\mathrm{B}}.

Magnetization in the Cr3 molecule with the nc SCAN functional

Figure 1. Magnetization in μB\mu_{\mathrm{B}} of the Cr3_3 molecule with the nc SCAN functional.

Conclusions. While it remains a challenge to assess whether a semi-local density functional may be derived to embrace all the relevant non-magnetic and magnetic (both collinear and non-collinearly polarized) exact constraints, the core findings of this work solves most urgent failures for magnetism and resumes progress towards improved density functional approximations.

Acknowledgments

This research has received funding from the Project CH4.0 under the MUR program “Dipartimenti di Eccellenza 2023-2027” (CUP: D13C22003520001). GV was supported by the Ministry of Education, Singapore, under its Research Centre of Excellence award to the Institute for Functional Intelligent Materials (I-FIM, project No. EDUNC-33-18-279-V12).

References

References

  1. [1]W. Heisenberg, Z. Physik 49, 619–636 (1928).
  2. [2]J. Sun, A. Ruzsinszky, and J. P. Perdew, Phys. Rev. Lett. 115, 036402 (2015), URL https://link.aps.org/doi/10.1103/PhysRevLett.115.036402.
  3. [3]J. Sun, R. C. Remsing, Y. Zhang, Z. Sun, A. Ruzsinszky, H. Peng, Z. Yang, A. Paul, U. Waghmare, X. Wu, et al., Nat Chem 8, 831 (2016).
  4. [4]T. Aschebrock and S. Kümmel, Phys. Rev. Res. 1, 033082 (2019), URL https://link.aps.org/doi/10.1103/PhysRevResearch.1.033082.
  5. [5]T. Lebeda, T. Aschebrock, and S. Kümmel, Phys. Rev. Lett. 133, 136402 (2024), URL https://link.aps.org/doi/10.1103/PhysRevLett.133.136402.
  6. [6]Y. Fu and D. J. Singh, Phys. Rev. Lett. 121, 207201 (2018), URL https://link.aps.org/doi/10.1103/PhysRevLett.121.207201.
  7. [7]M. Ekholm, D. Gambino, H. J. M. Jönsson, F. Tasnádi, B. Alling, and I. A. Abrikosov, Phys. Rev. B 98, 094413 (2018), URL https://link.aps.org/doi/10.1103/PhysRevB.98.094413.
  8. [8]Y. Fu and D. J. Singh, Phys. Rev. B 100, 045126 (2019), URL https://link.aps.org/doi/10.1103/PhysRevB.100.045126.
  9. [9]D. Mejía-Rodríguez and S. B. Trickey, Phys. Rev. B 100, 041113 (2019), URL https://link.aps.org/doi/10.1103/PhysRevB.100.041113.
  10. [10]F. Tran, G. Baudesson, J. Carrete, G. K. H. Madsen, P. Blaha, K. Schwarz, and D. J. Singh, Phys. Rev. B 102, 024407 (2020), URL https://link.aps.org/doi/10.1103/PhysRevB.102.024407.
  11. [11]A. Romero and M. Verstraete, Eur. Phys. J. B 91, 193 (2018), URL https://link.springer.com/article/10.1140/epjb/e2018-90275-5#citeas.
  12. [12]U. von Barth and L. Hedin, J. Phys. C 5, 1629 (1972).
  13. [13]O. Gunnarsson and B. I. Lundqvist, Phys. Rev. B 13, 4274 (1976), URL https://link.aps.org/doi/10.1103/PhysRevB.13.4274.
  14. [14]G. Vignale and M. Rasolt, Phys. Rev. B 37, 10685 (1988).
  15. [15]K. Bencheikh, J. Phys. A 36, 11929 (2003).
  16. [16]S. Pittalis, G. Vignale, and F. G. Eich, Phys. Rev. B 96, 035141 (2017).
  17. [17]For simplicity the expressions we assume integer occupation numbers but in applications fractional occupation may also occur.).
  18. [18]W. Kohn and L. J. Sham, Phys. Rev. 140, A1133 (1965).
  19. [19]J. P. Perdew, K. Burke, and M. Ernzerhof, Physical review letters 77, 3865 (1996).
  20. [20]J. P. Perdew, S. Kurth, A. c. v. Zupan, and P. Blaha, Phys. Rev. Lett. 82, 2544 (1999), URL https://link.aps.org/doi/10.1103/PhysRevLett.82.2544.
  21. [21]J. Tao, J. P. Perdew, V. N. Staroverov, and G. E. Scuseria, Phys. Rev. Lett. 91, 164401 (2003).
  22. [22]A. D. Kaplan and J. P. Perdew, Phys. Rev. Mater. 6, 083803 (2022), URL https://link.aps.org/doi/10.1103/PhysRevMaterials.6.083803.
  23. [23]We refrain from using the term “spin scaling relations” to avoid confusion between aspects of the spin separability with aspects of uniform coordinate scaling in the more general non-collinear framework.
  24. [25]J. K. Desmarais, G. Vignale, K. Bencheikh, A. Erba, and S. Pittalis, Physical Review Letters 133, 136401 (2024).
  25. [26]J. F. Dobson, The Journal of Chemical Physics 98, 8870 (1993).
  26. [27]A. D. Becke, The Journal of Chemical Physics 117, 6935 (2002).
  27. [28]J. Tao and J. P. Perdew, Phys. Rev. Lett. 95, 196403 (2005).
  28. [29]S. Pittalis, S. Kurth, N. Helbig, and E. K. U. Gross, Phys. Rev. A 74, 062511 (2006).
  29. [30]S. Pittalis, S. Kurth, S. Sharma, and E. K. U. Gross, The Journal of Chemical Physics 127, 124103 (2007).
  30. [31]E. Räsänen, S. Pittalis, C. R. Proetto, and E. K. U. Gross, Phys. Rev. B 79, 121305 (2009).
  31. [32]J. W. Furness, J. Verbeek, E. I. Tellgren, S. Stopkowicz, U. Ekström, T. Helgaker, and A. M. Teale, Journal of chemical theory and computation 11, 4169 (2015).
  32. [33]E. I. Tellgren, A. Laestadius, T. Helgaker, S. Kvaal, and A. M. Teale, The Journal of chemical physics 148 (2018).
  33. [34]A. Laestadius, E. I. Tellgren, M. Penz, M. Ruggenthaler, S. Kvaal, and T. Helgaker, Journal of Chemical Theory and Computation 15, 4003 (2019).
  34. [35]M. J. Pemberton, T. J. Irons, T. Helgaker, and A. M. Teale, The Journal of Chemical Physics 156 (2022).
  35. [36]J. K. Desmarais, J. Maul, B. Cavalleri, A. Erba, G. Vignale, and S. Pittalis, Physical Review Letters 132, 256401 (2024).
  36. [37]J. Kübler, K.-H. Hock, J. Slicht, and A. R. Williams, J. Phys. F 18, 469 (1988).
  37. [38]We stress that α̃nc differs from a previously reported non-collinear iso-orbital indicator underlying the recent non-collinear extension of the ELF.25 This is due to the fact that, in the present context by reusing the LSDA as the cornerstone of the new approximation, we must use as a reference a uniform gas that is polarized non-collinearly.
  38. [41]S. Lehtola, F. Blockhuys, and C. Van Alsenoy, Molecules (2020).
  39. [60]See Supplementary Material at INSERT URL HERE for a detailed derivation of α̃ and α̃nc, a proof that D̃xCSCAN exerts a torque on m⃗ that vanishes over the full space, expressions of the mSCAN, computational details, and all reported dissociation curves, lattice constants, bond lengths, as well as supplementary results on Cr₂, Graphene, and the Chern magnet.
  40. [61]For completeness, the collinear limit of Eqs. 13a and 13b is reported in the SM.
  41. [62]Also notice that a recent extension of SCAN by Ref. 99 breaks the U(1)×SU(2) gauge invariance and, by recovering the standard SCAN on collinear solutions, it does not resolve the cases discussed below.).
  42. [63]A. Erba, J. K. Desmarais, S. Casassa, B. Civalleri, L. Doná, I. J. Bush, B. Searle, L. Maschio, L. Edith-Daga, A. Cossard, et al., J. Chem. Theor. Comput. 19, 6891 (2023).
  43. [64]J. K. Desmarais, A. Erba, and R. Dovesi, Theor. Chem. Acc. 137, 1 (2018).
  44. [65]A. Seidl, A. Görling, P. Vogl, J. A. Majewski, and M. Levy, Phys. Rev. B 53, 3764 (1996).
  45. [66]R. Neumann, R. H. Nobes, and N. C. Handy, Mol. Phys. 87, 1 (1996).
  46. [67]J. K. Desmarais, G. Ambrogio, G. Vignale, A. Erba, and S. Pittalis, Physical Review Materials 8, 013802 (2024).
  47. [68]C. Chen, Y. Idzerda, H.-J. Lin, N. Smith, G. Meigs, E. Chaban, G. Ho, E. Pellegrin, and F. Sette, Physical review letters 75, 152 (1995).
  48. [69]R. Reck and D. Fry, Physical Review 184, 492 (1969).
  49. [70]A. Scherz, Spin Dependent X-ray Absorption Spectroscopy of 3d Transition Metals: Systematics and Applications (PhD dissertation, University of Berlin, 2004).
  50. [71]R. Moon, Physical Review 136, A195 (1964).
  51. [72]S. H. Vosko, L. Wilk, and M. Nusair, Canadian Journal of physics 58, 1200 (1980).
  52. [73]H. R. Larsson, H. Zhai, C. J. Umrigar, and G. K.-L. Chan, Journal of the American Chemical Society 144, 15932 (2022).
  53. [74]Y. Zhang, W. Zhang, and D. J. Singh, Physical Chemistry Chemical Physics 22, 19585 (2020).
  54. [75]S. Weber, B. Rao, P. Jena, V. Stepanyuk, W. Hergert, K. Wildberger, R. Zeller, and P. Dederichs, Journal of Physics: Condensed Matter 9, 10739 (1997).
  55. [76]Y. B. Kudasov and A. Korshunov, Physics Letters A 364, 348 (2007).
  56. [77]Y. Zhang, J. Furness, R. Zhang, Z. Wang, A. Zunger, and J. Sun, Physical Review B 102, 045112 (2020).
  57. [78]A. Annaberdiyerv, S. Mandal, L. Mitas, J. T. Krogel, and P. Ganesh, npj Quantum Materials 8, 50 (2023).
  58. [79]W. Roth, Physical Review 110, 1333 (1958).
  59. [80]P. Battle and A. Cheetham, Journal of Physics C: Solid State Physics 12, 337 (1979).
  60. [81]H. Fjellvåg, F. Grønvold, S. Stølen, and B. Hauback, Journal of Solid State Chemistry 124, 52 (1996).
  61. [82]D. Khan and R. Erickson, Physical Review B 1, 2243 (1970).
  62. [83]D. Herrmann-Ronzaud, P. Burlet, and J. Rossat-Mignod, Journal of Physics C: Solid State Physics 11, 2123 (1978).
  63. [84]W. Jauch, M. Reehüs, H. Bleif, F. Kubanek, and P. Pattison, Physical Review B 64, 052102 (2001).
  64. [85]V. Fernandez, C. Vettier, F. De Bergevin, C. Giles, and W. Neubeck, Physical Review B 57, 7870 (1998).
  65. [86]A. Cheetham and D. Hope, Physical Review B 27, 6964 (1983).
  66. [87]W. Neubeck, C. Vettier, V. Fernandez, F. De Bergevin, and C. Giles, Journal of applied physics 85, 4847 (1999).
  67. [88]H. Bowen, D. Adler, and B. Auker, Journal of Solid State Chemistry 12, 355 (1975).
  68. [89]M. Gvishi and D. Tannhauser, Solid State Communications 8, 485 (1970).
  69. [90]Y. M. Ksendzov and I. Drabkin, SOVIET PHYS SOLID STATE 7, 1519 (1965).
  70. [91]All calculations were performed with a spin-polarized starting guess from an antiferromagnetically coupled superposition of atomic Hartree Fock spin-densities.
  71. [94]X. Li and J. Paldus, The Journal of chemical physics 129 (2008).
  72. [95]M.-T. Huebsch, F. Tran, and M. Marsman, Capturing spin-torque effects with a semilocal exchange-correlation functional (2025), arXiv:2501.04124.arxiv.org/abs/2501.04124
  73. [96]G. Scalmani and M. J. Frisch, J. Chem. Theory and Comput. 8, 2193 (2012).
  74. [97]N. Tancogne-Dejean, A. Rubio, and C. A. Ullrich, Physical Review B 107, 165111 (2023).
  75. [98]Z. Pu, H. Li, N. Zhang, H. Jiang, Y. Gao, Y. Xiao, Q. Sun, Y. Zhang, and S. Shao, Physical Review Research 5, 013036 (2023).

Paper details

vlnc:W_synjrqay
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Physics and Astronomy

Not recorded for this paper

View statement
Looking for related papers…
Contents