HDR/Manuscript/Chapter3/chapter3.2.tex
2021-05-20 15:13:48 +02:00

380 lines
23 KiB
TeX

%****************************************************************
\section{
\label{sec:ExGLDA}
Exchange functionals based on finite uniform electron gases}
%****************************************************************
In this section, we show how to use these finite UEGs (FUEGs) to create a new type of exchange functionals applicable to atoms, molecules and solids \cite{ExGLDA17}.
We have successfully applied this strategy to one-dimensional systems \cite{1DChem15, SBLDA16, Leglag17, ESWC17}, for which we have created a correlation functional based on this idea \cite{gLDA14, Wirium14}.
%****************************************************************
\subsection{Theory}
%****************************************************************
Within DFT, one can write the total exchange energy as the sum of its spin-up ($\sigma=$ $\upa$) and spin-down ($\sigma=$ $\dwa$) contributions:
\begin{equation}
E_\text{x} = E_{\text{x},\upa} + E_{\text{x},\dwa},
\end{equation}
where
\begin{equation}
E_{\text{x},\sigma} = \int e_{\text{x},\sigma}(\rho_\sigma,\nabla \rho_\sigma, \tau_\sigma, \ldots) \, \rho_\sigma(\br) \, d\br,
\end{equation}
and $\rs$ is the electron density of the spin-$\sigma$ electrons.
Although, for sake of simplicity, we sometimes remove the subscript $\sigma$, we only use spin-polarised quantities from hereon.
The first-rung LDA exchange functional (or D30 \cite{Dirac30}) is based on the IUEG \cite{WIREs16} and reads
\begin{equation}
\exsLDA(\rs) = \CxLDA \rs^{1/3},
\end{equation}
where
\begin{equation}
\CxLDA = - \frac{3}{2} \qty( \frac{3}{4\pi} )^{1/3}.
\end{equation}
A GGA functional (second rung) is defined as
\begin{equation}
\label{eq:GGA}
\exsGGA(\rs,\xs)
= e_{\text{x},\sigma}^\text{LDA}(\rs) \FxGGA(\xs),
\end{equation}
where $\FxGGA$ is the GGA enhancement factor depending only on the reduced gradient
\begin{equation}
x = \frac{\abs{\nabla \rho}}{\rho^{4/3}},
\end{equation}
and
\begin{equation}
\lim_{x \to 0} \FxGGA(x) = 1,
\end{equation}
i.e.~a well thought-out GGA functional reduces to the LDA for homogeneous systems.
Similarly, motivated by the work of Becke \cite{Becke00} and our previous investigations \cite{gLDA14, Wirium14}, we define an alternative second-rung functional (see Fig.~\ref{fig:Jacob-rev}) that we call generalised LDA (GLDA)
\begin{equation}
\label{eq:GLDA}
\exsGLDA(\rs,\as) = \exsLDA(\rs) \FxGLDA(\as).
\end{equation}
%%% FIGURE 1 %%%
\begin{figure}
\centering
\includegraphics[width=0.5\linewidth]{../Chapter3/fig/Jacob-rev}
\caption{
\label{fig:Jacob-rev}
Jacob's ladder of DFT revisited.}
\end{figure}
%%%
By definition, a GLDA functional only depends on the electron density and the curvature of the Fermi hole (see Fig.~\ref{fig:Jacob-rev}):
\begin{equation}
\label{eq:eta-def}
\alpha = \frac{\tau - \tau_\text{W}}{\tau_\text{IUEG}} = \frac{\tau}{\tau_\text{IUEG}} - \frac{x^2}{4 \Cf},
\end{equation}
which measures the tightness of the exchange hole around an electron \cite{Becke83, Dobson91}.
In Eq.~\eqref{eq:eta-def},
\begin{equation}
\tau_\text{W} = \frac{\abs{\nabla\rho}^2}{4\,\rho}
\end{equation}
is the von Weizs{\"a}cker kinetic energy density \cite{vonWeizsacker35}, and
\begin{equation}
\tau_\text{IUEG} = \Cf \rho^{5/3}
\end{equation}
is the kinetic energy density of the IUEG \cite{WIREs16}, where
\begin{equation}
\Cf = \frac{3}{5} (6\pi^2)^{2/3}.
\end{equation}
The dimensionless parameter $\alpha$ has two characteristic features: i) $\alpha=0$ for any one-electron system, and ii) $\alpha=1$ for the IUEG.
Some authors call $\alpha$ the inhomogeneity parameter but we will avoid using this term as we are going to show that $\alpha$ can have distinct values in homogeneous systems.
For well-designed GLDA functionals, we must ensure that
\begin{equation}
\label{eq:limGLDA}
\lim_{\alpha \to 1} \FxGLDA(\alpha) = 1,
\end{equation}
i.e.~the GLDA reduces to the LDA for the IUEG.\footnote{While some functionals only use the variable $\tau$ \cite{Ernzerhof99, Eich14}, we are not aware of any functional only requiring $\alpha$.}
Although any functional depending on the reduced gradient $x$ and the kinetic energy density $\tau$ is said to be of MGGA type, here we will define a third-rung MGGA functional as depending on $\rho$, $x$ and $\alpha$:
\begin{equation}
\exsMGGA(\rs,\xs,\as) = \exsLDA(\rs) \FxMGGA(\xs,\as),
\end{equation}
where one should ensure that
\begin{equation}
\label{eq:limMGGA}
\lim_{x \to 0} \lim_{\alpha \to 1} \FxMGGA(x,\alpha) = 1,
\end{equation}
i.e.~the MGGA reduces to the LDA for an infinite homogeneous system.
The Fermi hole curvature $\alpha$ has been shown to be a better variable than the kinetic energy density $\tau$ as one can discriminate between covalent ($\alpha=0$), metallic ($\alpha \approx 1$) and weak bonds ($\alpha \gg 0$) \cite{Kurth99, TPSS, revTPSS, MS0, Sun13a, MS1_MS2, MVS, SCAN, Sun16b}.
The variable $\alpha$ is also related to the electron localisation function (ELF) designed to identify chemical bonds in molecules \cite{Becke83, ELF}.
Moreover, by using the variables $x$ and $\alpha$, we satisfy the correct uniform coordinate density-scaling behaviour \cite{Levy85}.
In conventional MGGAs, the dependence in $x$ and $\alpha$ can be strongly entangled, while, in GGAs for example, $\rho$ and $x$ are strictly disentangled as illustrated in Eq.~\eqref{eq:GGA}.
Therefore, it feels natural to follow the same strategy for MGGAs.
Thus, we consider a special class of MGGA functionals (rung $2.9$ in Fig.~\ref{fig:Jacob-rev}) that we call factorable MGGAs (FMGGAs)
\begin{equation}
\exsFMGGA(\rs,\xs,\as) = \exsLDA(\rs) \FxFMGGA(\xs,\as),
\end{equation}
where the enhancement factor is written as
\begin{equation}
\FxFMGGA(x,\alpha) = \FxGGA(x) \FxGLDA(\alpha).
\end{equation}
By construction, $\FxFMGGA$ fulfills Eq.~\eqref{eq:limMGGA} and the additional physical limits
\begin{subequations}
\begin{align}
\lim_{x \to 0} \FxFMGGA(x,\alpha) & = \FxGLDA(\alpha),
\\
\lim_{\alpha \to 1} \FxFMGGA(x,\alpha) & = \FxGGA(x).
\end{align}
\end{subequations}
The MVS functional designed by Sun, Perdew and Ruzsinszky is an example of FMGGA functional \cite{MVS}.
Unless otherwise stated, all calculations have been performed self-consistently with a development version of the Q-Chem4.4 package \cite{qchem4} using the aug-cc-pVTZ basis set \cite{Dunning89, Kendall92, Woon93, Woon94, Woon95, Peterson02}.
To remove quadrature errors, we have used a very large quadrature grids consisting of 100 radial points (Euler-MacLaurin quadrature) and 590 angular points (Lebedev quadrature).
As a benchmark, we have calculated the (exact) unrestricted HF (UHF) exchange energies.
%****************************************************************
\subsection{GLDA exchange functionals}
%****************************************************************
As stated in the previous section, the orbitals for an electron on a 3-sphere of unit radius are the normalised hyperspherical harmonics $Y_{\ell\mu}$, where $\ell$ is the principal quantum number and $\mu$ is a composite index of the remaining two quantum numbers \cite{Avery, Avery93}.
We confine our attention to ferromagnetic (i.e.~spin-polarised) systems in which each orbital with $\ell = 0, 1, \ldots , L_{\sigma}$ is occupied by one spin-up or spin-down electron.
As mentioned in the previous section, this yields an electron density that is uniform over the surface of the sphere.
Note that the present paradigm is equivalent to the jellium model \cite{WIREs16} for $L_{\sigma} \to \infty$.
We refer the reader to Ref.~\cite{Glomium11} for more details about this paradigm.
The number of spin-$\sigma$ electrons is
\begin{equation}
\ns = \frac{1}{3} (\Ls+1)(\Ls+3/2)(\Ls+2),
\end{equation}
and their one-electron uniform density around the 3-sphere is
\begin{equation}
\rs = \frac{\ns}{V} = \frac{(\Ls+2)(\Ls+3/2)(\Ls+1)}{6 \pi^2 R^3},
\end{equation}
where $V = 2 \pi^2 R^3$ is the surface of a 3-sphere of radius $R$.
Moreover, using Eq.~\eqref{eq:eta-def}, one can easily derive that \cite{gLDA14, Wirium14}
\begin{equation}
\label{eq:alpha}
\as = \frac{\Ls(\Ls+3)}{\qty[ (\Ls+1)(\Ls+3/2)(\Ls+2) ]^{2/3}},
\end{equation}
which yields
\begin{align}
\lim_{\ns \to 1 } \as & = 0,
&
\lim_{\ns \to \infty } \as & = 1.
\end{align}
We recover the results that $\alpha = 0$ in a one-electron system (here a one-electron FUEG), and that $\alpha = 1$ in the IUEG.
In particular, we have shown that the exchange energy of these systems can be written as \cite{Glomium11, Jellook12}
\begin{equation}
\label{eq:Ex}
E_{\text{x},\sigma}(\Ls) = \Cx(\Ls) \int \rs^{4/3} d\bm{r}.
\end{equation}
where
\begin{equation}
\label{eq:CxGLDA}
\Cx(L)
= \CxLDA \frac{\frac{1}{2} \qty( L+\frac{5}{4}) \qty(L+\frac{7}{4}) \qty[\frac{1}{2} H_{2 L+\frac{5}{2}} + \ln 2]
+ \qty(L+\frac{3}{2})^2 \qty(L^2+3 L+\frac{13}{8})}
{\qty[(L+1) \qty(L+\frac{3}{2}) (L+2)]^{4/3}}
\end{equation}
and $H_{k}$ is an harmonic number \cite{NISTbook}.
Therefore, thanks to the one-to-one mapping between $\Ls$ and $\as$ evidenced by Eq.~\eqref{eq:alpha}, we have created the \gX~functional
\begin{equation}
\label{eq:FxGLDA}
\FxgX(\alpha) = \frac{\CxGLDA(0)}{\CxGLDA(1)}
\\
+ \alpha \frac{c_0+c_1\,\alpha}{1+(c_0+c_1-1)\alpha} \qty[ 1 - \frac{\CxGLDA(0)}{\CxGLDA(1)} ],
\end{equation}
where $c_0 = +0.827 411$, $c_1 = -0.643 560$, and
\begin{align}
\CxGLDA(1) & = \CxLDA = - \frac{3}{2} \qty( \frac{3}{4\pi} )^{1/3},
\\
\CxGLDA(0) & = -\frac{4}{3} \qty(\frac{2}{\pi})^{1/3}.
\end{align}
The parameters $c_0$ and $c_1$ of the \gX~enhancement factor \eqref{eq:FxGLDA} have been obtained by fitting the exchange energies of these FUEGs for $ 1 \le L \le 10$ given by Eq.~\eqref{eq:Ex}.
$\FxgX$ automatically fulfils the constraint given by Eq.~\eqref{eq:limGLDA}.
Moreover, because $ 1\le \FxgX \le 1.233$, it breaks only slightly the tight Lieb-Oxford bound \cite{Lieb81, Chan99, Odashima09} $\Fx < 1.174$ derived by Perdew and coworkers for two-electron systems \cite{Perdew14, Sun16a}.
This is probably due to the non-zero curvature of these FUEGs.
Albeit very simple, the functional form \eqref{eq:FxGLDA} is an excellent fit to Eq.~\eqref{eq:CxGLDA}.
In particular, $\FxgX$ is linear in $\alpha$ for small $\alpha$, which is in agreement with Eq.~\eqref{eq:CxGLDA} \cite{Glomium11}.
Also, Eq.~\eqref{eq:CxGLDA} should have an infinite derivative at $\alpha=1$ and approached as $\sqrt{1-\alpha} \ln(1-\alpha)$.
Equation \eqref{eq:FxGLDA} does not behave that way.
However, it has a marginal impact on the numerical results.
As one can see in Fig.~\ref{fig:FxGLDA}, albeit being created with FUEGs, the \gX~functional has a fairly similar form to the common MGGA functionals, such as MS0 \cite{MS0}, MS1 \cite{MS1_MS2}, MS2 \cite{MS1_MS2}, MVS \cite{MVS}, and SCAN \cite{SCAN} for $ 0 \le \alpha \le 1$.
This is good news for DFT as it shows that we recover functionals with similar physics independently of the paradigm used to design them.
However, around $\alpha \approx 1$, the behaviour of $\FxgX$ is very different from other MGGAs (except for MVS) due to the constraint of the second-order gradient expansion (which is not satisfied in our case) \cite{Ma68}.
For $ 0 \le \alpha \le 1$, it is also instructive to note that the \gX~functional is an upper bound of all the MGGA functionals.
Taking into account the inhomogeneity of the system via the introduction of $x$ should have the effect of decreasing the MGGA enhancement factor (at least for $0 \le \alpha \le 1$).
Unlike other functionals, we follow a rather different approach and guide our functional between $\alpha=0$ and $1$ using FUEGs.
For example, the MS0 functional uses the exact exchange energies of non-interacting hydrogenic anions to construct the functional from $\alpha = 0$ to $1$ \cite{Staroverov04, MS0}, while revTPSS has no constraint to guide itself for this range of $\alpha$ \cite{revTPSS}.
Nonetheless, because these uniform systems only give valuable information in the range $0 \le \alpha \le 1$, we must find a different way to guide our functional for $\alpha > 1$.\footnote{Except for one- and two-electron systems, any atomic and molecular systems has region of space with $\alpha_\sigma > 1$, as discussed in details by Sun et al.\cite{Sun13a}}
To do so, we have extended the \gX~functional beyond $\alpha = 1$ using a simple one-parameter extrapolation:
\begin{equation}
\label{eq:FxGMVS}
\FxGX(\alpha) =
\begin{cases}
\FxgX(\alpha), & 0 \le \alpha \le 1,
\\
1 + (1-\alpha_\infty) \frac{1-\alpha}{1+\alpha}, & \alpha > 1,
\end{cases}
\end{equation}
where $\alpha_\infty$ is an adjustable parameter governing the value of $\FxGX$ when $\alpha \to \infty$.
For large $\alpha$, $\FxGX$ converges to $\alpha_\infty$ as $\alpha^{-1}$, similarly to the MVS functional \cite{MVS}.
Far from claiming that this choice is optimal, we have found that the simple functional form \eqref{eq:FxGMVS} for $\alpha > 1$ yields satisfactory results (see below).
%%% FIG 2 %%%
\begin{figure*}
\centering
\includegraphics[width=0.4\linewidth]{../Chapter3/fig/fig2a}
\includegraphics[width=0.4\linewidth]{../Chapter3/fig/fig2b}
\caption{
\label{fig:FxGLDA}
Enhancement factors $\FxGLDA(\alpha)$ or $\FxMGGA(x=0,\alpha)$ as a function of $\alpha$ for various GLDA and MGGA exchange functionals.
The TPSS functional is represented as a dot-dashed line, the MS family of functionals (MS0, MS1 and MS2) are represented as dashed lines, while the MVS and SCAN functionals are depicted with solid lines.
The new functionals \gX~and \PBEGX~are represented with thick black lines.
Note that $\FxgX(\alpha) = \FxPBEGX(0,\alpha)$ for $ 0 \le \alpha \le 1$.
For $\FxPBEGX$, $\alpha_\infty = +0.852$.}
\end{figure*}
%%% %%%
Following the seminal work of Sham \cite{Sham71} and Kleinman \cite{Kleinman84, Antoniewicz85, Kleinman88} (see also Ref.~\cite{Svendsen96}), it is also possible, using linear response theory, to derive a second-order gradient-corrected functional.
However, it does not provide any information for $\alpha >1$.
The performance of the \GX~functional is illustrated in Table \ref{tab:atoms}.
Although \GX~is an improvement compared to LDA, even for one- and two-electron systems, we observe that the \GX~functional cannot compete with GGAs and MGGAs in terms of accuracy.
%%% TABLE 1 %%%
\begin{table*}
\centering
\caption{
\label{tab:atoms}
Reduced (i.e.~per electron) mean error (ME) and mean absolute error (MAE) (in kcal/mol) of the error (compared to UHF) in the exchange energy of the hydrogen-like ions, helium-like ions and first 18 neutral atoms for various LDA, GGA, GLDA, FMGGA and MGGA functionals.
For the hydrogen-like ions, the exact density has been used for all calculations.}
\begin{tabular}{llcccccc}
\hline
& & \mc{2}{c}{hydrogen-like ions} & \mc{2}{c}{helium-like ions} & \mc{2}{c}{neutral atoms} \\
\cline{3-4} \cline{5-6} \cline{7-8}
& & ME & MAE & ME & MAE & ME & MAE \\
\hline
LDA & D30 & $153.5$ & $69.7$ & $150.6$ & $69.5$ & $70.3$ & $9.1$ \\
GGA & B88 & $9.5$ & $4.3$ & $9.3$ & $4.7$ & $2.8$ & $0.5$ \\
& G96 & $4.4$ & $2.0$ & $4.4$ & $2.2$ & $2.1$ & $0.5$ \\
& PW91 & $19.4$ & $8.8$ & $19.1$ & $9.3$ & $4.5$ & $0.8$ \\
& PBE & $22.6$ & $10.3$ & $22.3$ & $10.7$ & $7.4$ & $0.6$ \\
GLDA & \GX & $61.8$ & $123.5$ & $61.0$ & $122.0$ & --- & --- \\
FMGGA & MVS & $0.0$ & $0.0$ & $0.3$ & $0.2$ & $2.7$ & $0.9$ \\
& \PBEGX & $0.0$ & $0.0$ & $0.7$ & $0.4$ & $1.0$ & $1.1$ \\
MGGA & M06-L & $44.4$ & $88.8$ & $12.0$ & $24.0$ & $4.2$ & $2.9$ \\
& TPSS & $0.0$ & $0.0$ & $0.7$ & $0.4$ & $0.7$ & $1.1$ \\
& revTPSS & $0.0$ & $0.0$ & $0.5$ & $0.3$ & $3.5$ & $2.5$ \\
& MS0 & $0.0$ & $0.0$ & $0.4$ & $0.2$ & $1.3$ & $2.4$ \\
& SCAN & $0.0$ & $0.0$ & $0.3$ & $0.2$ & $1.2$ & $1.6$ \\
\hline
\end{tabular}
\end{table*}
%%%
%****************************************************************
\subsection{FMGGA exchange functionals}
%****************************************************************
One of the problem of GLDA functionals is that they cannot discriminate between homogeneous and inhomogeneous one-electron systems, for which we have $\alpha = 0$ independently of the value of the reduced gradient $x$.
For example, the \GX~functional is exact for one-electron FUEGs, while it is inaccurate for the hydrogen-like ions.
Unfortunately, it is mathematically impossible to design a GLDA functional exact for these two types of one-electron systems.
To cure this problem, we couple the \GX~functional designed above with a GGA enhancement factor to create a FMGGA functional.
We have chosen a PBE-like GGA factor, i.e.
\begin{equation}
\FxPBEGX(x,\alpha) =\Fx^\text{PBE}(x) \FxGX(\alpha),
\end{equation}
where
\begin{equation}
\label{eq:FxPBEGX}
\Fx^\text{PBE}(x) = \frac{1}{1+\mu\,x^2}.
\end{equation}
Similarly to various MGGAs (such as TPSS \cite{TPSS}, MVS \cite{MVS}, or SCAN \cite{SCAN}), we use the hydrogen atom as a ``norm'', and determine that $\mu = +0.001 015 549$ reproduces the exact exchange energy of the ground state of the hydrogen atom.
Also, we have found that $\alpha_\infty = +0.852$ yields excellent exchange energies for the first 18 neutral atoms.
Unlike \GX, \PBEGX~is accurate for both the (inhomogeneous) hydrogen-like ions and the (homogeneous) one-electron FUEGs, and fulfils the negativity constraint and uniform density scaling \cite{SCAN, Perdew16}.
The right graph of Fig.~\ref{fig:FxGLDA} shows the behaviour of the MGGA enhancement factor for $x = 0$ as a function of $\alpha$.
Looking at the curves for $\alpha > 1$, we observe that TPSS has a peculiar enhancement factor which slowly raises as $\alpha$ increases.
All the other functionals (including \PBEGX) decay more or less rapidly with $\alpha$.
We note that \PBEGX~and MVS behave similarly for $\alpha > 1$, though their functional form is different.
%%% FIG 3 %%%
\begin{figure}
\centering
\includegraphics[width=0.6\linewidth]{../Chapter3/fig/fig3}
\caption{
\label{fig:FxGGA}
Enhancement factors $\FxGGA(x)$ or $\FxMGGA(x,\alpha=1)$ as a function of $x$ for various GGA, FMGGA and MGGA exchange functionals.
The GGA functionals are represented in solid lines, while MGGAs are depicted in dashed lines.
The new functional \PBEGX~is represented with a thick black line.}
\end{figure}
%%% %%%
Figure \ref{fig:FxGGA} evidences a fundamental difference between GGAs and MGGAs: while the enhancement factor of conventional GGAs does increase monotonically with $x$ and favour inhomogeneous electron densities, $\FxMGGA$ decays monotonically with respect to $x$.
This is a well-known fact: the $x$- and $\alpha$-dependence are strongly coupled, as suggested by the relationship \eqref{eq:eta-def}.
Therefore, the $x$-dependence can be sacrificed if the $\alpha$-dependence is enhanced \cite{MS0, MVS, SCAN}.
Similarly to $\FxPBEGX$, $\Fx^\text{MVS}$ and $\Fx^\text{SCAN}$ decay monotonically with $x$ (although not as fast as \PBEGX), while earlier MGGAs such as TPSS and MS0 have a slowly-increasing enhancement factor.
We have observed that one needs to use a bounded enhancement factor at large $x$ (as in Eq.~\eqref{eq:FxPBEGX}) in order to be able to converge self-consistent field (SCF) calculations.
Indeed, using an unbounded enhancement factor (as in B88 \cite{B88} or G96 \cite{G96}) yields divergent SCF KS calculations.
Finally, we note that, unlike TPSS, \PBEGX~does not suffer from the order of limits problem \cite{regTPSS}.
%%% TABLE 2 %%%
\begin{table}
\centering
\caption{
\label{tab:molecules}
Reduced (i.e.~per electron) mean error (ME) and mean absolute error (MAE) (in kcal/mol) of the error (compared to the experimental value) in the atomisation energy ($E_\text{atoms} - E_\text{molecule}$) of diatomic molecules at experimental geometry for various LDA, GGA and MGGA exchange-correlation functionals.
Experimental geometries are taken from Ref.~\cite{HerzbergBook}.}
\begin{tabular}{lllcc}
\hline
& \mc{2}{c}{functional} & \mc{2}{c}{diatomics} \\
\cline{2-3} \cline{4-5}
& exchange & correlation & ME & MAE \\
\hline
LDA & D30 & VWN5 & $1.8$ & $3.7$ \\
GGA & B88 & LYP & $0.6$ & $1.2$ \\
& PBE & PBE & $0.7$ & $1.2$ \\
MGGA & M06-L & M06-L & $0.4$ & $0.7$ \\
& TPSS & TPSS & $0.6$ & $1.1$ \\
& revTPSS & revTPSS & $0.6$ & $1.2$ \\
& MVS & regTPSS & $0.5$ & $0.9$ \\
& SCAN & SCAN & $0.4$ & $0.7$ \\
& \PBEGX & PBE & $0.6$ & $1.2$ \\
& \PBEGX & regTPSS & $0.6$ & $1.1$ \\
& \PBEGX & LYP & $0.6$ & $1.1$ \\
& \PBEGX & TPSS & $0.7$ & $1.3$ \\
& \PBEGX & revTPSS & $0.8$ & $1.5$ \\
& \PBEGX & SCAN & $0.6$ & $1.0$ \\
\hline
\end{tabular}
\end{table}
%%%
%%% FIG 4 %%%
\begin{figure}
\centering
\includegraphics[width=0.6\linewidth]{../Chapter3/fig/fig4}
\caption{
\label{fig:error}
Reduced (i.e.~per electron) error (in kcal/mol) in atomic exchange energies of the first 18 neutral atoms of the periodic table for the B88 (red), TPSS (blue), MVS (orange), SCAN (purple) and \PBEGX~(thick black) functionals.}
\end{figure}
%%% %%%
How good are FMGGAs?
This is the question we would like to answer here.
In other word, we would like to know whether or not our new simple FMGGA functional called \PBEGX~is competitive within MGGAs.
Unlike GGAs and some of the MGGAs (like M06-L), by construction, \PBEGX~reproduces exactly the exchange energy of the hydrogen atom and the hydrogenic ions (\ce{He^+}, \ce{Li^2+}, \ldots) due to its dimensional consistency (see Table \ref{tab:atoms}).
\PBEGX~also reduces the error for the helium-like ions (\ce{H^-}, \ce{He}, \ce{Li^+}, \ldots) by one order of magnitude compared to GGAs, and matches the accuracy of MGGAs.
For the first 18 neutral atoms (Table \ref{tab:atoms} and Fig.~\ref{fig:error}), \PBEGX~is as accurate as conventional MGGAs with a mean error (ME) and mean absolute error (MAE) of $1.0$ and $1.1$ kcal/mol.
From the more conventional MGGAs, the TPSS and SCAN functionals are the best performers for neutral atoms with MEs of $0.7$ and $1.2$ kcal/mol, and MAEs of $1.1$ and $1.6$ kcal/mol.
\PBEGX~lies just in-between these two MGGAs.
We now turn our attention to diatomic molecules for which errors in the atomisation energy ($E_\text{atoms} - E_\text{molecule}$) are reported in Table \ref{tab:molecules} for various combinations of exchange and correlation functionals.
In particular, we have coupled our new \PBEGX~exchange functional with the PBE \cite{PBE}, regTPSS \cite{regTPSS} (also called vPBEc) and LYP \cite{LYP} GGA correlation functionals, as well as the TPSS, \cite{TPSS} revTPSS \cite{revTPSS} and SCAN \cite{SCAN} MGGA correlation functionals.
Although very lightly parametrised on atoms, \PBEGX~is also accurate for molecules.
Interestingly, the results are mostly independent of the choice of the correlation functional with MEs ranging from $0.6$ and $0.8$ kcal/mol, and MAEs from $1.0$ and $1.5$ kcal/mol.
\PBEGX~is only slightly outperformed by the SCAN functional and the highly-parametrized M06-L functional, which have both a ME of $0.4$ kcal/mol and a MAE of $0.7$ kcal/mol.