1109 lines
58 KiB
TeX
1109 lines
58 KiB
TeX
\documentclass[aip,jcp,reprint,noshowkeys,superscriptaddress]{revtex4-1}
|
|
\usepackage{graphicx,dcolumn,bm,xcolor,microtype,multirow,amsmath,amssymb,amsfonts,physics,mhchem,xspace,subfigure}
|
|
|
|
\usepackage[T1]{fontenc}
|
|
\usepackage[utf8]{inputenc}
|
|
|
|
\usepackage{txfonts}
|
|
|
|
\usepackage[
|
|
colorlinks=true,
|
|
citecolor=blue,
|
|
breaklinks=true
|
|
]{hyperref}
|
|
\urlstyle{same}
|
|
|
|
\newcommand{\ie}{\textit{i.e.}}
|
|
\newcommand{\eg}{\textit{e.g.}}
|
|
\newcommand{\alert}[1]{\textcolor{red}{#1}}
|
|
\definecolor{darkgreen}{HTML}{009900}
|
|
\usepackage[normalem]{ulem}
|
|
\newcommand{\toto}[1]{\textcolor{blue}{#1}}
|
|
\newcommand{\trashAS}[1]{\textcolor{blue}{\sout{#1}}}
|
|
\newcommand{\titou}[1]{\textcolor{red}{#1}}
|
|
\newcommand{\trashPFL}[1]{\textcolor{red}{\sout{#1}}}
|
|
\newcommand{\PFL}[1]{\titou{(\underline{\bf PFL}: #1)}}
|
|
|
|
\newcommand{\mc}{\multicolumn}
|
|
\newcommand{\fnm}{\footnotemark}
|
|
\newcommand{\fnt}{\footnotetext}
|
|
\newcommand{\tabc}[1]{\multicolumn{1}{c}{#1}}
|
|
|
|
\newcommand{\EPT}{E_{\text{PT2}}}
|
|
\newcommand{\EDMC}{E_{\text{FN-DMC}}}
|
|
\newcommand{\Ndet}{N_{\text{det}}}
|
|
\newcommand{\hartree}{$E_h$}
|
|
|
|
\newcommand{\LCT}{Laboratoire de Chimie Th\'eorique (UMR 7616), Sorbonne Universit\'e, CNRS, Paris, France}
|
|
\newcommand{\ANL}{Argonne Leadership Computing Facility, Argonne National Laboratory, Argonne, IL 60439, United States}
|
|
\newcommand{\LCPQ}{Laboratoire de Chimie et Physique Quantiques (UMR 5626), Universit\'e de Toulouse, CNRS, UPS, France}
|
|
|
|
|
|
\begin{document}
|
|
|
|
\title{Taming the fixed-node error in diffusion Monte Carlo via range separation}
|
|
%\title{Enabling high accuracy diffusion Monte Carlo calculations with
|
|
% range-separated density functional theory and selected configuration interaction}
|
|
|
|
\author{Anthony Scemama}
|
|
\email{scemama@irsamc.ups-tlse.fr}
|
|
\affiliation{\LCPQ}
|
|
\author{Emmanuel Giner}
|
|
\email{emmanuel.giner@lct.jussieu.fr}
|
|
\affiliation{\LCT}
|
|
\author{Anouar Benali}
|
|
\email{benali@anl.gov}
|
|
\affiliation{\ANL}
|
|
\author{Pierre-Fran\c{c}ois Loos}
|
|
\email{loos@irsamc.ups-tlse.fr}
|
|
\affiliation{\LCPQ}
|
|
|
|
|
|
\begin{abstract}
|
|
\end{abstract}
|
|
|
|
\maketitle
|
|
|
|
|
|
\section{Introduction}
|
|
\label{sec:intro}
|
|
|
|
Solving the Schr\"odinger equation for atoms and molecules is a complex task that has kept theoretical and computational chemists busy for almost hundred years now. \cite{Schrodinger_1926}
|
|
In order to achieve this formidable endeavour, various strategies have been carefully designed and implemented in quantum chemistry software packages.
|
|
|
|
One of this strategies consists in relying on wave function theory and, in particular, on the full configuration interaction (FCI) method.
|
|
However, FCI delivers only an approximate solution of the Schr\"odinger equation within a finite one-electron basis.
|
|
This solution is the eigenpair of an approximate Hamiltonian defined as
|
|
the projection of the exact Hamiltonian onto the finite many-electron basis of
|
|
all possible Slater determinants generated within this finite one-electron basis.
|
|
The FCI wave function can be interpreted as a constrained solution of the
|
|
true Hamiltonian forced to span the restricted space provided by the one-electron basis.
|
|
In the complete basis set (CBS) limit, the constraint is lifted and the
|
|
exact solution is recovered.
|
|
Hence, the accuracy of a FCI calculation can be systematically improved by increasing the size of the one-electron basis set.
|
|
Nevertheless, its exponential scaling with the number of electrons and with the size of the basis is prohibitive for most chemical systems.
|
|
In recent years, the introduction of new algorithms \cite{Booth_2009} and the
|
|
revival \cite{Abrams_2005,Bytautas_2009,Roth_2009,Giner_2013,Knowles_2015,Holmes_2016,Liu_2016,Garniron_2018}
|
|
of selected configuration interaction (SCI)
|
|
methods \cite{Bender_1969,Huron_1973,Buenker_1974} pushed the limits of
|
|
the sizes of the systems that could be computed at the FCI level. \cite{Booth_2010,Cleland_2010,Daday_2012,Chien_2018,Loos_2018a,Loos_2019,Loos_2020b,Loos_2020c}
|
|
However, the scaling remains exponential unless some bias is introduced leading
|
|
to a loss of size consistency. \cite{Evangelisti_1983,Cleland_2010,Tenno_2017}
|
|
|
|
Diffusion Monte Carlo (DMC) is another numerical scheme to obtain
|
|
the exact solution of the Schr\"odinger equation with a different
|
|
constraint. In DMC, the solution is imposed to have the same nodes (or zeroes)
|
|
as a given trial (approximate) wave function.
|
|
Within this so-called \emph{fixed-node} (FN) approximation,
|
|
the FN-DMC energy associated with a given trial wave function is an upper
|
|
bound to the exact energy, and the latter is recovered only when the
|
|
nodes of the trial wave function coincide with the nodes of the exact
|
|
wave function.
|
|
The polynomial scaling with the number of electrons and with the size
|
|
of the trial wave function makes the FN-DMC method particularly attractive.
|
|
In addition, the total energies obtained are usually far below
|
|
those obtained with the FCI method in computationally tractable basis
|
|
sets because the constraints imposed by the FN approximation
|
|
are less severe than the constraints imposed by the finite-basis
|
|
approximation.
|
|
|
|
%However, it is usually harder to control the FN error in DMC, and this
|
|
%might affect energy differences such as atomization energies.
|
|
%Moreover, improving systematically the nodal surface of the trial wave
|
|
%function can be a tricky job as \trashAS{there is no variational
|
|
%principle for the nodes}\toto{the derivatives of the FN-DMC energy
|
|
%with respect to the variational parameters of the wave function can't
|
|
%be computed}.
|
|
|
|
The qualitative picture of the electronic structure of weakly
|
|
correlated systems, such as organic molecules near their equilibrium
|
|
geometry, is usually well represented with a single Slater
|
|
determinant. This feature is in part responsible for the success of
|
|
density-functional theory (DFT) and coupled cluster theory.
|
|
DMC with a single-determinant trial wave function can be used as a
|
|
single-reference post-Hatree-Fock method, with an accuracy comparable
|
|
to coupled cluster.\cite{Dubecky_2014,Grossman_2002}
|
|
The favorable scaling of QMC, its very low memory requirements and
|
|
its adequacy with massively parallel architectures make it a
|
|
serious alternative for high-accuracy simulations of large systems.
|
|
|
|
As it is not possible to minimize directly the FN-DMC energy with respect
|
|
to the variational parameters of the trial wave function, the
|
|
fixed-node approximation is much more difficult to control than the
|
|
finite-basis approximation.
|
|
The conventional approach consists in multiplying the trial wave
|
|
function by a positive function, the \emph{Jastrow factor}, taking
|
|
account of the electron-electron cusp and the short-range correlation
|
|
effects. The wave function is then re-optimized within variational
|
|
Monte Carlo (VMC) in the presence of the Jastrow factor and the nodal
|
|
surface is expected to be improved. Using this technique, it has been
|
|
shown that the chemical accuracy could be reached within
|
|
FN-DMC.\cite{Petruzielo_2012}
|
|
|
|
Another approach consists in considering the FN-DMC method as a
|
|
\emph{post-FCI method}. The trial wave function is obtained by
|
|
approaching the FCI with a selected configuration interaction (sCI)
|
|
method such as CIPSI for instance.\cite{Giner_2013,Caffarel_2016_2}
|
|
\titou{When the basis set is increased, the trial wave function gets closer
|
|
to the exact wave function, so the nodal surface can be systematically
|
|
improved.\cite{Caffarel_2016} WRONG}
|
|
This technique has the advantage that using FCI nodes in a given basis
|
|
set is well defined, so the calculations are reproducible in a
|
|
black-box way without needing any expertise in QMC.
|
|
But this technique cannot be applied to large systems because of the
|
|
exponential scaling of the size of the trial wave function.
|
|
Extrapolation techniques have been used to estimate the FN-DMC energies
|
|
obtained with FCI wave functions,\cite{Scemama_2018} and other authors
|
|
have used a combination of the two approaches where highly truncated
|
|
CIPSI trial wave functions are re-optimized in VMC under the presence
|
|
of a Jastrow factor to keep the number of determinants
|
|
small,\cite{Giner_2016} and where the consistency between the
|
|
different wave functions is kept by imposing a constant energy
|
|
difference between the estimated FCI energy and the variational energy
|
|
of the CI wave function.\cite{Dash_2018,Dash_2019}
|
|
|
|
Nevertheless, finding a robust protocol to obtain high accuracy
|
|
calculations which can be reproduced systematically, and which is
|
|
applicable for large systems with a multi-configurational character is
|
|
still an active field of research. The present paper falls
|
|
within this context.
|
|
|
|
|
|
\section{Combining CIPSI with range-separated DFT}
|
|
\label{sec:rsdft-cipsi}
|
|
|
|
In single-determinant DMC calculations, the degrees of freedom used to
|
|
reduce the fixed-node error are the molecular orbitals on which the
|
|
Slater determinant is built.
|
|
Different molecular orbitals can be chosen:
|
|
Hartree-Fock (HF), Kohn-Sham (KS), natural (NO) orbitals of a
|
|
correlated wave function, or orbitals optimized under the
|
|
presence of a Jastrow factor.
|
|
The nodal surfaces obtained with the KS determinant are in general
|
|
better than those obtained with the HF determinant,\cite{Per_2012} and
|
|
of comparable quality to those obtained with a Slater determinant
|
|
built with NOs.\cite{Wang_2019} Orbitals obtained in the presence
|
|
of a Jastrow factor are generally superior to KS
|
|
orbitals.\cite{Filippi_2000,Scemama_2006,HaghighiMood_2017,Ludovicy_2019}
|
|
|
|
The description of electron correlation within DFT is very different
|
|
from correlated methods.
|
|
In DFT, one solves a mean field problem with a modified potential
|
|
incorporating the effects of electron correlation, whereas in
|
|
correlated methods the real Hamiltonian is used and the
|
|
electron-electron interactions are considered.
|
|
Nevertheless, as the orbitals are one-electron functions,
|
|
the procedure of orbital optimization in the presence of the
|
|
Jastrow factor can be interpreted as a self-consistent field procedure
|
|
with an effective Hamiltonian,\cite{Filippi_2000} similarly to DFT.
|
|
So KS-DFT can be viewed as a very cheap way of introducing the effect of
|
|
correlation in the orbital parameters determining the nodal surface
|
|
of a single Slater determinant.
|
|
Nevertheless, even when using the exact exchange correlation potential at the
|
|
CBS limit, a fixed-node error necessarily remains because the
|
|
single-determinant ansätz does not have enough flexibility to describe the
|
|
nodal surface of the exact correlated wave function of a generic $N$-electron
|
|
system.
|
|
If one wants to have to exact CBS limit, a multi-determinant parameterization
|
|
of the wave functions is required.
|
|
|
|
\subsection{CIPSI}
|
|
|
|
Beyond the single-determinant representation, the best
|
|
multi-determinant wave function one can obtain is the FCI. FCI is
|
|
a \emph{post-Hartree-Fock} method, and there exists several systematic
|
|
improvements between the Hartree-Fock and FCI wave functions:
|
|
increasing the maximum degree of excitation of CI methods (CISD, CISDT,
|
|
CISDTQ, \emph{etc}), or increasing the complete active space
|
|
(CAS) wave functions until all the orbitals are in the active space.
|
|
Selected CI methods take a shorter path between the Hartree-Fock
|
|
determinant and the FCI wave function by increasing iteratively the
|
|
number of determinants on which the wave function is expanded,
|
|
selecting the determinants which are expected to contribute the most
|
|
to the FCI eigenvector. At every iteration, the lowest eigenpair is
|
|
extracted from the CI matrix expressed in the determinant subspace,
|
|
and the FCI energy can be estimated by computing a second-order
|
|
perturbative correction (PT2) to the variational energy, $\EPT$.
|
|
The magnitude of $\EPT$ is a
|
|
measure of the distance to the exact eigenvalue, and is an adjustable
|
|
parameter controlling the quality of the wave function.
|
|
Within the \emph{Configuration interaction using a perturbative
|
|
selection made iteratively} (CIPSI)\cite{Huron_1973} method, the PT2
|
|
correction is computed along with the determinant selection. So the
|
|
magnitude of $\EPT$ can be made the only parameter of the algorithm,
|
|
and we choose this parameter as the convergence criterion of the CIPSI
|
|
algorithm.
|
|
|
|
Considering that the perturbatively corrected energy is a reliable
|
|
estimate of the FCI energy, using a fixed value of the PT2 correction
|
|
as a stopping criterion enforces a constant distance of all the
|
|
calculations to the FCI energy. In this work, we target the chemical
|
|
accuracy so all the CIPSI selections were made such that $|\EPT| <
|
|
1$~mE$_h$.
|
|
|
|
|
|
|
|
\subsection{Range-separated DFT}
|
|
\label{sec:rsdft}
|
|
|
|
Following the seminal work of Savin,\cite{Savin_1996,Toulouse_2004}
|
|
the Coulomb electron-electron interaction is split into a short-range
|
|
(sr) and a long range (lr) interaction as
|
|
\begin{equation}
|
|
\frac{1}{r_{ij}} = w_{\text{ee}}^{\text{lr}, \mu}(r_{ij}) + \qty(
|
|
\frac{1}{r_{ij}} - w_{\text{ee}}^{\text{lr}, \mu}(r_{ij}) )
|
|
\end{equation}
|
|
where
|
|
\begin{equation}
|
|
w_{\text{ee}}^{\text{lr},\mu}(r_{ij}) = \frac{\erf \qty( \mu\, r_{ij})}{r_{ij}}
|
|
\end{equation}
|
|
The main idea is to treat the short-range electron-electron
|
|
interaction with DFT, and the long range with wave function theory.
|
|
The parameter $\mu$ controls the range of the separation, and allows
|
|
to go continuously from the Kohn-Sham Hamiltonian ($\mu=0$) to
|
|
the FCI Hamiltinoan ($\mu = \infty$).
|
|
|
|
To rigorously connect wave function theory and DFT, the universal
|
|
Levy-Lieb density functional\cite{Lev-PNAS-79,Lie-IJQC-83} is
|
|
decomposed as
|
|
\begin{equation}
|
|
\mathcal{F}[n] = \mathcal{F}^{\mathrm{lr},\mu}[n] + \bar{E}_{\mathrm{Hxc}}^{\mathrm{sr,}\mu}[n],
|
|
\label{Fdecomp}
|
|
\end{equation}
|
|
where $n$ is a one-particle density,
|
|
$\mathcal{F}^{\mathrm{lr},\mu}$ is a long-range universal density
|
|
functional and $\bar{E}_{\mathrm{Hxc}}^{\mathrm{sr,}\mu}$ is the
|
|
complementary short-range Hartree-exchange-correlation (Hxc) density
|
|
functional\cite{Savin_1996,Toulouse_2004}.
|
|
One obtains the following expression for the ground-state
|
|
electronic energy
|
|
\begin{equation}
|
|
\label{min_rsdft} E_0= \min_{\Psi} \left\{
|
|
\left
|
|
\langle\Psi|\hat{T}+\hat{W}_\mathrm{{ee}}^{\mathrm{lr},\mu}+\hat{V}_{\mathrm{ne}}|\Psi\right
|
|
\rangle
|
|
+ \bar{E}^{\mathrm{sr},\mu}_{\mathrm{Hxc}}[n_\Psi]\right\}
|
|
\end{equation}
|
|
with $\hat{T}$ the kinetic energy operator,
|
|
$\hat{W}_\mathrm{ee}^{\mathrm{lr}}$ the long-range
|
|
electron-electron interaction,
|
|
$n_\Psi$ the one-particle density associated with $\Psi$,
|
|
and $\hat{V}_{\mathrm{ne}}$ the electron-nucleus potential.
|
|
The minimizing multi-determinant wave function $\Psi^\mu$
|
|
can be determined by the self-consistent eigenvalue equation
|
|
\begin{equation}
|
|
\label{rs-dft-eigen-equation}
|
|
\hat{H}^\mu[n_{\Psi^{\mu}}] \ket{\Psi^{\mu}}= \mathcal{E}^{\mu} \ket{\Psi^{\mu}},
|
|
\end{equation}
|
|
with the long-range interacting Hamiltonian
|
|
\begin{equation}
|
|
\label{H_mu}
|
|
\hat{H}^\mu[n_{\Psi^{\mu}}] = \hat{T}+\hat{W}_{\mathrm{ee}}^{\mathrm{lr},\mu}+\hat{V}_{\mathrm{ne}}+ \hat{\bar{V}}_{\mathrm{Hxc}}^{\mathrm{sr},\mu}[n_{\Psi^{\mu}}],
|
|
\end{equation}
|
|
where
|
|
$\hat{\bar{V}}_{\mathrm{Hxc}}^{\mathrm{sr},\mu}$
|
|
is the complementary short-range Hartree-exchange-correlation
|
|
potential operator.
|
|
Once $\Psi^{\mu}$ has been calculated, the electronic ground-state
|
|
energy is obtained by
|
|
\begin{equation}
|
|
\label{E-rsdft}
|
|
E_0= \mel{\Psi^{\mu}}{\hat{T}+\hat{W}_\mathrm{{ee}}^{\mathrm{lr},\mu}+\hat{V}_{\mathrm{ne}}}{\Psi^{\mu}}+\bar{E}^{\mathrm{sr},\mu}_{\mathrm{Hxc}}[n_{\Psi^\mu}].
|
|
\end{equation}
|
|
|
|
Note that, for $\mu=0$, the long-range interaction vanishes,
|
|
$w_{\mathrm{ee}}^{\mathrm{lr},\mu=0}(r_{12}) = 0$, and thus
|
|
range-separated DFT (RS-DFT) reduces to standard KS-DFT and $\Psi^\mu$
|
|
is the KS determinant. For $\mu\to\infty$, the long-range
|
|
interaction becomes the standard Coulomb interaction,
|
|
$w_{\mathrm{ee}}^{\mathrm{lr},\mu\to\infty}(r_{12}) = 1/r_{12}$, and
|
|
thus RS-DFT reduces to standard wave-function theory and $\Psi^\mu$ is
|
|
the FCI wave function.
|
|
|
|
\begin{figure*}
|
|
\centering
|
|
\includegraphics[width=0.7\linewidth]{algorithm.pdf}
|
|
\caption{Algorithm showing the generation of the RS-DFT wave
|
|
function.}
|
|
\label{fig:algo}
|
|
\end{figure*}
|
|
|
|
Hence we have a continuous path connecting the KS determinant to the
|
|
FCI wave function, and as the KS nodes are of higher quality than the
|
|
HF nodes, we expect that using wave functions built along this path
|
|
will always provide reduced fixed-node errors compared to the path
|
|
connecting HF to FCI using an increasing number of selected
|
|
determinants.
|
|
|
|
We can follow this path by performing FCI calculations using the
|
|
RS-DFT Hamiltonian with different values of $\mu$. In this work, we
|
|
have used the CIPSI algorithm to perform approximate FCI calculations
|
|
with the RS-DFT Hamiltonians,\cite{GinPraFerAssSavTou-JCP-18}
|
|
$\hat{H}^\mu$ as shown in figure~\ref{fig:algo}. In the outer loop
|
|
(red), a CIPSI selection is performed with a RS-Hamiltonian
|
|
parameterized using the current density.
|
|
An inner loop (blue) is introduced to accelerate the
|
|
convergence of the self-consistent calculation, in which the set of
|
|
determinants is kept fixed, and only the diagonalization of the
|
|
RS-Hamiltonian is performed iteratively with the updated density.
|
|
The convergence of the algorithm was further improved
|
|
by introducing a direct inversion in the iterative subspace (DIIS)
|
|
step to extrapolate the density both in the outer and inner loops.
|
|
Note that any range-separated post-Hartree-Fock method can be
|
|
implemented using this scheme by just replacing the CIPSI step by the
|
|
post-HF method of interest.
|
|
|
|
\section{Computational details}
|
|
\label{sec:comp-details}
|
|
|
|
|
|
All the calculations were made using BFD
|
|
pseudopotentials\cite{Burkatzki_2008} with the associated double-,
|
|
triple-, and quadruple-$\zeta$ basis sets (BFD-VXZ).
|
|
CCSD(T) and KS-DFT calculations were made with
|
|
\emph{Gaussian09},\cite{g16} using an unrestricted Hartree-Fock
|
|
determinant as a reference for open-shell systems.
|
|
|
|
All the CIPSI calculations were made with \emph{Quantum
|
|
Package}.\cite{Garniron_2019,qp2_2020} We used the short-range version
|
|
of the Perdew-Burke-Ernzerhof (PBE)~\cite{PerBurErn-PRL-96} exchange
|
|
and correlation functionals of
|
|
Ref.~\onlinecite{GolWerStoLeiGorSav-CP-06} (see also
|
|
Refs.~\onlinecite{TouColSav-JCP-05,GolWerSto-PCCP-05}).
|
|
The convergence criterion for stopping the CIPSI calculations
|
|
was $\EPT < 1$~m\hartree{} $\vee \Ndet > 10^7$.
|
|
All the wave functions are eigenfunctions of the $S^2$ operator, as
|
|
described in ref~\onlinecite{Applencourt_2018}.
|
|
|
|
Quantum Monte Carlo calculations were made with QMC=Chem,\cite{scemama_2013}
|
|
in the determinant localization approximation (DLA),\cite{Zen_2019}
|
|
where only the determinantal component of the trial wave
|
|
function is present in the expression of the wave function on which
|
|
the pseudopotential is localized. Hence, in the DLA the fixed-node
|
|
energy is independent of the Jastrow factor, as in all-electron
|
|
calculations. Simple Jastrow factors were used to reduce the
|
|
fluctuations of the local energy.
|
|
|
|
|
|
|
|
|
|
\section{Influence of the range-separation parameter on the fixed-node
|
|
error}
|
|
\label{sec:mu-dmc}
|
|
\begin{table}
|
|
\caption{Fixed-node energies (in hartree) and number of determinants in \ce{H2O} and \ce{F2} with various trial wave functions.}
|
|
\label{tab:h2o-dmc}
|
|
\centering
|
|
\begin{ruledtabular}
|
|
\begin{tabular}{ccrlrl}
|
|
& & \multicolumn{2}{c}{BFD-VDZ} & \multicolumn{2}{c}{BFD-VTZ} \\
|
|
\cline{3-4} \cline{5-6}
|
|
System & $\mu$ & $\Ndet$ & $\EDMC$ & $\Ndet$ & $\EDMC$ \\
|
|
\hline
|
|
\ce{H2O}
|
|
& $0.00$ & $11$ & $-17.253\,59(6)$ & $23$ & $-17.256\,74(7)$ \\
|
|
& $0.20$ & $23$ & $-17.253\,73(7)$ & $23$ & $-17.256\,73(8)$ \\
|
|
& $0.30$ & $53$ & $-17.253\,4(2)$ & $219$ & $-17.253\,7(5)$ \\
|
|
& $0.50$ & $1\,442$ & $-17.253\,9(2)$ & $16\,99$ & $-17.257\,7(2)$ \\
|
|
& $0.75$ & $3\,213$ & $-17.255\,1(2)$ & $13\,362$ & $-17.258\,4(3)$ \\
|
|
& $1.00$ & $6\,743$ & $-17.256\,6(2)$ & $256\,73$ & $-17.261\,0(2)$ \\
|
|
& $1.75$ & $54\,540$ & $-17.259\,5(3)$ & $207\,475$ & $-17.263\,5(2)$ \\
|
|
& $2.50$ & $51\,691$ & $-17.259\,4(3)$ & $858\,123$ & $-17.264\,3(3)$ \\
|
|
& $3.80$ & $103\,059$ & $-17.258\,7(3)$ & $1\,621\,513$ & $-17.263\,7(3)$ \\
|
|
& $5.70$ & $102\,599$ & $-17.257\,7(3)$ & $1\,629\,655$ & $-17.263\,2(3)$ \\
|
|
& $8.50$ & $101\,803$ & $-17.257\,3(3)$ & $1\,643\,301$ & $-17.263\,3(4)$ \\
|
|
& $\infty$ & $200\,521$ & $-17.256\,8(6)$ & $1\,631\,982$ & $-17.263\,9(3)$ \\
|
|
\\
|
|
\ce{F2}
|
|
& $0.00$ & $23$ & $-48.419\,5(4)$ \\
|
|
& $0.25$ & $8$ & $-48.421\,9(4)$ \\
|
|
& $0.50$ & $1743$ & $-48.424\,8(8)$ \\
|
|
& $1.00$ & $11952$ & $-48.432\,4(3)$ \\
|
|
& $2.00$ & $829438$ & $-48.441\,0(7)$ \\
|
|
& $5.00$ & $5326459$ & $-48.445(2)$ \\
|
|
& $\infty$ & $8302442$ & $-48.437(3)$ \\
|
|
\end{tabular}
|
|
\end{ruledtabular}
|
|
\end{table}
|
|
|
|
\begin{figure}
|
|
\centering
|
|
\includegraphics[width=\columnwidth]{h2o-dmc.pdf}
|
|
\caption{Fixed-node energies of the water molecule for different
|
|
values of $\mu$, using the sr-LDA or sr-PBE short-range density
|
|
functionals to build the trial wave function.}
|
|
\label{fig:h2o-dmc}
|
|
\end{figure}
|
|
|
|
\begin{figure}
|
|
\centering
|
|
\includegraphics[width=\columnwidth]{f2-dmc.pdf}
|
|
\caption{Fixed-node energies of difluorine for different
|
|
values of $\mu$.}
|
|
\label{fig:f2-dmc}
|
|
\end{figure}
|
|
The first question we would like to address is the quality of the
|
|
nodes of the wave functions $\Psi^{\mu}$ obtained with an intermediate
|
|
range separation parameter (\textit{i.e.} $0 < \mu < +\infty$).
|
|
We generated trial wave functions $\Psi^\mu$ with multiple values of
|
|
$\mu$, and computed the associated fixed node energy keeping all the
|
|
parameters having an impact on the nodal surface fixed.
|
|
We considered two weakly correlated molecular systems: the water
|
|
molecule and the fluorine dimer, near their equilibrium
|
|
geometry\cite{Caffarel_2016}.
|
|
|
|
\subsection{Fixed node energy of $\Psi^\mu$}
|
|
\label{sec:fndmc_mu}
|
|
From table~\ref{tab:h2o-dmc} and figures~\ref{fig:h2o-dmc}
|
|
and~\ref{fig:f2-dmc}, one can clearly observe that using FCI trial
|
|
wave functions ($\mu = \infty$) gives FN-DMC energies which are lower
|
|
than the energies obtained with a single Kohn-Sham determinant ($\mu=0$):
|
|
a gain of $3.2 \pm 0.6$~m\hartree{} at the double-zeta level and $7.2 \pm
|
|
0.3$~m\hartree{} at the triple-zeta level are obtained for water, and
|
|
a gain of $18 \pm 3$~m\hartree{} for F$_2$.
|
|
Coming now to the nodes of the trial wave functions $\Psi^{\mu}$ with intermediate values of $\mu$,
|
|
the figures~\ref{fig:h2o-dmc} show that a smooth behaviour is obtained:
|
|
starting from $\mu=0$ (\textit{i.e.} the KS determinant),
|
|
the FN-DMC error is reduced continuously until it reaches a minimum for an optimal value of $\mu$,
|
|
and then the FN-DMC error raises until it reaches the $\mu=\infty$ limit (\textit{i.e.} the FCI wave function).
|
|
For instance, with respect to the FN-DMC energy of the FCI trial wave function in the double zeta basis set,
|
|
with the optimal value of $\mu$, one can obtain a lowering of the FN-DMC energy of $2.6 \pm 0.7$~m\hartree{}
|
|
and $8 \pm 3$~m\hartree{} for the water and difluorine molecules, respectively.
|
|
The optimal value of $\mu$ is $\mu=1.75$~bohr$^{-1}$ and $\mu=5$~bohr$^{-1}$ for the water and fluorine dimer, respectively.
|
|
When the basis set is increased, the gain in FN-DMC energy with respect to the FCI trial wave function is reduced, and the optimal value of $\mu$ is shifted towards large $\mu$.
|
|
Eventually, the nodes of the wave functions $\Psi^\mu$ obtained with short-range
|
|
LDA exchange-correlation functional give very similar FN-DMC energy with respect
|
|
to those obtained with the short-range PBE functional, even if the RS-DFT energies obtained
|
|
with these two functionals differ by several tens of m\hartree{}.
|
|
|
|
\subsection{Link between RS-DFT and jastrow factors }
|
|
\label{sec:rsdft-j}
|
|
The data obtained in \ref{sec:fndmc_mu} show that RS-DFT can provide CI coefficients
|
|
giving trial wave functions with better nodes than FCI wave functions.
|
|
Such behaviour can be compared to the common practice of
|
|
re-optimizing the Slater part of a trial wave function in the presence of the Jastrow factor.
|
|
In the present paragraph, we would like therefore to elaborate some more on the link between RS-DFT
|
|
and wave function optimization within the presence of a jastrow factor.
|
|
|
|
Let us assume a fixed jastrow factor $J({\bf{r}_1}, \hdots , {\bf{r}_N})$,
|
|
and a corresponding Slater-jatrow wave function $\Phi = e^J \Psi$,
|
|
where $\Psi = \sum_I c_I D_I$ is a general linear combination of Slater determinants $D_I$.
|
|
The only remaining variational parameters in $\Phi$ are therefore the Slater part $\Psi$.
|
|
Let us call $\Psi^J$ the linear combination of Slater determinant minimizing the variational energy
|
|
\begin{equation}
|
|
\Psi^J = \text{argmin}_{\Psi}\frac{ \langle \Psi | e^{J} H e^{J} |\Psi \rangle}{\langle \Psi | e^{2J} |\Psi \rangle}.
|
|
\end{equation}
|
|
Such a wave function $\Psi^J$ satisfies the generalized hermitian eigenvalue equation
|
|
\begin{equation}
|
|
e^{J} H e^{J} \Psi^J = E e^{2J} \Psi^J,
|
|
\end{equation}
|
|
but also the non-hermitian transcorrelated eigenvalue problem\cite{many_things}
|
|
\begin{equation}
|
|
\label{eq:transcor}
|
|
e^{-J} H e^{J} \Psi^J = E \Psi^J,
|
|
\end{equation}
|
|
which is much easier to handle despite its non-hermicity.
|
|
Of course, the FN-DMC energy of $\Phi$ depends therefore only on the nodes of $\Psi^J$.
|
|
In a finite basis set and with a quite accurate jastrow factor, it is known that the nodes
|
|
of $\Psi^J$ are better than that of the FCI wave function, and therefore, we would like to compare $\Psi^J$ and $\Psi^\mu$.
|
|
|
|
To do so, we have made the following numerical experiment.
|
|
First, we extract the 200 determinants with the largest weights in the FCI wave
|
|
function out of a large CIPSI calculation. Within this set of determinants,
|
|
we solve the self-consistent equations of RS-DFT (see Eq.~\eqref{rs-dft-eigen-equation})
|
|
with different values of $\mu$. This gives the CI expansions $\Psi^\mu$.
|
|
Then, within the same set of determinants we optimize the CI coefficients in the presence of
|
|
a simple one- and two-body Jastrow factor. This gives the CI expansion $\Psi^J$.
|
|
Then, we can easily compare $\Psi^\mu$ and $\Psi^J$ as they are developed
|
|
on the same Slater determinant basis.
|
|
In figure~\ref{fig:overlap}, we plot the overlaps
|
|
$\braket{\Psi^J}{\Psi^\mu}$ obtained for water and the fluorine dimer,
|
|
and in figure~\ref{dmc_small} the FN-DMC energy of the wave functions
|
|
$\Psi^\mu$ together with that of $\Psi^J$.
|
|
|
|
\begin{figure}
|
|
\centering
|
|
\includegraphics[width=\columnwidth]{overlap.pdf}
|
|
\caption{Overlap of the RS-DFT CI expansion with the
|
|
CI expansion optimized in the presence of a Jastrow factor.}
|
|
\label{fig:overlap}
|
|
\end{figure}
|
|
|
|
\begin{figure}
|
|
\centering
|
|
\includegraphics[width=\columnwidth]{x.png}
|
|
\caption{H$_2$O, double-zeta basis set. FN-DMC energy of $\Psi^\mu$ built with the 200 largest determinants of a large CIPSI expansion, together with the FN-DMC energy of $\Psi^J$ (see \ref{sec:rsdft-j}).}
|
|
\label{dmc_small}
|
|
\end{figure}
|
|
|
|
|
|
|
|
In the case of H$_2$O, there is a clear maximum of overlap at
|
|
$\mu=1$~bohr$^{-1}$, which coincide with the minimum of the FN-DMC energy of $\Psi^\mu$.
|
|
???? hypothetic: Also, it is interesting to notice that the FN-DMC energy of $\Psi^J$ is very close to that of $\Psi^\mu$ with $\mu=1$~bohr$^{-1}$.
|
|
This confirms that introducing short-range correlation with DFT has
|
|
an impact on the CI coefficients similar to the Jastrow factor.
|
|
In the case of F$_2$, the Jastrow factor has
|
|
very little effect on the CI coefficients, as the overlap
|
|
$\braket{\Psi^J}{\Psi^{\mu=\infty}}$ is very close to
|
|
$1$.
|
|
Nevertheless, a slight maximum is obtained for
|
|
$\mu=5$~bohr$^{-1}$.
|
|
In order to refine the comparison between $\Psi^\mu$ and $\Psi^J$,
|
|
we report, in the case of the water molecule in the double-zeta basis set,
|
|
several quantities related to the one- and two-body density of $\Psi^J$ and $\Psi^\mu$ with different values of $\mu$.
|
|
First, we report in table~\ref{table_on_top} the integrated on-top pair density $\langle n_2({\bf r},{\bf r}) \rangle$
|
|
\begin{equation}
|
|
\langle n_2({\bf r},{\bf r}) \rangle = \int \text{d}{\bf r} \,\,n_2({\bf r},{\bf r})
|
|
\end{equation}
|
|
where $n_2({\bf r}_1,{\bf r}_2)$ is the two-body density (normalized to $N(N-1)$ where $N$ is the number of electrons)
|
|
obtained for both $\Psi^\mu$ and $\Psi^J$.
|
|
Then, in order to have a pictorial representation of both the on-top pair density and the density, we report in figures~\ref{fig:n1} and ~\ref{fig:n2}
|
|
the plots of the total density $n({\bf r})$ and on-top pair density $n_2({\bf r},{\bf r})$ along the OH axis of the water molecule.
|
|
From these data, one can clearly observe several trends.
|
|
First, from table~\ref{table_on_top}, we can observe that the overall on-top pair density decreases
|
|
when one increases $\mu$, which is expected as the two-electron interaction increases in $H^\mu[n]$.
|
|
Second, the relative variation of the on-top pair density with $\mu$ are much more important than that of the one-body density, the latter being essentially unchanged between $\mu=0$ and $\mu=\infty$ while the former can vary by about 10$\%$ in some regions.
|
|
??? Hypothetic: the value of the on-top pair density in $\Psi^\mu$ are closer for certain values of $\mu$ to that of $\Psi^J$ than the
|
|
FCI wave function.
|
|
|
|
These data suggest that the wave functions $\Psi^\mu$ and $\Psi^J$ are similar,
|
|
and therefore that the operators that produced these wave functions (\textit{i.e.} $H^\mu[n]$ and $e^{-J}He^J$) contain similar physics.
|
|
Considering the form of $\hat{H}^\mu[n]$ (see Eq.~\eqref{H_mu}),
|
|
one can notice that the differences with respect to the usual Hamiltonian come
|
|
from the non-divergent two-body interaction $\hat{W}_{\mathrm{ee}}^{\mathrm{lr},\mu}$
|
|
and the effective one-body potential $\hat{\bar{V}}_{\mathrm{Hxc}}^{\mathrm{sr},\mu}[n]$ which is the functional derivative of the Hartree-exchange-correlation functional.
|
|
The role of these two terms are therefore very different: with respect
|
|
to the exact ground state wave function $\Psi$, the non divergent two body interaction
|
|
increases the probability to find electrons at short distances in $\Psi^\mu$,
|
|
while the effective one-body potential $\hat{\bar{V}}_{\mathrm{Hxc}}^{\mathrm{sr},\mu}[n_{\Psi^{\mu}}]$,
|
|
provided that it is exact, maintains the exact one-body density.
|
|
This is clearly what has been observed from the plots in figures ~\ref{fig:n1} and~\ref{fig:n2} in the case of the water molecule.
|
|
Regarding now the transcorrelated Hamiltonian $e^{-J}He^J$, as pointed out by Ten-No\cite{Ten-no2000Nov},
|
|
the effective two-body interaction induced by the presence of a jastrow factor
|
|
can be non-divergent when a proper Jastrow factor is chosen.
|
|
Therefore, one can understand the similarity between the eigenfunctions of $H^\mu$ and the jastrow-Slater optimization:
|
|
they both deal with an effective non-divergent interaction but still produce reasonable one-body density.
|
|
|
|
\begin{table}
|
|
\caption{H$_2$O, double-zeta basis set. Integrated on-top pair density $\langle n_2({\bf r},{\bf r}) \rangle$
|
|
for $\Psi^J$ and $\Psi^\mu$ with different values of $\mu$. }
|
|
\label{table_on_top}
|
|
\begin{ruledtabular}
|
|
\begin{tabular}{cc}
|
|
$\mu$ & $\langle n_2({\bf r},{\bf r}) \rangle$ \\
|
|
\hline
|
|
0.00 & 1.443 \\
|
|
0.25 & 1.438 \\
|
|
0.50 & 1.423 \\
|
|
1.00 & 1.378 \\
|
|
2.00 & 1.325 \\
|
|
5.00 & 1.288 \\
|
|
$\infty$ & 1.277 \\
|
|
\hline
|
|
$\Psi^J$ & \\
|
|
\end{tabular}
|
|
\end{ruledtabular}
|
|
\end{table}
|
|
|
|
\begin{figure}
|
|
\centering
|
|
\includegraphics[width=\columnwidth]{on-top-mu.pdf}
|
|
\caption{H$_2$O, double-zeta basis set. Integrated on-top pair density $\langle n_2({\bf r},{\bf r}) \rangle$ along the OH axis, for $\Psi^J$ and $\Psi^\mu$ with different values of $\mu$. }
|
|
\label{fig:n2}
|
|
\end{figure}
|
|
\begin{figure}
|
|
\centering
|
|
\includegraphics[width=\columnwidth]{density-mu.pdf}
|
|
\caption{H$_2$O, double-zeta basis set. Density $n({\bf r})$ along the OH axis, for $\Psi^J$ and $\Psi^\mu$ with different values of $\mu$. }
|
|
\label{fig:n1}
|
|
\end{figure}
|
|
|
|
|
|
As a conclusion of the first part of this study, we can notice that:
|
|
i) with respect to the nodes of a KS determinant or a FCI wave function,
|
|
one can obtain a multi determinant trial wave function $\Psi^\mu$ with a smaller
|
|
fixed node error by properly choosing an optimal value of $\mu$
|
|
in RS-DFT calculations, ii) the value of the optimal $\mu$ depends
|
|
on the system and the basis set, and the larger the basis set, the larger the optimal value of $\mu$,
|
|
iii) numerical experiments (such as computation of overlap, FN-DMC energies) indicates
|
|
that the RS-DFT scheme essentially plays the role of a simple Jastrow factor,
|
|
\textit{i.e.} mimicking short-range correlation effects.
|
|
The latter statement can be qualitatively understood by noticing that both RS-DFT
|
|
and transcorrelated approach deal with an effective non-divergent electron-electron interaction, while keeping the density constant.
|
|
|
|
|
|
|
|
\section{Energy differences in FN-DMC: atomization energies}
|
|
\label{sec:atomization}
|
|
|
|
Atomization energies are challenging for post-Hartree-Fock methods
|
|
because their calculation requires a perfect balance in the
|
|
description of atoms and molecules. Basis sets used in molecular
|
|
calculations are atom-centered, so they are always better adapted to
|
|
atoms than molecules and atomization energies usually tend to be
|
|
underestimated with variational methods.
|
|
In the context of FN-DMC calculations, the nodal surface is imposed by
|
|
the trial wavefunction which is expanded on an atom-centered basis
|
|
set, so we expect the fixed-node error to be also tightly related to
|
|
the basis set incompleteness error.
|
|
Increasing the size of the basis set improves the description of
|
|
the density and of electron correlation, but also reduces the
|
|
imbalance in the quality of the description of the atoms and the
|
|
molecule, leading to more accurate atomization energies.
|
|
|
|
\subsection{Size consistency}
|
|
|
|
An extremely important feature required to get accurate
|
|
atomization energies is size-consistency (or strict separability),
|
|
since the numbers of correlated electron pairs in the isolated atoms
|
|
are different from those of the molecules.
|
|
The energy computed within density functional theory is size-consistent, and
|
|
as it is a mean-field method the convergence to the complete basis set
|
|
(CBS) limit is relatively fast. Hence, DFT methods are very well adapted to
|
|
the calculation of atomization energies, especially with small basis
|
|
sets. But going to the CBS limit will converge to biased atomization
|
|
energies because of the use of approximate density functionals.
|
|
|
|
On the other hand, FCI is also size-consistent, but the convergence of
|
|
the FCI energies to the CBS limit is much slower because of the
|
|
description of short-range electron correlation using atom-centered
|
|
functions. But ultimately the exact energy will be reached.
|
|
|
|
In the context of selected CI calculations, when the variational energy is
|
|
extrapolated to the FCI energy\cite{Holmes_2017} there is no
|
|
size-consistency error. But when the truncated sCI wave function is used
|
|
as a reference for post-Hartree-Fock methods such as sCI+PT2
|
|
or for QMC calculations, there is a residual size-consistency error
|
|
originating from the truncation of the wave function.
|
|
|
|
QMC energies can be made size-consistent by extrapolating the
|
|
FN-DMC energy to estimate the energy obtained with the FCI as a trial
|
|
wave function.\cite{Scemama_2018,Scemama_2018b} Alternatively, the
|
|
size-consistency error can be reduced by choosing the number of
|
|
selected determinants such that the sum of the PT2 corrections on the
|
|
fragments is equal to the PT2 correction of the molecule, enforcing that
|
|
the variational potential energy surface (PES) is
|
|
parallel to the perturbatively corrected PES, which is a relatively
|
|
accurate estimate of the FCI PES.\cite{Giner_2015}
|
|
|
|
Another source of size-consistency error in QMC calculations originates
|
|
from the Jastrow factor. Usually, the Jastrow factor contains
|
|
one-electron, two-electron and one-nucleus-two-electron terms.
|
|
The problematic part is the two-electron term, whose simplest form can
|
|
be expressed as
|
|
\begin{equation}
|
|
J_\text{ee} = \sum_i \sum_{j<i} \frac{a\, r_{ij}}{1 + b\, r_{ij}}.
|
|
\end{equation}
|
|
The parameter
|
|
$a$ is determined by cusp conditions, and $b$ is obtained by energy
|
|
or variance minimization.\cite{Coldwell_1977,Umrigar_2005}
|
|
One can easily see that this parameterization of the two-body
|
|
interation is not size-consistent: the dissociation of a
|
|
diatomic molecule $AB$ with a parameter $b_{AB}$
|
|
will lead to two different two-body Jastrow factors, each
|
|
with its own optimal value $b_A$ and $b_B$. To remove the
|
|
size-consistency error on a PES using this ansätz for $J_\text{ee}$,
|
|
one needs to impose that the parameters of $J_\text{ee}$ are fixed:
|
|
$b_A = b_B = b_{AB}$.
|
|
|
|
When pseudopotentials are used in a QMC calculation, it is common
|
|
practice to localize the non-local part of the pseudopotential on the
|
|
complete wave function (determinantal component and Jastrow).
|
|
If the wave function is not size-consistent,
|
|
so will be the locality approximation. Within, the determinant
|
|
localization approximation,\cite{Zen_2019} the Jastrow factor is
|
|
removed from the wave function on which the pseudopotential is localized.
|
|
The great advantage of this approximation is that the FN-DMC energy
|
|
only depends on the parameters of the determinantal component. Using a
|
|
non-size-consistent Jastrow factor, or a non-optimal Jastrow factor will
|
|
not introduce an additional error in FN-DMC calculations, although it
|
|
will reduce the statistical errors by reducing the variance of the
|
|
local energy. Moreover, the integrals involved in the pseudo-potential
|
|
are computed analytically and the computational cost of the
|
|
pseudo-potential is dramatically reduced (for more detail, see
|
|
Ref.~\onlinecite{Scemama_2015}).
|
|
|
|
%\begin{squeezetable}
|
|
\begin{table}
|
|
\caption{FN-DMC energies (in hartree) using the VDZ-BFD basis set
|
|
and pseudo-potential of the fluorine atom and the dissociated fluorine
|
|
dimer, and size-consistency error. }
|
|
\label{tab:size-cons}
|
|
\begin{ruledtabular}
|
|
\begin{tabular}{cccc}
|
|
$\mu$ & \ce{F} & Dissociated \ce{F2} & Size-consistency error \\
|
|
\hline
|
|
0.00 & $-24.188\,7(3)$ & $-48.377\,7(3)$ & $-0.000\,3(4)$ \\
|
|
0.25 & $-24.188\,7(3)$ & $-48.377\,2(4)$ & $+0.000\,2(5)$ \\
|
|
0.50 & $-24.188\,8(1)$ & $-48.376\,9(4)$ & $+0.000\,7(4)$ \\
|
|
1.00 & $-24.189\,7(1)$ & $-48.380\,2(4)$ & $-0.000\,8(4)$ \\
|
|
2.00 & $-24.194\,1(3)$ & $-48.388\,4(4)$ & $-0.000\,2(5)$ \\
|
|
5.00 & $-24.194\,7(4)$ & $-48.388\,5(7)$ & $+0.000\,9(8)$ \\
|
|
$\infty$ & $-24.193\,5(2)$ & $-48.386\,9(4)$ & $+0.000\,1(5)$ \\
|
|
\end{tabular}
|
|
\end{ruledtabular}
|
|
\end{table}
|
|
|
|
In this section, we make a numerical verification that the produced
|
|
wave functions are size-consistent for a given range-separation
|
|
parameter.
|
|
We have computed the energy of the dissociated fluorine dimer, where
|
|
the two atoms are at a distance of 50~\AA. We expect that the energy
|
|
of this system is equal to twice the energy of the fluorine atom.
|
|
The data in table~\ref{tab:size-cons} shows that it is indeed the
|
|
case, so we can conclude that the proposed scheme provides
|
|
size-consistent FN-DMC energies for all values of $\mu$ (within
|
|
$2\times$ statistical error bars).
|
|
|
|
|
|
\subsection{Spin invariance}
|
|
|
|
Closed-shell molecules often dissociate into open-shell
|
|
fragments. To get reliable atomization energies, it is important to
|
|
have a theory which is of comparable quality for open-shell and
|
|
closed-shell systems. A good test is to check that all the components
|
|
of a spin multiplet are degenerate.
|
|
FCI wave functions have this property and give degenrate energies with
|
|
respect to the spin quantum number $m_s$, but the multiplication by a
|
|
Jastrow factor introduces spin contamination if the parameters
|
|
for the same-spin electron pairs are different from those
|
|
for the opposite-spin pairs.\cite{Tenno_2004}
|
|
Again, when pseudo-potentials are used this tiny error is transferred
|
|
in the FN-DMC energy unless the determinant localization approximation
|
|
is used.
|
|
|
|
Within DFT, the common density functionals make a difference for
|
|
same-spin and opposite-spin interactions. As DFT is a
|
|
single-determinant theory, the density functionals are designed to be
|
|
used with the highest value of $m_s$, and therefore different values
|
|
of $m_s$ lead to different energies.
|
|
So in the context of RS-DFT, the determinantal expansions will be
|
|
impacted by this spurious effect, as opposed to FCI.
|
|
|
|
\begin{table}
|
|
\caption{FN-DMC energies (in hartree) of the triplet carbon atom (BFD-VDZ) with
|
|
different values of $m_s$.}
|
|
\label{tab:spin}
|
|
\begin{ruledtabular}
|
|
\begin{tabular}{cccc}
|
|
$\mu$ & $m_s=1$ & $m_s=0$ & Spin-invariance error \\
|
|
\hline
|
|
0.00 & $-5.416\,8(1)$ & $-5.414\,9(1)$ & $+0.001\,9(2)$ \\
|
|
0.25 & $-5.417\,2(1)$ & $-5.416\,5(1)$ & $+0.000\,7(1)$ \\
|
|
0.50 & $-5.422\,3(1)$ & $-5.421\,4(1)$ & $+0.000\,9(2)$ \\
|
|
1.00 & $-5.429\,7(1)$ & $-5.429\,2(1)$ & $+0.000\,5(2)$ \\
|
|
2.00 & $-5.432\,1(1)$ & $-5.431\,4(1)$ & $+0.000\,7(2)$ \\
|
|
5.00 & $-5.431\,7(1)$ & $-5.431\,4(1)$ & $+0.000\,3(2)$ \\
|
|
$\infty$ & $-5.431\,6(1)$ & $-5.431\,3(1)$ & $+0.000\,3(2)$ \\
|
|
\end{tabular}
|
|
\end{ruledtabular}
|
|
\end{table}
|
|
|
|
In this section, we investigate the impact of the spin contamination
|
|
due to the short-range density functional on the FN-DMC energy. We have
|
|
computed the energies of the carbon atom in its triplet state
|
|
with BFD pseudo-potentials and the corresponding double-zeta basis
|
|
set. The calculation was done with $m_s=1$ (3 $\uparrow$ electrons
|
|
and 1 $\downarrow$ electrons) and with $m_s=0$ (2 $\uparrow$ and 2
|
|
$\downarrow$ electrons).
|
|
|
|
The results are presented in table~\ref{tab:spin}.
|
|
Although using $m_s=0$ the energy is higher than with $m_s=1$, the
|
|
bias is relatively small, more than one order of magnitude smaller
|
|
than the energy gained by reducing the fixed-node error going from the single
|
|
determinant to the FCI trial wave function. The highest bias, close to
|
|
2~m\hartree, is obtained for $\mu=0$, but the bias decreases quickly
|
|
below 1~m\hartree{} when $\mu$ increases. As expected, with $\mu=\infty$
|
|
there is no bias (within the error bars), and the bias is not
|
|
noticeable with $\mu=5$~bohr$^{-1}$.
|
|
|
|
|
|
|
|
|
|
\subsection{Benchmark}
|
|
|
|
\begin{squeezetable}
|
|
\begin{table*}
|
|
\caption{Mean absolute errors (MAE), mean signed errors (MSE) and
|
|
standard deviations (RMSD) obtained with various methods and
|
|
basis sets.}
|
|
\label{tab:mad}
|
|
\begin{ruledtabular}
|
|
\begin{tabular}{ll ddd ddd ddd}
|
|
& & \mc{3}{c}{VDZ-BFD} & \mc{3}{c}{VTZ-BFD} & \mc{3}{c}{VQZ-BFD} \\
|
|
\cline{3-5} \cline{6-8} \cline{9-11}
|
|
Method & $\mu$ & \tabc{MAE} & \tabc{MSE} & \tabc{RMSD} & \tabc{MAE} & \tabc{MSE} & \tabc{RMSD} & \tabc{MAE} & \tabc{MSE} & \tabc{RMSD} \\
|
|
\hline
|
|
PBE & 0 & 5.02 & -3.70 & 6.04 & 4.57 & 1.00 & 5.32 & 5.31 & 0.79 & 6.27 \\
|
|
BLYP & 0 & 9.53 & -9.21 & 7.91 & 5.58 & -4.44 & 5.80 & 5.86 & -4.47 & 6.43 \\
|
|
PBE0 & 0 & 11.20 & -10.98 & 8.68 & 6.40 & -5.78 & 5.49 & 6.28 & -5.65 & 5.08 \\
|
|
B3LYP & 0 & 11.27 & -10.98 & 9.59 & 7.27 & -5.77 & 6.63 & 6.75 & -5.53 & 6.09 \\
|
|
\\
|
|
CCSD(T) & \(\infty\) & 24.10 & -23.96 & 13.03 & 9.11 & -9.10 & 5.55 & 4.52 & -4.38 & 3.60 \\
|
|
\\
|
|
RS-DFT-CIPSI & 0 & 4.53 & -1.66 & 5.91 & 6.31 & 0.91 & 7.93 & 6.35 & 3.88 & 7.20 \\
|
|
& 1/4 & 5.55 & -4.66 & 5.52 & 4.58 & 1.06 & 5.72 & 5.48 & 1.52 & 6.93 \\
|
|
& 1/2 & 13.42 & -13.27 & 7.36 & 6.77 & -6.71 & 4.56 & 6.35 & -5.89 & 5.18 \\
|
|
& 1 & 17.07 & -16.92 & 9.83 & 9.06 & -9.06 & 5.88 & & & \\
|
|
& 2 & 19.20 & -19.05 & 10.91 & & & & & & \\
|
|
& 5 & 22.93 & -22.79 & 13.24 & & & & & & \\
|
|
& \(\infty\) & 23.63(4) & -23.49(4) & 12.81(4) & 8.43(39) & -8.43(39) & 4.87(7) & 4.51(78) & -4.18(78) & 4.19(20) \\
|
|
\\
|
|
DMC@ & 0 & 4.61(34) & -3.62(34) & 5.30(09) & 3.52(19) & -1.03(19) & 4.39(04) & 3.16(26) & -0.12(26) & 4.12(03) \\
|
|
RS-DFT-CIPSI & 1/4 & 4.04(37) & -3.13(37) & 4.88(10) & 3.39(77) & -0.59(77) & 4.44(34) & 2.90(25) & 0.25(25) & 3.745(5) \\
|
|
& 1/2 & 3.74(35) & -3.53(35) & 4.03(23) & 2.46(18) & -1.72(18) & 3.02(06) & 2.06(35) & -0.44(35) & 2.74(13) \\
|
|
& 1 & 5.42(29) & -5.14(29) & 4.55(03) & 4.38(94) & -4.24(94) & 5.11(31) & & & \\
|
|
& 2 & 5.98(83) & -5.91(83) & 4.79(71) & & & & & & \\
|
|
& 5 & 6.18(84) & -6.13(84) & 4.87(55) & & & & & & \\
|
|
& \(\infty\) & 7.38(1.08) & -7.38(1.08) & 5.67(68) & & & & & & \\
|
|
& Opt. & 5.85(1.75) & -5.63(1.75) & 4.79(1.11) & & & & & & \\
|
|
\end{tabular}
|
|
\end{ruledtabular}
|
|
\end{table*}
|
|
\end{squeezetable}
|
|
|
|
The 55 molecules of the benchmark for the Gaussian-1
|
|
theory\cite{Pople_1989,Curtiss_1990} were chosen to test the
|
|
performance of the RS-DFT-CIPSI trial wave functions in the context of
|
|
energy differences. Calculations were made in the double-, triple-
|
|
and quadruple-zeta basis sets with different values of $\mu$, and using
|
|
natural orbitals of a preliminary CIPSI calculation.
|
|
For comparison, we have computed the energies of all the atoms and
|
|
molecules at the DFT level with different density functionals, and at
|
|
the CCSD(T) level. Table~\ref{tab:mad} gives the corresponding mean
|
|
absolute errors (MAE), mean signed errors (MSE) and standard
|
|
deviations (RMSD). For FCI (RS-DFT-CIPSI, $\mu=\infty$) we have
|
|
given extrapolated values at $\EPT\rightarrow 0$, and the error bars
|
|
correspond to the difference between the energies computed with a
|
|
two-point and with a three-point linear extrapolation.
|
|
|
|
In this benchmark, the great majority of the systems are well
|
|
described by a single determinant. Therefore, the atomization energies
|
|
calculated at the DFT level are relatively accurate, even when
|
|
the basis set is small. The introduction of exact exchange (B3LYP and
|
|
PBE0) make the results more sensitive to the basis set, and reduce the
|
|
accuracy. Thanks to the single-reference character of these systems,
|
|
the CCSD(T) energy is an excellent estimate of the FCI energy, as
|
|
shown by the very good agreement of the MAE, MSE and RMSD of CCSDT(T)
|
|
and FCI energies.
|
|
The imbalance of the quality of description of molecules compared
|
|
to atoms is exhibited by a very negative value of the MSE for
|
|
CCSD(T) and FCI/VDZ-BFD, which is reduced by a factor of two
|
|
when going to the triple-zeta basis, and again by a factor of two when
|
|
going to the quadruple-zeta basis.
|
|
|
|
This large imbalance at the double-zeta level affects the nodal
|
|
surfaces, because although the FN-DMC energies obtained with near-FCI
|
|
trial wave functions are much lower than the single-determinant FN-DMC
|
|
energies, the MAE obtained with FCI (7.38~$\pm$ 1.08~kcal/mol) is
|
|
larger than the single-determinant MAE (4.61~$\pm$ 0.34 kcal/mol).
|
|
Using the FCI trial wave function the MSE is equal to the
|
|
negative MAE which confirms that all the atomization energies are
|
|
underestimated. This confirms that some of the basis-set
|
|
incompleteness error is transferred in the fixed-node error.
|
|
|
|
Within the double-zeta basis set, the calculations could be done for the
|
|
whole range of values of $\mu$, and the optimal value of $\mu$ for the
|
|
trial wave function was estimated for each system by searching for the
|
|
minimum of the spline interpolation curve of the FN-DMC energy as a
|
|
function of $\mu$.
|
|
This corresponds the the line of the table labelled by the \emph{Opt}
|
|
value of $\mu$. Using the optimal value of $\mu$ clearly improves the
|
|
MAE, the MSE an the RMSD compared the the FCI wave function. This
|
|
result is in line with the common knowledge that re-optimizing
|
|
the determinantal component of the trial wave function in the presence
|
|
of electron correlation reduces the errors due to the basis set incompleteness.
|
|
These calculations were done only for the smallest basis set
|
|
because of the expensive computational cost of the QMC calculations
|
|
when the trial wave function is expanded on more than a few million
|
|
determinants.
|
|
At the RS-DFT-CIPSI level, we can remark that with the triple-zeta
|
|
basis set the MAE are larger for $\mu=1$~bohr$^{-1}$ than for the
|
|
FCI. For the largest systems, as shown in figure~\ref{fig:g2-ndet}
|
|
there are many systems which did not reach the threshold
|
|
$\EPT<1$~m\hartree{}, and the number of determinants exceeded
|
|
10~million so the calculation stopped. In this regime, there is a
|
|
small size-consistency error originating from the imbalanced
|
|
truncation of the wave functions, which is not present in the
|
|
extrapolated FCI energies. The same comment applies to
|
|
$\mu=0.5$~bohr$^{-1}$ with the quadruple-zeta basis set.
|
|
|
|
|
|
\begin{figure}
|
|
\centering
|
|
\includegraphics[width=\columnwidth]{g2-dmc.pdf}
|
|
\caption{Errors in the FN-DMC atomization energies with the different
|
|
trial wave functions. Each dot corresponds to an atomization
|
|
energy.
|
|
The boxes contain the data between first and third quartiles, and
|
|
the line in the box represents the median. The outliers are shown
|
|
with a cross.}
|
|
\label{fig:g2-dmc}
|
|
\end{figure}
|
|
|
|
Searching for the optimal value of $\mu$ may be too costly, so we have
|
|
computed the MAD, MSE and RMSD for fixed values of $\mu$. The results
|
|
are illustrated in figure~\ref{fig:g2-dmc}. As seen on the figure and
|
|
in table~\ref{tab:mad}, the best choice for a fixed value of $\mu$ is
|
|
0.5~bohr$^{-1}$ for all three basis sets. It is the value for which
|
|
the MAE (3.74(35), 2.46(18) and 2.06(35) kcal/mol) and RMSD (4.03(23),
|
|
3.02(06) and 2.74(13)~kcal/mol) are minimal. Note that these values
|
|
are even lower than those obtained with the optimal value of
|
|
$\mu$. Although the FN-DMC energies are higher, the numbers show that
|
|
they are more consistent from one system to another, giving improved
|
|
cancellations of errors.
|
|
|
|
\begin{figure}
|
|
\centering
|
|
\includegraphics[width=\columnwidth]{g2-ndet.pdf}
|
|
\caption{Number of determinants in the different trial wave
|
|
functions. Each dot corresponds to an atomization energy.
|
|
The boxes contain the data between first and third quartiles, and
|
|
the line in the box represents the median. The outliers are shown
|
|
with a cross.}
|
|
\label{fig:g2-ndet}
|
|
\end{figure}
|
|
|
|
|
|
The number of determinants in the trial wave functions are shown in
|
|
figure~\ref{fig:g2-ndet}. As expected, the number of determinants
|
|
is smaller when $\mu$ is small and larger when $\mu$ is large.
|
|
It is important to remark that the median of the number of
|
|
determinants when $\mu=0.5$~bohr$^{-1}$ is below 100~000 determinants
|
|
with the quadruple-zeta basis set, making these calculations feasilble
|
|
with such a large basis set. At the double-zeta level, compared to the
|
|
FCI trial wave functions the median of the number of determinants is
|
|
reduced by more than two orders of magnitude.
|
|
Moreover, going to $\mu=0.25$~bohr$^{-1}$ gives a median close to 100
|
|
determinants at the double-zeta level, and close to 1~000 determinants
|
|
at the quadruple-zeta level for only a slight increase of the
|
|
MAE. Hence, RS-DFT-CIPSI trial wave functions with small values of
|
|
$\mu$ could be very useful for large systems to go beyond the
|
|
single-determinant approximation at a very low computational cost
|
|
while keeping the size-consistency.
|
|
|
|
Note that when $\mu=0$ the number of determinants is not equal to one because
|
|
we have used the natural orbitals of a first CIPSI calculation, and
|
|
not the sr-PBE orbitals.
|
|
So the Kohn-Sham determinant is expressed as a linear combination of
|
|
determinants built with natural orbitals. It is possible to add
|
|
an extra step to the algorithm to compute the natural orbitals from the
|
|
RS-DFT/CIPSI wave function, and re-do the RS-DFT/CIPSI calculation with
|
|
these orbitals to get an even more compact expansion. In that case, we would
|
|
have converged to the Kohn-Sham orbitals with $\mu=0$, and the
|
|
solution would have been the PBE single determinant.
|
|
|
|
|
|
|
|
\section{Conclusion}
|
|
|
|
We have seen that introducing short-range correation via
|
|
a range-separated Hamiltonian in a full CI expansion yields improved
|
|
nodal surfaces, especially with small basis sets. The effect of sr-DFT
|
|
on the determinant expansion is similar to the effect of re-optimizing
|
|
the CI coefficients in the presence of a Jastrow factor, but without
|
|
the burden of performing a stochastic optimization.
|
|
|
|
Varying the range-separation parameter $\mu$ and approaching the
|
|
RS-DFT-FCI with CIPSI provides a way to adapt the number of
|
|
determinants in the trial wave function, leading always to
|
|
size-consistent FN-DMC energies.
|
|
We propose two methods. The first one is for the computation of
|
|
accurate total energies by a one-parameter optimization of the FN-DMC
|
|
energy via the variation of the parameter $\mu$.
|
|
The second method is for the computation of energy differences, where
|
|
the target is not the lowest possible FN-DMC energies but the best
|
|
possible cancellation of errors. Using a fixed value of $\mu$
|
|
increases the consistency of the trial wave functions, and we have found
|
|
that $\mu=0.5$~bohr$^{-1}$ is the value where the cancellation of
|
|
errors is the most effective.
|
|
Moreover, such a small value of $\mu$ gives extermely
|
|
compact wave functions, making this recipe a good candidate for
|
|
the accurate description of the whole potential energy surfaces of
|
|
large systems. If the number of determinants is still too large, the
|
|
value of $\mu$ can be further reduced to $0.25$~bohr$^{-1}$ to get
|
|
extremely compact wave functions at the price of less efficient
|
|
cancellations of errors.
|
|
|
|
|
|
|
|
|
|
%%
|
|
\begin{acknowledgments}
|
|
This work was performed using HPC resources from GENCI-TGCC (Grand
|
|
Challenge 2019-gch0418) and from CALMIP (Toulouse) under allocation
|
|
2019-0510.
|
|
\end{acknowledgments}
|
|
|
|
|
|
\bibliography{rsdft-cipsi-qmc}
|
|
|
|
\begin{enumerate}
|
|
\item Total energies and nodal quality:
|
|
\begin{itemize}
|
|
% \item Facts: KS occupied orbitals closer to NOs than HF
|
|
% \item Even if exact functional, complete basis set, still approximated nodes for KS
|
|
\item KS -> exponentially fast convergence (as HF) with basis because of non divergence of effective KS potential (citer le papier de Gill)
|
|
% \item With correlation consistent basis set, FCI nodes (which include correlation) are better than KS
|
|
% \item With FCI, good limit at CBS ==> exact energy
|
|
% \item But slow convergence with basis set because of divergence of the e-e interaction not well represented in atom centered basis set
|
|
% \item Exponential increase of number of Slater determinants
|
|
\item Cite papiers RS-DFT: there exists an hybrid scheme combining fast convergence wr to basis set (non divergent basis set) and short expansion in SCI (cite papier Ferté)
|
|
\item Question: does such a scheme provide better nodal quality ?
|
|
%<<<<<<< HEAD
|
|
\item In RSDFT we cannot optimize energy with $\mu$ , but in FNDMC
|
|
\item Factual stuffs: with optimal $\mu$, lower FNDMC energy than HF/KS/FCI
|
|
\begin{itemize}
|
|
\item less determinants $\Rightarrow$ large systems
|
|
\item only one parameter to optimize $\Rightarrow$ deterministic
|
|
\item $\Rightarrow$ reproducible
|
|
\end{itemize}
|
|
\item with the optimal $\mu$:
|
|
\begin{itemize}
|
|
\item Direct optimization of FNDMC with one parameter
|
|
\item Do we improve energy differences ?
|
|
\item system dependent
|
|
\item basis set dependent: $\mu \rightarrow \infty$ when $\mathcal{B}\rightarrow \text{CBS}$
|
|
\item large wave functions
|
|
\end{itemize}
|
|
% \item Invariance with m_s
|
|
%=======
|
|
\item In RS-DFT we cannot optimize energy with $\mu$ , but in FNDMC
|
|
% \item Factual stuffs: with optimal $\mu$, lower FNDMC energy than HF/KS/FCI
|
|
% \begin{itemize}
|
|
% \item less determinants $\Rightarrow$ large systems
|
|
% \item only one parameter to optimize $\Rightarrow$ deterministic
|
|
% \item $\Rightarrow$ reproducible
|
|
% \end{itemize}
|
|
% \item with the optimal $\mu$:
|
|
% \begin{itemize}
|
|
% \item Direct optimization of FNDMC with one parameter
|
|
% \item Do we improve energy differences ?
|
|
% \item system dependent
|
|
% \item basis set dependent: $\mu \rightarrow \infty$ when $\mathcal{B}\rightarrow \text{CBS}$
|
|
% \item large wave functions
|
|
% \end{itemize}
|
|
%>>>>>>> 44470b89936d1727b638aefd982ce83be9075cc8
|
|
|
|
\end{itemize}
|
|
\end{enumerate}
|
|
|
|
|
|
|
|
\end{document}
|
|
|
|
|
|
% * Recouvrement avec Be : Optimization tous electrons
|
|
% impossible. Abandon. On va prendre H2O.
|
|
% * Manu doit faire des programmes pour des plots de ensite a 1 et 2
|
|
% corps le long des axes de liaison, et l'integrale de la densite a
|
|
% 2 corps a coalescence.
|
|
% 5 Toto optimise les coefs en presence e jastrow
|
|
% 6 Toto renvoie a manu psicoef
|
|
% 7 Manu fait tourner ses petits programmes avec psi_J
|
|
% 8 Toto calcule les energies FN-DMC de
|
|
% +) toutes les fonctions d'ondes psi_mu
|
|
% +) la fonction d'onde psi_J
|
|
|