abinit/doc/theory/pseudopotentials.md

143 lines
6.1 KiB
Markdown

---
authors: MG
---
$$
\newcommand{\aa}{\alpha}
\newcommand{\PS}{{\text{PS}}}
\newcommand{\Ylm}{{Y_m^l}}
\newcommand{\Vloc}{V_{\text{loc}}}
\newcommand{\vv}{\hat v}
\newcommand{\vloc}{\vv_{\text{loc}}}
\newcommand{\Vnl}{V_{\text{nl}}}
\newcommand{\lm}{{lm}}
\newcommand{\KK}{{\bf K}}
\newcommand{\KKp}{{\bf{K'}}}
\newcommand{\KKhat}{\widehat \KK}
\newcommand{\KKphat}{\widehat{\KK}'}
\newcommand{\jl}{j_l}
\newcommand{\rrhat}{{\widehat\rr}}
\newcommand{\dd}{{\,\text{d}}}
$$
## Pseudopotentials in the Kleynman Bylander form
This page reports the basic definitions and equations needed to evaluate the nonlocal part of the Hamiltonian.
We mainly focus on norm-conserving pseudopotentials.
Providing a consistent introduction to the theory of pseudopotentials is beyond the purpose of this section,
a more complete discussion of the topic can be found in specialized articles available in the literature
[[cite:Hamann1979]], [[cite:Bachelet1982]], [[cite:Kleinman1982]], [[cite:Hamann1989]], [[cite:Troullier1991]]
[[cite:Gonze1991]], [[cite:Fuchs1999]].
For the generalization of norm conservation to multiple projectors (Optimized norm-conserving Vanderbilt pseudopotentials)
we refer the reader to [[cite:Hamann2013]].
The generation and validation of ONCVPSP pseudopotentials is described in
the [PseudoDojo](http://www.pseudo-dojo.org/) paper [[cite:Setten2018]].
For our purpose, we can summarize by saying that a pseudopotential is constructed in order to replace
the atomic all-electron potential such that core states are eliminated and valence electrons are
described by pseudo wavefunctions whose representation in the Fourier space decays rapidly.
Modern norm-conserving pseudopotentials are constructed such that the scattering properties
of the all-electron atom are reproduced around the reference energy configuration up to first order in energy,
following the procedure described in [[cite:Hamann1979]], [[cite:Hamann1989]], and [[cite:Troullier1991]].
In modern ab initio codes, the fully separable KB form, proposed by Kleynman and Bylander, is usually employed
as it drastically reduces the number of operations required for the application of the nonlocal part of the Hamiltonian
as well as the memory required to store the operator [[cite:Kleinman1982]].
In the KB form, the interaction between a valence electron and an ion of type $\aa$ is described by means of the operator:
\begin{equation}\label{eq:KBsingleatom}
\vv_\PS^\aa =
\vloc^\aa(r) + \sum_\lm |\chi_\lm^\aa\ra\,E_l^\aa\,\la\chi_\lm^\aa|,
% \vloc_\aa(\rr) + \sum_\lm |\Ylm\,u_{l\aa}^\KB\ra E_{l\aa}^\KB \lau_{l\aa}^\KB\,\Ylm|
\end{equation}
where $\vloc^\aa(r)$ is a purely local potential with a Coulomb tail $\gamma/r$.
!!! note
For simplicity, the discussion is limited to pseudopotentials with a single projector per angular channel.
The generalization to multi-projector pseudopotentials requires introducing an additional $n$ index
in the KB energies i.e. $E_{nl}^\aa$ and in the projectors $\chi_{n\lm}^\aa(\rr)$.
The so-called Kleinman-Bylander energies, $E_l^\aa$,
measure the strength of the nonlocal component with respect to the local part.
The projectors $\chi_\lm^\aa(\rr)$ are
short-ranged functions, expressed in terms of a complex spherical harmonic $\Ylm(\theta,\phi)$
multiplied by an $l$- and atom-dependent radial function:
\begin{equation}
\label{eq:KB_projectors}
\chi_\lm^\aa(\rr) = \dfrac{1}{r}\,\chi_l^\aa(r)\,\Ylm(\theta,\phi).
\end{equation}
<!--
where
\begin{equation}\label{eq:KBfunctionU}
u_{l\aa}^\KB (\rr) =
\frac{\Delta v_{l\aa}(\rr)\,u_{l\aa}^\PS (\rr)}
{\norm{u_{l\aa}(\rr)\,\Delta_{l\aa}}^{1/2}}
\end{equation}
are localized functions defined in terms of the short-ranged ...
and the pseudo eigenfuncions of the reference atom.
\begin{equation}\label{eq:KBenergy}
E_{l\aa}^\KB =
\frac{\la u_{l\aa}^\PS \Delta v_{l\aa} | \Delta v_{l\aa} u_{l\aa}^\PS \ra }
{\la u_{l\aa}^\PS | \Delta v_{l\aa} | u_{l\aa}^\PS \ra}
\end{equation}
-->
The total nonlocal part of the Hamiltonian is obtained by summing the different atom-centered contributions.
The final expression in real space reads:
\begin{equation}
\Vnl(\rr_1,\rr_2)=
\sum_{\substack{\RR \\ \aa\tt_\aa}}
\sum_\lm \la\rr_1-\RR-\tt_\aa|\chi_\lm^\aa\ra E_l^\aa\la\chi_\lm^\aa|\rr_2-\RR-\tt_\aa\ra,
\end{equation}
where $\RR$ runs over the sites of the Bravais lattice, and $\tt_\aa$ over the positions of the atoms with the same type $\aa$ located inside the unit cell.
Due to the nonlocality of the operator, its Fourier transform depends on two separate indices, $\kk+\GG_1$ and $\kk+\GG_2$, instead of the simple difference $\GG_1-\GG_2$.
The shorthand notation $\KK = \kk + \GG$ is used in the following to simplify the derivation.
Using the Rayleigh expansion of planewaves in terms of spherical harmonics $\Ylm(\theta,\phi)$ and spherical Bessel functions $\jl(Kr)$:
\begin{equation}\label{eq:PWinYlmPl}
e^{i\KK\cdot\rr} =
4\pi\,\sum_\lm i^l \jl(Kr) \Ylm^\*(\KKhat)\,\Ylm(\rrhat),
\end{equation}
the Fourier representation of the projector, defined by \ref{eq:KB_projectors}, can be expressed as:
\begin{equation}
\la\KK|\chi_\lm^\aa\ra =
\frac{4\pi}{\Omega^{1/2}}\, (-1)^l\,\Ylm(\KKhat)
\int_0^\infty\, r\jl(Kr)\,\chi_l^\aa(r)r\,\dd r =
\frac{4\pi}{\Omega^{1/2}}\, (-1)^l\,\Ylm(\KKhat) F_l^\aa(K).
\end{equation}
<!--
%where the form factors $F_l^\aa(K)$ related to the atom of type $\aa$ is defined by
%\begin{equation}\label{wq:defformfactors}
% F_l^\aa(K) \df
% \frac{\int_0^\infty r\,j_l (Kr)\,u_{l\aa} \Delta v_{l\aa}\,\dd r}
% {\norm{u_{l\aa} \Delta v_{l\aa}}^{1/2} }
%\end{equation}
-->
Finally, the expression for the total nonlocal operator in reciprocal space is:
\begin{equation}
\label{eq:VnlKBmatrixelements}
\Vnl(\KK,\KKp)= \frac{(4\pi)^2}{\Omega} \sum_{\aa \tt_\aa} \sum_\lm
\,e^{-i(\KK-\KKp)\cdot\tt_\aa}\,
\Ylm(\KKhat)\Ylm^\*(\KKphat)\, E_l^\aa F_l^\aa(K) F_l^\aa(K').
\end{equation}
!!! note
Abinit employs iterative eigenvalue solvers to solve the KS equations.
This means that we only need to compute $\Vnl |\Psi\ra$, i.e., we only
need to **apply** the Hamiltonian onto a set of trial eigenvectors.
Therefore the full $\Vnl(\KK,\KKp)$ matrix is never constructed explicitly.