Quantum Monte Carlo (QMC) is in principle
an exact technique to obtain the ground state properties of a multicomponent
quantum system.
We mention here two ways of doing QMC
calculation. The first is the so called Variational Monte Carlo (VMC),
which is a stochastic way of calculating expectation values of observables
with a given trial wavefunction. The second is the Diffusion Monte Carlo
(DMC), which is a technique to project out the ground state wavefunction
of the system out of a trial wavefunction (provided that the two are not
orthogonal).
With VMC one can only obtain exact properties
if the trial wavefuction is an exact wavefunction of the system, but for
the ground state it is a variational method.
With DMC it is possible, at least in principle,
to obtain the exact ground state wavefuction, and therefore the exact ground
state properties of the system. In practice, however, it is often necessary
to introduce two approximations. The first is the fixed node approximation,
which is related to the fermionic nature of the electrons. The wavefunction
of a system of fermions changes sign at the nodal surface, and this
is usually very difficult to handle as the ground state nodal surface is
usually unknown. With the fixed node approximation the nodal surface is
fixed to that of the trial wavefuction. In this way the problem is much
more manageable, but one only obtains the ground state wavefuction compatible
with this constrain. However, the variational principle holds, and the
error in the ground state energy is only second order in the error in the
nodal surface, being usually very small as very good trial wavefuctions
can ofted be produced using Density Functional Theory (DFT) of Hartree
Fock (HF) methods. The second approximation is the introduction of pseudopotentials
to
eliminate the core states from the calculations. There is a pletora of
data, mainly based on DFT calculations, showing that the pseudopotential
approximation (PP) usually works very well. However, in DMC the use of
non-local PP is responsible for an additional error, due to a difficult
term present in the evaluation of the energy, which is usually neglected.
The error is second order in the error in the trial wavefunction and makes
the DMC energy weakly dependent on it. However, this error is usually very
small, and in principle it is possible to estimate its magnitude by using
different trial wavefunctions (with the same nodal surface).
DFT methods have had enourmous success
since they introduction by Hohenberg and Kohn and Kohn and Sham in the
sixties, and in the last few years their applications have been extended
also to the calculation of finite temperature properties of liquid and
solid. DFT works incredibly well in a large number of cases, and it is
particularly suitable when cancellation of errors is expected, like in
the comparison of the energy (or the free energy) of similar structures.
Our work on the melting temperature of pure iron is a clear example of
that, as both solid and liquid iron have a similar close packed structure.
However, the success of DFT is not comprehensive,
with known failures for strongly correlated materials like transition metal
oxydes for example. An other example of non perfect agreement
with the experiments is the DFT with the Local Density Approximation calculation
of the zero pressure melting point of silicon, for which agreement
with the experiments is only within 20 % (see Sugino and Car, Phys. Rev.
Lett., 74, 1823 (1995).)
The goal of our project is to go beyond
DFT and start using QMC to calculate the properties of materials, including
finite temperature properties with the calculation of free energies. We
will concentrate on those problems for which DFT fails. As a test case
we have chosen to calculate the melting point of silicon at zero pressure,
for which the DFT-LDA error is significantly large.
The melting point will be determined by
calculating the free energy of solid and liquid with DFT first, and then
adding the small correction QMC - DFT, which will be determined using thermodynamic
integration.
This project is being carried out together
with Mike Gillan, and involves
a collaboration with Mike
Towler and Richard Needs
at the Cavendish Laboratory in Cambridge. The calculations will be performed
with the code CASINO.