QMC Practice in a Nutshell
The aim of this section is to provide a very brief overview of the essential concepts undergirding Quantum Monte Carlo calculations of electronic structure with a particular focus on the key approximations and quantities to converge to achieve high accuracy. The discussion here is not intended to be comprehensive. For deeper perspectives on QMC, please see the review articles listed in Quantum Monte Carlo: Theory and Practice.
VMC and DMC in the abstract
Ground state QMC methods, such as Variational (VMC) and Diffusion (DMC) Monte Carlo, attempt to obtain the ground state energy of a many-body quantum system.
The VMC method obtains an upper bound on the ground state energy (guaranteed by the Variational Principle) by introducing a guess at the ground state wavefunction, known as the trial wavefunction \(\Psi_T\):
The DMC method improves on this variational bound by projecting out component eigenstates of the trial wavefunction lying higher in energy than the ground state. The operator that acts as a projector is the imaginary time, or thermodynamic, density matrix:
The DMC energy approaches the ground state energy from above as the imaginary time becomes large.
However from the equations above, one can already anticipate that the DMC method will struggle in the face of degeneracy or near-degeneracy.
In principle, the DMC method is exact for the ground state, but further complications arise for systems that are extended, comprised of fermions, or contain heavy nuclei, pseudized or otherwise. Approximations arising from the numerical implementation of the method also require care to keep under control.
From expectation values to random walks
Evaluating expectation values of a many-body system involves performing high dimensional integrals (the dimensionality is at least the dimensions of the physical space times the number of particles). In VMC, for example, the expectation value of the total energy is represented succinctly as:
where \(E_L\) is the local energy \(E_L=\Psi_T^{-1}\hat{H}\Psi_T\). The other factor in the integral \(|{\Psi_T}|^2\) can clearly be thought of as a probability distribution and can therefore be sampled by Monte Carlo methods (such as the Metropolis algorithm) to evaluate the integral exactly.
The sampling procedure takes the form of random walks. A “walker” is just a set of particle positions, along with a weight, that evolves (or moves) to new positions according to a set of statistical rules. In VMC as few walkers are used as possible to reduce the equilibration time (the number of steps or moves required to lose a memory of the potentially poor starting guess for particle positions). In DMC, the walker population is a dynamic feature of the calculation and must be large enough to avoid introducing bias in expectation values.
The tradeoff of moving to a the sampling procedure for the integration is that it introduces statistical error into the calculation which diminishes slowly with the number of samples (it falls off like \(1/(\# of samples)\) by the Central Limit Theorem). The good news for ground state QMC is that this error can be reduced more rapidly through the discovery of better guesses at the detailed nature of the many-body wavefunction.
Quality orbitals: planewaves, cutoffs, splines, and meshes
Acting on an understanding of perturbation theory, the zeroth order representation of the wavefunction of an interacting system takes the form of a Slater determinant of single particle orbitals. In practice, QMC calculations often obtain a starting guess at these orbitals from Hartree-Fock or Density Functional Theory calculations (which already contain non-perturbative contributions from correlation). An important factor in the generation and use of these orbitals is to ensure that they are described to high accuracy within the parent theory.
For example, when taking orbitals from a planewave DFT calculation, one
must take care to converge the planewave energy cutoff to a sufficient
level of accuracy (usually far beyond what is required to obtain an
accurate DFT energy). One criterion to use it to converge the kinetic
energy of the Kohn-Sham wavefunction with respect to the planewave
energy cutoff until it is accurate to the energy scale you care about in
your production QMC calcuation. For systems with a small number of
valence electrons, a cutoff of around 200 Ry is often sufficient. To
obtain the kinetic energy from a PWSCF calculation the pw2casino.x
post-processing tool can be used. In Nexus one has the option to compute
the kinetic energy by setting the kinetic_E
flag in the
standard_qmc
or basic_qmc
convenience functions.
For efficiency reasons, QMC codes often use a real-space representation of the wavefunction. It is common to represent the orbitals in terms of B-splines which have control points, or knots, that fall on a regular 3-D mesh. Analogous to the planewave cutoff, the fineness of the B-spline mesh controls the quality of the represented orbitals. To verify that the quality of the orbitals has not been compromised during the conversion process from planewave to B-spline, one often performs a VMC calculation with the B-spline Slater determinant wavefunction to obtain the kinetic energy. This value should agree with the kinetic energy of the planewave representation within the energy scale of interest.
In QMCPACK, the B-spline mesh is controlled with the meshfactor
keyword. Larger values correspond to finer meshes. A value of
\(1.0\) usually gives a similar quality representation as the
original planewave calculation. Control of this parameter is made
available in Nexus through the meshfactor
keyword in the
standard_qmc
or basic_qmc
convenience functions.
Quality Jastrows: less variance = more efficient
aking a further cue from perturbation theory, the first order correction to the Slater determinant wavefunction is the Jastrow correlation prefactor.
In a quantum liquid, an appropriate form for the Jastrow factor is:
This form is often used without modification in electronic structure calculations. Note that the correlation factors \(u_{ij}\) can be different for particles of differing species, or, if one of the particles in the pair is classical (such as a heavy atomic nucleus), the local electronic environment varies across the system.
The primary role of the Jastrow factor is to increase the efficiency of the QMC calculation. The variance of the local energy across all samples of the random walk is directly related to the statistical error of the final results:
The variance of local energy is usually minimized by performing a statistical optimization of the Jastrow factor with QMC.
In addition to selecting a good form for the pair correlation functions \(u_{ij}\) (which are represented in QMCPACK as 1-D B-spline functions with a finite cutoff radius), the (iterative) optimization procedure must be performed with a sufficient number of samples to converge all the free parameters. Starting with a small number of samples (\(\approx 20,000\)) is usually preferable for early iterations, followed by a larger number for later iterations. This larger number is something close to \(100,000\times (\#~of~free~parameters)^2\). For B-spline functions, the number of free parameters is the number of control points, or knots.
The number of samples is controlled with the samples
keyword in
QMCPACK. Control of this parameter is made available in Nexus through
the samples
keyword in the linear
or cslinear
convenience
functions (Which are often used in conjunction with standard_qmc
or
basic_qmc
). For a B-spline correlation factor, the number of free
parameters/knots is indicated by the size
keyword in either QMCPACK
or Nexus.
Finite size effects: k-points, supercells, and corrections
For extended systems, finite size errors are a key consideration. In addition to the finite size effects that are typically seen in DFT (k-points related). Correlated, many-body methods such as QMC also must contend with correlation-related finite size effects. Both types of finite-size effects are reduced by simply using larger supercells. The complete elimination of finite size effects using this approach can be prohibitively costly since the finite size error typically falls off like \(1/\Omega_C\), where \(\Omega_C\) is the volume of the supercell. A more sophisticated approach involves a combination of the supercell size, k-point grid, and additional estimated corrections for correlation finite size effects.
Although there is no firm rule on the selection of these three elements, adhering to some general guidelines is usually helpful. For a production calculation of an extended system, the minimum supercell size is around 50 atoms. The size of the supercell k-point grid can then be determined by proxy with a DFT calculation (converge the energy down to the scale of interest). Note that although the cost of a DFT calculation scales linearly with the number of k-points, the cost of the corresponding QMC calculation is hardly increased due to the statistical averaging of the results (the QMC calculation at each separate supercell k-point is simply performed with fewer samples so that the total number of samples remains fixed w.r.t. the number of k-points). Finally, corrections for correlation-related finite size effects are computed during the QMC run and added to the result by hand in post-processing the data.
In Nexus, the supercell size is controlled through the tiling
parameter in the generate_physical_system
, generate_structure
,
Structure
, or Crystal
convenience functions. Supercells can also
be constructed by tiling existing structures through the tile
member
function of Structure
or PhysicalSystem
objects. The k-point
grid is controlled through the kgrid
parameter in the
generate_physical_system
, generate_structure
, Structure
, or
Crystal
convenience functions. K-point grids can also be added to
existing structures through the add_kmesh
member function of
Structure
or PhysicalSystem
objects.
Imaginary time discretization: the DMC timestep
An analytic form for the imaginary time projection operator is not known, but real-space approximations to it can be obtained in the small time limit. With importance sampling included (not covered here), the short-time projector splits into two parts, known as the drift-diffusion and branching factors (shown below in atomic units):
The long-time projector is found as the product of many approximate short-time solutions, which takes the form of a many-body path integral in real space:
The short-time parameter \(\tau\) is known as the DMC timestep and accurate quantities are obtained only in the limit as \(\tau\) approaches zero.
Ensuring that the timestep error is sufficiently small usually involves performing many DMC calculations over a range of timesteps (sometimes on a smaller supercell than the production calculation). The largest timestep is chosen that produces a bias smaller than the energy scale of interest. For very high accuracy, one uses the total energy as a function of timestep to extrapolate to the zero time limit.
The DMC timestep is made available in Nexus through the timestep
parameter of the dmc
convenience function (which is often used in
conjunction with the standard_qmc
, basic_qmc
,
generate_qmcpack
, or Qmcpack
functions).
Population control bias: safety in numbers
While the drift-diffusion factor \(G_d(R',R;\tau)\) can be sampled exactly using Gaussian distributed random numbers (this generates the DMC random walk), the branching factor \(G_b(R',R;\tau)\) is handled a different way for efficiency. The product of branching factors over an imaginary time trajectory (random walk) serves as a statistical weight for each walker. The fluctuations in this weight rapidly become quite large as the random walk progresses (because it approaches an infinite product of real numbers). As its name suggests, this weight factor is used to “branch” walkers every few steps. If the weight is small the walker is deleted, but if the weight is large the walker is copied many times (“branched”) with each copy carrying a weight close to unity. This is more efficient because more walkers are created (and thus more statistics are gathered) in the high weight regions of phase space that contribute most to the integral.
The branching process in DMC naturally leads to a fluctuating population of walkers. The fluctuations in the walker population, if left to its own dynamics, are unbounded. This means that the walker population can grow very large, or even become zero. To prevent collapse of the walker population, population control techniques (not covered here) are added to the algorithm. The practical upshot of population control is that it introduces a systematic bias in the DMC results that scales like \(1/(\# of walkers)\) (Although note that another route to reduce the population control bias is to improve the trial wavefunction, since the fluctuations in the branching weights will become zero for the exact ground state).
For many production calculations, population control bias is not much of an issue because the simulations are performed on supercomputers with thousands of cores per run, and thus tens of thousands of walkers. As a rule of thumb, the walker population should at least number in the thousands. One should occasionally explicitly check the magnitude of the population control bias for the system under study since predictions have been made that it will eventually diverge exponentially with the number of particles in the system.
The DMC walker population can be directly controlled in QMCPACK or Nexus
through the samples
(total walker population) or
samplesperthread
(walkers per OpenMP thread) keywords in the VMC
block directly proceeding DMC (vmc
convenience function in Nexus).
If you opt to use the samples
keyword, check that each thread in the
calculation will have at least a few walkers.
The fixed node/phase approximation: varying the nodes/phase
For every fermionic system, the bosonic ground state lies lower in energy than the fermionic ground state. This means that projection methods like DMC will approach the bosonic ground state exponentially fast in imaginary time if unconstrained (this would show up as an exponentially diverging statistical error). In order to guarantee that the projected wavefunction remains in the space of fermionic functions (and consequently that the projected energy remains an upper bound to the fermionic ground state energy), the projected wavefunction is constrained to share the nodes (if it is real-valued) or the phase (if it is complex-valued) of the trial wavefunction. The fixed node/phase approximation represents one of the two most important approximations for electronic structure calculations (the other is the pseudopotential approximation covered in the next section).
The fixed node/phase error can be reduced, but it cannot be completely eliminated unless the exact nodes/phase is known. A common approach to reduce the fixed node/phase error is to perform several DMC calculations (sometimes on a smaller supercell) with different sets of orbitals (perhaps generated with different functionals). Another, more expensive approach, is to include the backflow transformation (this is the second order correction to the wavefunction; it is not covered in any detail here) to get a lower bound on how large the fixed node error is in standard Slater-Jastrow calculations.
To perform a calculation of this type (scanning over orbitals from
different functionals) with Nexus, the DFT functional can be selected
with the functional
keyword in the standard_qmc
or basic_qmc
convenience functions. If you are using pseudopotentials generated for
use in DFT, you should maintain consistency between the functional and
pseudopotential. Even if such consistency is maintained, the impact of
using DFT pseudopotentials (or those made with many other theories) in
QMC can be significant.
Pseudopotentials: theoretical dissonance, the locality approximation, and T-moves
The accurate use of pseudopotentials in electronic structure QMC calculations remains one of the largest challenges in current practice. The necessity for pseudopotentials arises from the rapidly increasing computational cost with increasing nuclear charge (it scales like \(Z^6\), compared with the \(N_{electrons}^3\) scaling with \(Z\) fixed). The challenge in using pseudopotentials in QMC is that practically no pseudopotentials exist that have been generated self-consistently with QMC. In other words, QMC is currently reliant on other theories to provide the pseudopotentials, which can be a critical source of error.
The current state-of-the-art is not without rigor, however. One source of Dirac-Fock based pseudopotentials, the Burkatzki-Filippi-Dolg database (see http://www.burkatzki.com/pseudos/index.2.html), has been explicitly vetted against quantum chemistry calculations of atoms (a higher-fidelity proxy for QMC calculations of small systems). It must be stressed that these pseudopotentials should still be validated for use in a particular target system. Another collection of Dirac-Fock pseudopotentials that have been created for use in QMC can be found in the Trail-Needs database (see http://www.tcm.phy.cam.ac.uk/~mdt26/casino2_pseudopotentials.html). Many current calculations also use the OPIUM package (see http://opium.sourceforge.net/) to generate DFT pseudopotentials and then port them directly to QMC.
Whatever the source of pseudopotentials (but perhaps especially so for those derived from DFT), testing and validation remains an important step preceding production calculations. One option is to perform parallel pseudopotential and all-electron DMC calculations of atoms with varying electron count (i.e. ionization potential/electron affinity calculations). As with any electronic structure calculation, it is also advisable to devise a test in or close to the target host environment. Validating pseudopotentials remains a difficult task, and while the suggestions presented here may be of some help, they do not amount to a panacea for the issue.
Beyond the central approximation of using a pseudopotential at all, two approximations unique to pseudopotential use in DMC merit discussion. The direct use of non-local pseudopotentials in DMC leads to a second sign-problem (akin to the fixed-node issue) in the imaginary time projector. One solution, devised first, is known as the locality approximation. In the locality approximation, the non-local pseudopotential is replaced by a “localized” form: \(V_{NLPP}\rightarrow \Psi_T^{-1}V_{NLPP}\Psi_T\). This approximation becomes exact as the trial wavefunction approaches the pseudo ground state, however the Variational Principle of the pseudo-system is lost (though it should be acknowledged that a non-variational portion of the energy has been discarded by using pseudopotentials at all). The Variational Principle for the pseudo-system can be restored with an advanced sampling technique known as T-moves (although the first incarnation of the technique reduces to the locality approximation as the system becomes larger than several atoms, the second version fixes this oversight).
One can select whether to use the locality approximation or T-moves
(version 1!) in QMCPACK from within Nexus by setting the parameter
nonlocalmoves
to True or False in the dmc
convenience function.
Other approximations: what else is missing?
Though a few points could be selected for mention at this point, only one additional approximation will be highlighted here. In most modern QMC calculations of electronic structure, relativistic effects have been neglected entirely (there have been a few exceptions) or simply assumed to be covered by the pseudopotential. Clearly this will become an issue for systems with large effective core charges. At present, relativistic corrections are not available within QMCPACK.