Finite-temperature DFT calculations using the Ensemble-DFT method

Author:: Álvaro Ruiz Serrano, University of Southampton
Author:: Extended to spin relaxation by Kevin Duff, University of Cambridge
Author:: Extended to include Grand Canonical Ensemble by Arihant Bhandari, University of Southampton
Author:: Extended to include Pulay mixing and input trial step by Gabriel Bramley, University of Southampton
Date:: August 2013
Date:: Extended by Kevin Duff April 2018
Date:: Extended by Arihant Bhandari September 2020
Date:: Extended by Gabriel Bramley November 2020

Basic principles

This manual describes how to run finite-temperature calculations using ONETEP. A recent implementation uses a direct minimisation technique based on the Ensemble-DFT method [Marzari1997]. The Helmholtz free energy functional is minimised in two nested loops. The inner loop performs a line-search in the space of Hamiltonian matrices (in a similar fashion as described in Ref. [Freysoldt2009]), for a fixed set of NGWFs. Then, the outer loop optimises the NGWFs psinc expansion coefficients for a fixed density kernel. For a more detailed description and discussion of this method in ONETEP, see Ref. [Ruiz-Serrano2013].

Using the NGWF representation, the Helmholtz free energy functional becomes:

(3)\[\begin{aligned} A_{\mathcal{T}}\left[{\left\lbrace H_{\alpha\beta} \right\rbrace},{\left\lbrace \lvert\phi_\alpha\rangle \right\rbrace}\right] = E\left[{\left\lbrace H_{\alpha\beta} \right\rbrace},{\left\lbrace \lvert\phi_\alpha\rangle \right\rbrace}\right] - {\mathcal{T}}S\left[{\left\lbrace f_i \right\rbrace}\right].\end{aligned}\]

where \({\left\lbrace H_{\alpha\beta} \right\rbrace}\) is the NGWF representation of the Hamiltonian matrix, \({\left\lbrace \lvert\phi_\alpha\rangle \right\rbrace}\) is the current set of NGWFs, \({\mathcal{T}}\) is the electronic temperature, \(E\left[{\left\lbrace H_{\alpha\beta} \right\rbrace},{\left\lbrace \lvert\phi_\alpha\rangle \right\rbrace}\right]\) is the energy functional and \(S\left[{\left\lbrace f_i \right\rbrace}\right]\) is an entropy term. The occupancies of the Kohn-Sham states, \({\left\lbrace f_i \right\rbrace}\) are calculated from the energy levels, \({\left\lbrace \epsilon_i \right\rbrace}\), using the Fermi-Dirac distribution:

\[\label{eq:fermidirac} f_i\left(\epsilon_i\right) = \left( 1 + \exp\left[\dfrac{\epsilon_i - \mu}{{k_\textrm{B}}{\mathcal{T}}}\right] \right)^{-1}.\]

where \(\mu\) is the Fermi level. To obtain the energy eigenvalues, \({\left\lbrace \epsilon_i \right\rbrace}\), the Hamiltonian matrix is diagonalised as:

\[\label{eq:hamdiag} H_{\alpha\beta} {M^\beta_i} = S_{\alpha\beta} {M^\beta_i} \epsilon_i,\]

where \({\left\lbrace S_{\alpha\beta} \right\rbrace}\) are the elements of the NGWF overlap matrix, and \({\left\lbrace {M^\beta_i} \right\rbrace}\) are the expansion coefficients of the Kohn-Sham eigenstates in the NGWF basis set. At the moment, this is a cubic-scaling operation that requires dealing dense matrices, which makes it memory-demanding.

Free- and fixed-spin EDFT

By default in spin polarized runs, the total occupancy of each spin channel is held fixed; each spin channel has its own Fermi level determined by this constraint. Alternatively the whole system can be held at one Fermi level dictated by the conservation of the total number of electrons in the system, allowing the net spin to freely relax.

Free-spin EDFT should be appropriate for most applications unless there’s a reason to hold the system fixed at a given net spin. As with any minimization with potentially many minima, the final state may depend on initial conditions. As a special case, free-spin EDFT may not be able to symmetry-break a system that wants to have any kind of spin polarization but that is initialized to have 0 net spin. The general advice for simple systems like basic ferromagnets (though this should not replace good system-specific judgment) is to slightly over-specify the expected net spin on each atom and hold the spin fixed for a few iterations before being allowed to relax. For example a cobalt cluster is expected to have a net spin per atom lower than that of an isolated atom, that decreases to bulk-like as a function of cluster size. A good initialization may be to give each atom atomic-like net spin and hold the net spin fixed for 3-5 NGWF CG iterations, then allow it to relax.

Compilation

By default, ONETEP is linked against the Lapack library [lapack_web] for linear algebra. The Lapack eigensolver DSYGVX [DSYGVX], can only be executed in one CPU at a time. Therefore, EDFT calculations with Lapack are limited to small systems (a few tens of atoms). Calculations on large systems are possible if, instead, ONETEP is linked against ScaLapack library [scalapack_web] during compilation time. The ScaLapack eigensolver, PDSYGVX, can be run in parallel using many CPUs simultaneously. Moreover, ScaLapack can distribute the storage of dense matrices across many CPUs, thus allowing to increase the total memory allocated to a given calculation in a systematic manner, simply by requesting more processors. For the compilation against ScaLapack to take effect, the flag -DSCALAPACK must be specified during the compilation of ONETEP.

Pulay Mixing EDFT

In default EDFT, the Hamiltonian is updated using a damped fixed point update routine:

\[\label{linearmixing} H_{\alpha\beta}^{(m+1)} = H_{\alpha\beta}^{(m)} + \lambda \, R[H_{\alpha\beta}^{(m)}]\]

Where the \(\lambda\) defines the mixing parameter and residual is defined as:

\[\label{residual} R[H_{\alpha\beta}^{(m)}] = \tilde{H}_{\alpha\beta}^{(m)} - H_{\alpha\beta}^{(m)}\]

Where \(\tilde{H}_{\alpha\beta}^{(m)}\) is the diagonlised Hamiltonian obtained at step m. At a sufficiently low value of \(\lambda\), most systems will achieve convergence, but at an increasingly slow rate as the system increases in size. Convergence can be accelerated using quasi-Newton update methods such as Broyden or Pulay methods, the latter of which is implemented in EDFT as an alternative to the damped fixed point method.

The implementation in ONETEP uses a similar logic to other DFT implementations of Pulay’s method, except the Hamiltonian is optimised instead of the density:

\[H_{\alpha\beta}^{(m+1)} = \sum_{j=m-n+1}^{m} c_j H_{\alpha\beta}^{(j)} + \lambda \sum_{j=m-n+1}^{m} c_j R[H_{\alpha\beta}^{(j)}]\]

Where the history length is defined \(n\) and the co-efficients \(c_j\) are obtained through the procedure outlined by Ref. [Kresse1996]. For the systems tested, this method leads to improved convergence, especially for larger metallic systems. Further information can be found in Ref. [Woods2019].

Increased Calculation Speed Using Fixed Step Sizes

As described in the Section on Pulay mixing, \(\lambda\) defines the step length taken at each inner loop iteration. In the default algorithm, an optimal \(\lambda\) value which gives the greatest decrease in the Lagrangian is determined by a line search routine. Although this improves the robustness of the algorithm, the line search requires two or more energy evaluations per inner loop step to obtain the optimum \(\lambda\) value. If \(\lambda\) varies very little over the course of the calculation, this can double the computational expense of each inner loop iteration for a negligible increase in the accuracy for each step.

Alternatively, one can fix the \(\lambda\) to a reasonable value for a significant speed-up by ensuring only one energy evaluation is performed per inner loop iteration. However, this option is less robust than the default line search algorithm, as the fixed \(\lambda\) value may produce either sub-optimal energy decreases or energy increases for certain steps. Furthermore, if \(\lambda\) is chosen to be too high, your answer may diverge from the ground state by taking several consecutive positive Lagrangian steps (A warning will be provided if this occurs too often). Conversely, convergence will be very slow if \(\lambda\) is chosen to be too low. \(\lambda\) is set with the edft_trial_step keyword, which switches from the line search algorithm if greater than 0, and uses the fixed \(\lambda\) value specified.

User input values of \(\lambda\) can be determined by running a standard EDFT calculation for a single NGWF iteration with line search and plotting the ’step’ value printed at each iteration (in VERBOSE output mode). The safest option is to choose a value close to the minimum step value, but a slightly higher value can be selected, especially if larger step values are common. The first two steps of your calculation choose \(\lambda\) with line search regardless of your input, as optimal step sizes for these iterations are significantly larger than subsequent inner loop iterations. As such, these two iterations should be disregarded from your \(\lambda\) value selection analysis. As step sizes which yield stable convergence are system dependent, it is recommended to manually determine different \(\lambda\) values for systems with large differences in species or size.

Commands for the inner loop

Basic setup

edft: T/F [Boolean, default edft: F]. If true, it enables Ensemble-DFT calculations.
edft_maxit: n [Integer, default edft_maxit: 10]. Number of EDFT iterations in the ONETEP inner loop.
edft_smearing_width: x units [Real physical, default edft_smearing_width: 0.1 eV]. Sets the value of the smearing width, \({k_\textrm{B}}{\mathcal{T}}\), of the Fermi-Dirac distribution. It takes units of energy (eV, Hartree) or temperature. For example, edft_smearing_width: 1500 K will set \({\mathcal{T}}=\) 1500 degree Kelvin.
edft_update_scheme: damp_fixpoint/pulay_mix [Character, default dft_update_scheme: damp_fixpoint]. Defines the mixing scheme for EDFT in the ONETEP inner loop.
edft_ham_diis_size: x [Integer, default edft_ham_diis_size: 10]. Specifies the maximum number of Hamiltonians used from previous iterations to generate the new guess through Pulay mixing.
spin: x [Real, default spin: 0.0]. For EDFT runs this value does not need to be an integer. Because we are considering an ensemble of states it can have any real value between \(-\frac{n_\mathrm{elec}}{2}\) to \(\frac{n_\mathrm{elec}}{2}\). Make sure you have enough bands to cover the more populated spin channel.
edft_spin_fix [Integer, default edft_spin_fix: -1]. Control for whether the net spin of the system should remain fixed at spin, or relax during the run. Any negative number will fix the net spin. Nonnegative numbers \(n\) will hold the net spin fixed for \(n\) iterations then let it relax for the rest of the calculation.
edft_trial_step [Integer, default edft_trial_step: 0]. Sets the value of \(\lambda\), which fixes the step size in the EDFT inner loop, and switches off the line search for optimum \(\lambda\) values. If set to 0, the normal line search routine is used.

Tolerance thresholds

edft_free_energy_thres: x units [Real physical, default edft_free_energy_thres: 1.0e-6 Ha/Atom]. Maximum difference in the Helmholtz free energy functional per atom between two consecutive iterations.
edft_energy_thres: x units [Real physical, default edft_energy_thres: 1.0e-6 Ha/Atom]. Maximum difference in the energy functional per atom between two consecutive iterations.
edft_entropy_thres: x units [Real physical, default edft_entropy_thres: 1.0e-6 Ha/Atom]. Maximum difference in the entropy per atom functional between two consecutive iterations.
edft_rms_gradient_thres: x [Real, default edft_rms_gradient_thres: 1.0e-4]. Maximum RMS gradient \(\dfrac{d A_{\mathcal{T}}}{d f_i}\).
edft_commutator_thres: x units [Real physical, default edft_commutator_thres: 1.0e-5 Hartree]. Maximum value of the Hamiltonian-Kernel commutator.
edft_fermi_thres: x units [Real physical, default edft_fermi_thres: 1.0e-3 Hartree]. Maximum change in the Fermi energy between two consecutive iterations.

Advanced setup

edft_extra_bands: n [Integer, default edft_extra_bands: -1]. Number of extra energy bands. The total number of bands is equal to the number of NGWFs plus edft_extra_bands. When set to a negative number, no extra bands are added.
edft_round_evals: n [Integer, default edft_round_evals: -1]. When set to a positive integer value, the occupancies that result from the Fermi-Dirac distribution are rounded to n significant figures. This feature can reduce some numerical errors arising from the grid-based representation of the NGWFs.
edft_write_occ: T/F [Boolean, default edft_write_occ: F]. Save fractional occupancies in a file.
edft_max_step: x [Real, default edft_max_step: 1.0]. Maximum step during the EDFT line search.

Commands for the outer loop

The standard ONETEP commands for NGWF optimisation apply to the EDFT calculations as well. The only flag that is different is:

ngwf_cg_rotate: T/F [Integer, default ngwf_cg_rotate: T]. This flag is always true in EDFT calculations. It ensures that the eigenvectors \({M^\beta_i}\) are rotated to the new NGWF representation once these are updated.

Restarting an EDFT calculation

write_hamiltonian: T/F [Boolean, default write_hamiltonian: F]. Save the last Hamiltonian matrix on a file.
read_hamiltonian: T/F [Boolean, default read_hamiltonian: F]. Read the Hamiltonian matrix from a file, and continue the calculation from this point.
write_tightbox_ngwfs: T/F [Boolean, default write_tightbox_ngwfs: T]. Save the last NGWFs on a file.
read_tightbox_ngwfs: T/F [Boolean, default read_tightbox_ngwfs: F]. Read the NGWFs from a file and continue the calculation from this point.

If a calculation is intended to be restarted at some point in the future, then run the calculation with

write_tightbox_ngwfs: T

write_hamiltonian: T

to save the Hamiltonian and the NGWFs on disk. Two new files will be created, with extensions .ham and .tightbox_ngwfs, respectively. Then, to restart the calculation, set

read_tightbox_ngwfs: T

read_hamiltonian: T

to tell ONETEP to read the files that were previously saved on disk. Remember to keep a backup of the output of the first run before restarting the calculation.

the density kernel is not necessary to restart an EDFT calculation. However, it is necessary to calculate the electronic properties of the system, once the energy minimisation has completed. To save the density kernel on a file, set: write_denskern: T

to generate a .dkn file containing the density kernel. To read in the density kernel, set

read_denskern: T

Controlling the parallel eigensolver

Currently, only the ScaLapack PDSYGVX parallel eigensolver is available. A complete manual to this routine can be found by following the link in Ref. [PDSYGVX]. If ONETEP is interfaced to ScaLapack, the following directives can be used:

eigensolver_orfac: x [Real, default eigensolver_orfac: 1.0e-4]. Precision to which the eigensolver will orthogonalise degenerate Hamiltonian eigenvectors. Set to a negative number to avoid reorthogonalisation with the ScaLapack eigensolver.
eigensolver_abstol: x [Real, default eigensolver_abstol: 1.0e-9]. Precision to which the parallel eigensolver will calculate the eigenvalues. Set to a negative number to use ScaLapack defaults.

The abovementioned directives are useful in calculations where the ScaLapack eigensolver fails to orthonormalise the eigenvectors. In such cases, the following error will be printed in the input file:

(P)DSYGVX in subroutine dense_eigensolve returned info= 2.

Many times (although not always) this error might cause the
calculation to fail. If this situation occurs, set
eigensolver_orfac: -1
eigensolver_abstol: -1
in the input file and restart the calculation. ScaLapack will not
reorthonormalise the eigenvectors. Instead, an external Löwdin
orthonormalisation process [Lowdin1950] will be
triggered. This is usually more efficient for larger systems.

Grand Canonical Ensemble DFT

In simulations of electrochemical electrodes, the electrons can freely exchange between the electrode and the electrical circuit. So, there is no constraint on the number of electrons \(N\). Rather, the electrode potential \(U\) is fixed, with respect to a reference electrochemical potential \(\mu_{ref}\) which fixes the chemical potential of electrons \(\mu\):

\[\mu = \mu_{ref} -eU\]

Typical experiments use a standard hydrogen electrode as the reference electrode with \(\mu_{ref}^{SHE}=-4.44\) eV. Once the chemical potential of electrons is fixed, the number of electrons changes as a dependent variable according to the Fermi-Dirac distribution in eq. .

\[N = \sum_i f_i\]

Thermodynamically, this corresponds to switching the electrons from the finite-temperature, fixed-number canonical ensemble to the finite-temperature, fixed-potential grand-canonical ensemble. Correspondingly, the relevant free energy minimized at equilibrium is the grand potential [Sundararaman2017]:

\[\Omega = A -\mu N\]

The following keywords are used for the grand-canonical ensemble DFT:

edft_grand_canonical: T/F [Boolean, default edft_grand_canonical: F]. Switch to fixed-potential grand-canonical ensemble.
edft_reference_potential: x units [Real physical, default edft_reference_potential: -4.44 eV]. Set the reference potential \(\mu_{ref}\). If no units are given, atomic units are considered: Ha (hartrees).
edft_electrode_potential: x units [Real physical, default edft_electrode_potential: 0.0 V]. Set the electrode potential \(U\). If no units are given, atomic units are considered: Ha/e, hartrees per elementary charge.
edft_nelec_thres: x [Real, default edft_nelec_thres: 1.0e-06 per atom]. Convergence threshold on the change in number of electrons per spin channel per atom.

[Sundararaman2017] R. Sundararaman, W. Goddard, and T. Arias. J. Chem. Phys., 146(11):114104, 2017.

[Marzari1997] N. Marzari, D. Vanderbilt, and M. C. Payne. Phys. Rev. Lett., 79(7):1337–1340, 1997.

[Freysoldt2009] C. Freysoldt, S. Boeck, and J. Neugebauer. Phys. Rev. B, 79(24):241103, 2009.

[Ruiz-Serrano2013] A. Ruiz-Serrano and C.-K. Skylaris. A variational method for density functional theory calculations on metallic systems with thousands of atoms. J. Chem. Phys., 139(5):054107, 2013.

[Lapack_web] Lapack. http://www.netlib.org/lapack/.

[DSYGVX] Lapack DSYGVX eigensolver. http://netlib.org/lapack/double/dsygvx.f.

[Scalapack_web] ScaLapack. http://www.netlib.org/scalapack/.

[PDSYGVX] ScaLapack PDSYGVX eigensolver. http://www.netlib.org/scalapack/double/pdsygvx.f.

[Lowdin1950] Per-Olov Lowdin. On the non-orthogonality problem connected with the use of atomic wave functions in the theory of molecules and crystals. J. Chem. Phys., 18(3):365–375, 1950.

[Kresse1996] G. Kresse and J. Furthmüller. Efficient iterative schemes for ab initio total-energy calculations using a plane-wave basis set. Phys. Rev. B, 54:11169, 1996.

[Woods2019] N. Woods, M. Payne and P. Hasnip. Computing the self-consistent field in Kohn–Sham density functional theory J. Phys. Condens. Matter, 31:453001, 2019.