7.3 GW and BSE: a staged advanced workflow

These are unexecuted teaching inputs and starting settings to test. Original figures are schematics, not computed results. Use licensed VASP and PAW data, replace every placeholder, record the executable version and validate convergence.

7.3.1 Model, units and provenance

Use eV for energy, Å for length and eV/Å for force; 1 kbar = 0.1 GPa. State normalization per atom, molecule, primitive cell or simulation cell. Record PAW identifiers, release, ZVAL, ENMAX and permitted hashes; never redistribute POTCAR. SCF convergence addresses the chosen electronic problem; convergence of the target property requires separate tests.

GW and BSE: a staged advanced workflow — original conceptual schematic

Original schematic. Curves explain concepts; blank data areas await verified learner results. No calculation is claimed.

7.3.2 Worked case: procedure, interpretation and checks

The goal and the boundary of this tutorial

A quasiparticle energy describes adding or removing an electron while the other electrons respond. An optical excitation creates an electron and a hole together. GW approximates the electronic self-energy for the former; the Bethe–Salpeter equation (BSE) treats correlated electron–hole response for the latter within its selected kernel. Thus a quasiparticle gap and an optical excitation energy are different objects. Subtracting an optical peak from a gap is only a well-defined exciton binding estimate when the gap, transition, momentum, and state assignments are consistent.

This is an advanced workflow map for a small three-dimensional nonmagnetic insulator, not a turnkey recipe for metals, molecules, charged cells, 2D layers, strongly correlated compounds, or every SOC combination. Those cases need additional electrostatic, starting-point, and implementation checks. Begin with the official training examples before scaling to a large target. The VASP BSE course includes diamond, LiF, and reciprocal-space sampling exercises; its purpose-built cases are better validation targets than an expensive unexplained first run. Official BSE course, official Si staged exercise

Stage A: trustworthy ground-state orbitals

Converge a static semilocal ground state on a regular mesh using a suitable PAW set, ideally assessed for high-energy unoccupied-state accuracy. GW-labelled PAW datasets can be important, but the label alone does not guarantee convergence. Keep the actual dataset release, valence configuration, cutoff, and any semicore choice in the method record. Converge occupied orbitals tightly, for example starting with EDIFF=1E-8, and save WAVECAR and CHGCAR. Use a small occupation width consistent with an insulating state.

Retain the geometry, cell, mesh, spin treatment, and PAW choices throughout the subsequent compatible stages. Each time you change a quantity that affects the basis or orbitals, regenerate dependent files rather than combining leftovers from unrelated jobs. Create stage-specific directories and a manifest listing which run produced every input binary.

Stage B: empty states and optical derivatives

Restart from the converged ground state and obtain many empty orbitals. A conventional semilocal preparation branch is:

The single electronic step is appropriate only because the input ground-state Hamiltonian is already converged. It is not a way to obtain a ground state from scratch. Check that the new WAVECAR has the intended number of bands. For this insulating branch, LOPTICS supplies WAVEDER, containing orbital derivatives used in the long-wavelength response. Verify its creation and compatibility rather than making an empty file with that name. The metal-specific workflow differs and is outside this template. Practical GW preparation, WAVEDER

Stage C: a clearly named GW approximation

For modern VASP releases, the following illustrates one-shot G0W0 through one eigenvalue-update step with fixed W. Replace the placeholders with converged values and use a clean stage INCAR rather than accumulating every old flag:

The important version distinction is that NELMGW is available from VASP 6.3.0; the corresponding older staged guide uses NELM for 6.2 and earlier, and VASP 5 naming uses GW0 for this algorithm family. Do not mix old and new recipes without checking the installed release. More iterations change the approximation, not just the accuracy tolerance of a fixed G0W0 result. NELMGW, ALGO variants

Read the quasiparticle energy/shift tables in OUTCAR, including which bands were corrected; a DFT EIGENVAL plot is not proof that GW energies were plotted. Save the stage-C WAVECAR and screened-interaction files needed downstream. Record the starting functional and distinguish G0W0, eigenvalue-self-consistent variants, and orbital-updating schemes in the method name. They can give different answers for legitimate methodological reasons.

Stage D: BSE with compatible screening

The BSE step requires compatible orbitals, transition derivatives, and screened interaction data from the preceding workflow. Preserve WFULLxxxx.tmp when available and required; the diagonal Wxxxx.tmp alternative is a different approximation. For atoms, molecules, and surfaces the full kernel is essential in the documented route, so a silent fallback is not harmless. Low-scaling GW has its own screening-dump requirements. Do not remove all .tmp files as generic “scratch cleanup” before BSE. Screened-interaction files

NBANDSO and NBANDSV define the smaller occupied/virtual electron–hole space, while NBANDS describes the compatible underlying orbital set. Their roles are different. OMEGAMAX can truncate the transition space here, not merely the x-axis of a finished plot. Inspect the selected bands in OUTCAR. BSE spectra and excitonic energies are available in vasprun.xml for the documented route; solver choice affects which eigenvector information is available. NBANDSO, NBANDSV, BSE workflow

Convergence must be layered, not guessed

For GW, test the orbital basis, response basis ENCUTGW, empty states, frequency sampling NOMEGA, and k mesh. These controls are coupled. A stable gap difference can hide unstable absolute quasiparticle energies because errors cancel. For demanding benchmarks, increase the orbital cutoff, response cutoff, and accessible empty-state space consistently rather than declaring convergence after changing only one parameter. Frequency-grid convergence is separate from plotting-grid smoothness. ENCUTGW, NOMEGA

For BSE, converge k sampling and electron–hole basis, as well as screening. A weakly bound exciton can require especially dense momentum sampling. Compare at least an independent-particle spectrum on the same quasiparticle energies with the interacting spectrum to isolate the electron–hole effect. A model-screening BSE or scissors-shifted shortcut must be explicitly named and validated, not presented as a full GW+BSE calculation. If retaining the Tamm–Dancoff approximation, state it; testing beyond it requires a documented solver branch, not just silently reusing the same result. ANTIRES

What can still go wrong after a successful run

Memory and disk demands scale rapidly with bands, response vectors, and transitions. Pilot a small compatible problem before allocation-scale production. Confirm algorithm support for the actual executable, hardware, version, spin treatment, and functional. A completed job with an unintended fallback can be less informative than an obvious failure. For 2D systems, image screening, vacuum convergence, Coulomb treatment, and the conversion from a supercell dielectric function to a sheet observable require specialist care; the bulk template cannot resolve these by adding vacuum alone.

Quasiparticle approximations can have starting-point dependence; BSE commonly uses approximate static screening and a restricted transition basis. Lifetime, phonon, temperature, and strong-correlation effects may remain outside the model. Do not promise that an expensive spectrum must coincide with experiment. Exercise: build a stage manifest, run a small validated insulator, and produce separate convergence plots for a quasiparticle gap, a selected absolute quasiparticle energy, and a BSE feature. Label unfinished convergence as unfinished, and give no numerical “expected result” unless it comes from a cited, method-matched benchmark.

7.3.3 Unexecuted inputs and analysis scaffolds

These are unexecuted teaching inputs and starting settings to test. Original figures are schematics, not computed results. Use licensed VASP and PAW data, replace every placeholder, record the executable version and validate convergence.

7.3.3.1 Input block 1

# Stage B only: exact diagonalization after an already converged SCF run
ALGO    = Exact
NELM    = 1
NSW     = 0
ISMEAR  = 0
SIGMA   = 0.01
LOPTICS = .TRUE.
LWAVE   = .TRUE.
# NBANDS = <large integer selected for the GW convergence study>

7.3.3.2 Input block 2

# Stage C, VASP >= 6.3: explicit staged calculation
ALGO   = EVGW0
NELMGW = 1
LWAVE  = .TRUE.
ISMEAR = 0
SIGMA  = 0.01
# NBANDS  = <same prepared large band count>
# ENCUTGW = <tested response cutoff in eV>
# NOMEGA  = <tested frequency-grid count>

7.3.3.3 Input block 3

# Stage D: conventional BSE skeleton, not a universal complete input
ALGO   = BSE
ISMEAR = 0
SIGMA  = 0.01
# NBANDS   = <same compatible band count as the GW files>
# NBANDSO  = <number of occupied bands included in the BSE basis>
# NBANDSV  = <number of virtual bands included in the BSE basis>
# OMEGAMAX = <tested electron-hole transition-energy cutoff in eV>
# CSHIFT   = <stated optical broadening in eV>