7. Embedded GW-BSE: Screening by a Polarizable Environment

An electronic excitation of a molecule in a solid is screened by its surroundings. When the excitation rearranges charge, the neighbouring molecules polarize in response, and that response acts back on the excitation. For charged excitations (electron or hole) the effect is of the order of an electronvolt; for neutral excitations it is smaller, but large enough to reorder states or to shift them by more than the accuracy one expects of GW-BSE.

VOTCA-XTP offers two ways to include this response in a QM/MM job with polarizable regions:

  • Iteratively, by treating the excited state like a ground state: the QM/MM loop converges the environment’s induced dipoles self-consistently against the density of the excited state, with a GW-BSE calculation in every outer iteration. This is the default for a job whose QM region is in an excited state.

  • Embedded, by folding the environment’s response into the screened Coulomb interaction of GW-BSE itself, following [Li:2016] and [Li:2018]. The QM/MM loop converges only the ground state; one GW-BSE calculation then runs with the environment as an additional, static screening medium. This is switched on with environment_screening in the qmregion options.

The embedded scheme is cheaper, since it needs one GW-BSE calculation instead of one per outer iteration. It also contains two contributions that the iterative scheme cannot represent (see Relation to the iterative scheme). The radial_dielectric correction of the legacy code is available within it as shell regions.

7.1. The reaction field

The environment is the set of polarizable sites of the job’s polar regions: Thole-damped inducible point dipoles [Thole:1981]. A change \(\delta\rho\) of the QM charge density produces a field \(\mathbf{E} = F\,\delta\rho\) at the sites, the sites respond with induced dipoles \(\boldsymbol{\mu} = A^{-1}\mathbf{E}\), where

\[A = \alpha^{-1} + T_\text{Thole}\]

is the same operator the polar region solves in the QM/MM loop, and the induced dipoles act back on the QM subsystem. The resulting additional interaction between two QM charge distributions is the reaction field

(7.1)\[v_\text{reac} = -F A^{-1} F^{T},\]

and the QM electrons interact through

\[u = v + v_\text{reac}\]

instead of the bare Coulomb interaction \(v\). \(v_\text{reac}\) is negative semidefinite: the environment screens.

In the resolution-of-identity representation used by GW-BSE, a product density \(\rho_{mn}\) is expanded in auxiliary functions, \(F\) becomes the field at the sites of each auxiliary function taken as a charge density, and Eq.7.1 turns into a matrix \(B = -F A^{-1} F^{T}\) over the auxiliary basis. In the metric of the stored three-centre integrals \(M_{mn}\) (in which the bare \(v\) is the identity) it becomes

\[R = T^{T} B\, T, \qquad (mn|v_\text{reac}|kl) = M_{mn}\, R\, M_{kl}^{T},\]

with \(T\) the inverse square root of the auxiliary Coulomb metric that is folded into \(M\). \(R\) is built once per job: it depends only on the QM atoms, their auxiliary basis and the polar sites, not on the orbitals.

\(A\) is assembled and factorized densely, which is far cheaper than iterating a linear solve for every auxiliary function, and \(B\) is formed from its Cholesky factor, so it is symmetric and negative semidefinite by construction. The field \(F\) is obtained from nuclear-attraction (or, for smeared sites, two-centre Coulomb) integrals of each auxiliary function by a four-point central difference.

7.2. Quasiparticle energies (GW)

With the environment, the screened interaction is

\[W = \left[u^{-1} - \chi_0\right]^{-1},\]

which is split as \(W = v + v_\text{reac} + (W - u)\):

  • \(v\) gives the exchange self-energy \(\Sigma^\text{x}\), as without an environment.

  • \(v_\text{reac}\) gives a static COH+SEX self-energy,

    \[\begin{split}\Sigma^\text{reac}_{mn} = \tfrac{1}{2} \sum_{l} s_l\, (ml|v_\text{reac}|ln), \qquad s_l = \begin{cases} -1 & l \text{ occupied} \\ +1 & l \text{ virtual,} \end{cases}\end{split}\]

    that is, a Coulomb hole \(\tfrac{1}{2}\sum_l (ml|v_\text{reac}|ln)\) plus a screened exchange \(-\sum_{l\in\text{occ}} (ml|v_\text{reac}|ln)\). It is part of the correlation (it belongs to \(W - v\)) and enters wherever \(\Sigma^\text{c}\) does: the quasiparticle equation, evGW and QSGW. It is printed as S-R next to S-X and S-C, and the log lists its diagonal together with the resulting HOMO, LUMO and gap shifts.

  • \(W - u\) is the correlation of the QM electrons, which now interact through \(u\) and are screened by the environment as well. With \(S = (1 + R)^{1/2}\),

    \[W = S \left[1 - S\chi_0 S\right]^{-1} S ,\]

    so the unchanged RPA and correlation machinery runs on the dressed integrals \(M S\).

For an electron added to or removed from the molecule, \(\Sigma^\text{reac}\) contains the polarization energy of the environment by that charge. The HOMO and LUMO quasiparticle energies therefore include it directly, which is how the embedded scheme provides hole and electron site energies (see Usage).

7.3. Neutral excitations (BSE)

The BSE kernel becomes

\[K = K^\text{x}[v] + K^\text{reac}[v_\text{reac}] + K^\text{d}[W_\text{tot}], \qquad W_\text{tot} = \left[(1 + R)^{-1} + \varepsilon - 1\right]^{-1},\]

with \(\varepsilon\) the RPA dielectric matrix of the QM subsystem in the same metric.

  • \(K^\text{d}[W_\text{tot}]\) is the direct (electron-hole) term with the environment-screened interaction. It describes the environment’s response to the charge that the excitation rearranges: a state-specific effect, in the language of solvation models [Duchemin:2018].

  • \(K^\text{reac} = (ia|v_\text{reac}|jb)\) has the index structure of the exchange term. It is the environment’s response to the transition density, i.e. a linear-response effect, and it mainly moves bright singlets. Triplets have no exchange term and therefore no \(K^\text{reac}\).

With include_kreac (the default), the whole kernel runs on dressed integrals, where \(K^\text{x} + K^\text{reac}\) and \(K^\text{d}[W_\text{tot}]\) come out of the unchanged BSE operators. Without it, the integrals stay bare and \(W_\text{tot}\) is diagonalized in place of \(\varepsilon\), so that \(K^\text{x}\) keeps the bare \(v\). In both cases the first-order shift \(\langle K^\text{reac}\rangle = 2\,(X+Y)^{T} K^\text{reac} (X+Y)\) of each singlet is logged after the BSE is solved, as an estimate of the linear-response contribution. With include_kreac the exchange column of the state analysis is labelled <K_x+K_reac>.

7.4. Site widths and stability

For a charge density and inducible points outside it, the reaction energy \(\langle\rho|v_\text{reac}|\rho\rangle\) is bounded by the dielectric limit, above \(-\langle\rho|v|\rho\rangle\). In the metric of \(M\) this means that all eigenvalues of \(R\) lie above \(-1\), i.e. that \(1 + R\) is positive definite. A point dipole inside the tail of a diffuse auxiliary function is not bounded in this way, and in dense molecular solids the closest sites of neighbouring molecules do sit in those tails. An eigenvalue of \(R\) at or below \(-1\) would mean that the environment screens a charge fluctuation by more than the fluctuation itself; the dressing \(S = (1+R)^{1/2}\) then does not exist.

Each site therefore responds to the QM field averaged over a normalized Gaussian of width

\[R_j = w\,\alpha_{\text{iso},j}^{1/3}, \qquad \alpha_{\text{iso},j} = \tfrac{1}{3}\operatorname{tr}\alpha_j ,\]

with \(w\) the option site_width – the length scale that Thole damping itself uses. Outside the auxiliary functions a spherical charge acts as a point, so this changes the response only where point sites over-respond. In a production QM/MM job (55 QM atoms, def2-TZVP with aux-def2-TZVP, 3900 Thole sites) the lowest eigenvalue of \(R\) was \(-1.056\) with point sites, carried by three carbon atoms of a neighbouring C60 at 3.05–3.3 Å; with the default site_width of 0.5 it was \(-0.755\), while the reaction energies of the HOMO, LUMO and HOMO-LUMO densities changed by \(10^{-4}\) relative.

Because \(R\) does not depend on the orbitals, a job checks \(1 + R\) before the QM/MM loop starts. If it is not positive definite, the job fails right away with a report of the lowest modes: where their charge sits on the QM side and which polar sites carry their reaction, with distances.

7.5. Shell regions

Polar regions listed in shell_regions respond without mutual induction, each site with the polarizability \(\alpha_j/\varepsilon_\text{shell}\):

\[B_\text{shell} = -\sum_{j} F_j\, \frac{\alpha_j}{\varepsilon_\text{shell}}\, F_j^{T} .\]

This needs no linear solve, so it is a cheap treatment of an outer shell beyond the explicit Thole region. The factor \(1/\varepsilon_\text{shell}\) (shell_dielectric) stands in for the mutual induction that is left out. It is the radial_dielectric correction of the legacy code in this framework. Because the sum runs over actual sites, the geometry is that of the morphology (a slab simply has no sites in the vacuum), which is why this rather than an analytic continuum term is used for the tail. The shell and the explicit regions do not couple to each other.

The explicit (non-shell) polar regions respond together, as one coupled Thole system, and must therefore use the same exp_damp.

An ewaldregion contributes the electrostatic potential of the periodic background to the ground state, as in any QM/MM job (see Polarizable Ewald Embedding). It is frozen and does not respond to the excitation, so it is not part of \(R\).

7.6. Relation to the iterative scheme

With a linear environment response, the total energy of the QM/MM system in a state \(X\) with QM density \(\rho_X\) is, at self-consistency,

\[E_X = E_\text{QM}[\rho_X] + \langle\rho_X|\phi_\text{p}\rangle + \tfrac{1}{2}\langle\rho_X|v_\text{reac}|\rho_X\rangle ,\]

with \(\phi_\text{p}\) the potential of the permanent multipoles. The excitation energy of the iterative scheme is then

\[\Delta E = E_X - E_g = \Omega_g + \tfrac{1}{2}\langle\Delta\rho|v_\text{reac}|\Delta\rho\rangle ,\]

where \(\Omega_g\) is the excitation energy in the environment polarized by the ground state and \(\Delta\rho = \rho_X - \rho_g\) is the density change of the excitation. The iterative scheme thus sees the environment’s response to the mean density change of the excitation.

The embedded scheme starts from the same ground state, so its \(\Omega_g\) is the same, and through \(K^\text{d}[W_\text{tot}]\) it contains the mean-field response as well. In addition, it contains

  • the response to the correlated motion of electron and hole: the environment responds to where the electron and the hole are relative to each other, not only to their averaged densities; and

  • \(K^\text{reac}\), the response to the transition density.

Neither can be represented by a density that is iterated against the environment. Conversely, both schemes treat the environment’s response as static (instantaneous); see [Amblard:2024] for when a dynamically polarizable environment matters.

The same distinction applies to charged excitations. In a charged QM/MM job (state e or h) the environment polarizes against the averaged density of the extra electron or hole. The static self-energy \(\Sigma^\text{reac}\) responds to the carrier as an instantaneous point charge in each orbital it occupies. The two can differ noticeably for orbitals spread over several separated parts of a molecule.

7.7. Usage

The embedded scheme is switched on per QM region:

<qmregion>
  <id>0</id>
  <state>jobfile</state>
  <segments>jobfile</segments>
  <dftpackage> ... </dftpackage>
  <gwbse> ... </gwbse>
  <statetracker> ... </statetracker>
  <environment_screening>
    <include_kreac>true</include_kreac>
    <shell_regions>2</shell_regions>
    <shell_dielectric>4.0</shell_dielectric>
    <site_width>0.5</site_width>
  </environment_screening>
</qmregion>

option

default

meaning

include_kreac

true

Include \(K^\text{reac}\) in the BSE kernel. Without it, \(K^\text{reac}\) is only estimated to first order and logged.

shell_regions

(none)

Ids of polar regions that respond as an uncoupled shell, with polarizabilities \(\alpha/\varepsilon_\text{shell}\).

shell_dielectric

4.0

\(\varepsilon_\text{shell}\) for the shell regions.

site_width

0.5

Width of the polar sites as seen by the QM density, in units of \(\alpha_\text{iso}^{1/3}\); 0 gives point sites.

The state of the QM region must be a GW-BSE state:

  • an exciton (s1, t2, …), tracked with the statetracker as usual, or

  • a quasiparticle level (pqpN or dqpN, with N the HOMO or LUMO index), for hole and electron site energies.

Charged DFT states (e, h) are refused. Their extra charge would be polarized in the QM/MM loop and screened a second time in GW. Unrestricted excitons and truncated (DFT-in-DFT) active regions are not supported.

A job proceeds as follows:

  1. The screening environment is assembled from the job’s polar regions, and \(1 + R\) is checked.

  2. The QM/MM loop converges the ground state: DFT and the polar regions, as in an n job.

  3. One GW-BSE calculation runs on top of it, screened by the environment. The tracked state’s energy replaces the QM region’s energy.

  4. The job writes checkpoint_screened.hdf5 and reports, in its output, the ground-state total energy E_ground and the site energies screened_site_energies, all relative to its own ground state, in eV:

    • h \(= -\varepsilon_\text{HOMO}\) and e \(= \varepsilon_\text{LUMO}\), the quasiparticle energies (diagonalized ones if available),

    • s or t, the excitation energy of the tracked exciton, if the job’s state is one.

A screened job therefore needs no separate ground-state job. When writing jobs (-j write) no n job is added for the segments, and -j read takes the site energies of screened jobs directly. Several screened jobs of the same segment (an s1 and a quasiparticle job, say) report the same e and h; -j read checks that they agree to 1 meV.

The auxiliary basis of the screening is the one GW-BSE uses: the DFT auxiliary basis if one is set, otherwise gwbse.auxbasisset, otherwise aux-<basisset>.

7.8. Example

For a dicyanovinyl-oligothiophene donor in a morphology with C60 (55 QM atoms, def2-TZVP/aux-def2-TZVP, evGW, full BSE; polar region of 2.4 nm with 68 segments and 3900 Thole sites, exp_damp 0.39; optionally a periodic Ewald background of 1212 segments; site_width 0.5, no shell regions), the lowest singlet excitation energy (eV) is

scheme

cut-off

with Ewald background

QM only (vacuum)

2.378

iterative QM/MM, \(E_{s_1} - E_n\)

2.319

2.286

embedded, without \(K^\text{reac}\)

2.218

2.186

embedded, with \(K^\text{reac}\)

2.102

2.068

The Ewald background shifts all three schemes by the same \(-0.033\) eV: it acts through the ground state only. Relative to the QM-only value (cut-off environment), the ground-state embedding contributes \(+0.027\) eV and the mean-field response to the density change \(-0.087\) eV in the iterative and \(-0.083\) eV in the embedded scheme. These two are common to both schemes. The embedded scheme adds \(-0.104\) eV from the correlated electron-hole pair and \(-0.116\) eV from \(K^\text{reac}\), which the logged first-order estimate reproduces to about 10%.