7. Vibronic Spectra

When an electron is removed from a molecule (photoemission, UPS), added to it (inverse photoemission, IPES) or excited (absorption), the nuclei respond to the new charge distribution. A spectrum therefore does not consist of sharp lines at the quasiparticle (QP) or excitation energies but of vibrational progressions, whose shape and width are set by the coupling of the electronic level to the molecular vibrations. VOTCA-XTP computes these couplings from analytic GW and GW-BSE gradients and a ground-state Hessian, and turns them into vibronically resolved spectra of single molecules and of morphologies.

The tools are in the Python package pyxtp.vibronic and in the script xtp_vibronic; both are installed with XTP ASE Interface. The tutorial Vibronic spectra from GW gradients goes through all steps for formaldehyde.

7.1. Overview

A vibronic spectrum needs three ingredients, all at the minimum of the electronic ground state and with the same DFT settings:

  • the normal modes (from the ground-state Hessian),

  • the gradients of the energies whose spectrum is wanted (QP levels for UPS and IPES, BSE excitation energies for absorption), and

  • the vertical energies at which the bands are placed: the QP or excitation energies of the molecule, or the site energies of a morphology (e.g. from QM/MM).

The work is split into steps that each leave a file behind. Only the first three are expensive, and none of them has to be repeated to change temperature, broadening, levels or plots:

Step

Cost

Result

Repeat when

  1. ground-state geometry

one DFT optimization

minimum.xyz

the method changes

  1. Hessian

\(6N+1\) DFT force runs

one directory per displacement, modes.npz

never (interrupted runs resume)

  1. GW(-BSE) with gradients

one GW(-BSE) run + gradients

orb file

never; further levels or states are added to it

  1. spectrum

seconds

text file, one column per level

parameters, levels or site energies change

  1. figure

seconds

PDF/SVG/PNG

broadening, axis range, experiment overlay change

On the command line:

xtp_vibronic hessian -x minimum.xyz -w hessian -t 4 -o options.xml
xtp_vibronic modes -w hessian -m modes.npz
xtp_tools -e dftgwbse -o dftgwbse.xml -n 8   # with qp_gradients, see below
xtp_vibronic spectrum -m modes.npz -g molecule.orb -p ionization -o ups.dat
xtp_vibronic plot -s ups.dat --broadening 0.1 --onset -o ups.pdf

xtp_vibronic <command> --help lists all options of a command.

7.2. Theory

7.2.1. Couplings

The model is that of displaced harmonic oscillators: the initial state (the neutral ground state) has the harmonic potential of its Hessian, and the final state (cation, anion or excited state) has the same frequencies and normal modes, displaced by the linear term of the energy gap \(\Delta E = E_\mathrm{final} - E_\mathrm{initial}\). With the gradient \(\mathbf g = \nabla \Delta E\) at the ground-state minimum and the mass-weighted normal modes \(\mathbf e_\nu\) with frequencies \(\omega_\nu\), the projections

\[q_\nu = \mathbf e_\nu \cdot \mathbf M^{-1/2} \mathbf g\]

give the reorganization energy and the Huang-Rhys factor of every mode,

\[\lambda_\nu = \frac{q_\nu^2}{2\omega_\nu^2}, \qquad S_\nu = \frac{\lambda_\nu}{\hbar\omega_\nu}.\]

\(\lambda = \sum_\nu \lambda_\nu\) is the relaxation energy of the final state from the vertical point: the vertical and the adiabatic (0-0) energy differ by \(\lambda\). For charge transfer between two molecules the reorganization energy is the sum of the \(\lambda\) of charging and discharging. The sign of the gradient does not matter.

For a QP level \(p\), \(\Delta E\) is the QP energy \(E_p\) itself: removing an electron from level \(p\) costs \(-E_p\), adding one to an empty level gains it. For a BSE state it is the excitation energy \(\Omega_n\). Both gradients are analytic and include the response of the QP corrections (see Excited-State Gradients and Geometry Optimization).

Translations and rotations are projected out of the Hessian exactly (Eckart conditions at the reference geometry): there are always \(3N-6\) (\(3N-5\) for linear molecules) vibrations, and no frequency threshold is needed to remove the rigid-body motions.

7.2.2. Line shape

The thermal line shape of the transition, with the initial state in thermal equilibrium at temperature \(T\), is the Fourier transform of the generating function (Lax, Kubo),

\[F(E) = \frac{1}{2\pi}\int d\tau\, e^{i(E-E_{00})\tau} \exp\Big(\sum_\nu S_\nu\big[(n_\nu+1)e^{-i\omega_\nu\tau} + n_\nu e^{i\omega_\nu\tau} - (2n_\nu+1)\big]\Big) \, e^{-i\lambda_\mathrm{cl}\tau - \lambda_\mathrm{cl}k_BT\tau^2} \, e^{-\sigma^2\tau^2/2 - \gamma|\tau|},\]

with \(\tau = t/\hbar\), \(\hbar = 1\) in the exponents, and the Bose occupations \(n_\nu\). At \(T=0\) it is a Poisson progression of lines \(E_{00} + k\hbar\omega_\nu\) with weights \(e^{-S}S^k/k!\); at high temperature it becomes the Marcus Gaussian. The further terms are

  • \(\lambda_\mathrm{cl}\): a classical reorganization energy, for modes that are not in the Hessian (the environment, outer sphere) or that are better treated classically, such as torsions that are hindered in a solid (option classical_below). It adds a Gaussian of variance \(2\lambda_\mathrm{cl}k_BT\).

  • \(\sigma\) and \(\gamma\): Gaussian (standard deviation) and Lorentzian (half width) broadening of every vibronic line, e.g. for the lifetime of the final state.

\(F\) is normalized, its mean is the vertical energy \(E_\mathrm{vert} = E_{00} + \lambda\), and it is evaluated with one fast Fourier transform. A warning is printed if the time integral has not decayed on the grid (no broadening and no classical part).

7.2.3. Process and energy axis

Which side of \(E_{00}\) the progression lies on depends on the measurement and on the energy axis it is plotted on. pyxtp.vibronic names the physical process instead:

process

energy axis (and site energies)

progression

absorption

photon energy, \(S_0\to S_n\) (\(\Omega_n\))

above

emission

photon energy, \(S_n\to S_0\) (\(\Omega_n\))

below

electron-addition

QP energy of the added electron (LUMO, LUMO+1, …; inverse photoemission)

above

electron-removal

QP energy of the removed electron (HOMO, HOMO-1, …; photoemission)

below

ionization

ionization energy, \(-E_\mathrm{QP}\) of an occupied level (photoemission, binding energy)

above

electron-affinity

electron affinity, \(-E_\mathrm{QP}\) of an empty level (inverse photoemission)

below

For emission the modes should strictly be those of the excited state.

7.2.4. Ensembles and spectra of several levels

A spectrum of a film is the sum of the line shapes of all molecules, each placed at its own vertical energy \(E_i\) (e.g. QP or excitation energies from QM/MM):

\[I(E) = \sum_i w_i F(E - E_i),\]

with optional weights \(w_i\), e.g. \(e^{-z_i/\ell}\) for the surface sensitivity of photoemission (depth \(z_i\), inelastic mean free path \(\ell\)). The line shape \(F\) is that of one molecule; the disorder of the environment enters through the distribution of the \(E_i\), the polarization of the environment through their mean.

For UPS and IPES the bands of all levels are added, each with its own couplings. Often site energies are known only for the frontier levels; the other levels then use the distribution of the nearest level that has them, shifted rigidly by the difference of the molecular QP energies. All levels count equally: photoemission cross sections are not included.

The onset of a spectrum is determined as in experiment: the tangent at the steepest point of the outer edge of the first band is extrapolated to zero intensity.

7.3. Step by step

7.3.1. 1. Ground-state geometry

The couplings are those at the ground-state minimum, so the geometry has to be optimized tightly with the same functional and basis set as the following steps, e.g. with analytic DFT forces through XTP ASE Interface and ASE:

from ase.io import read, write
from ase.optimize import BFGS
from pyxtp.vibronic import xtp_ground_state_factory

settings = {'dftpackage/functional': 'XC_HYB_GGA_XC_PBEH',   # PBE0
            'dftpackage/basisset': 'def2-tzvp',
            'dftpackage/auxbasisset': 'aux-def2-tzvp'}
factory = xtp_ground_state_factory(settings, nthreads=8)
atoms = read('molecule.xyz')
atoms.calc = factory('.', None)
BFGS(atoms).run(fmax=5e-4)
write('minimum.xyz', atoms)

Functionals are given by their libxc names (XC_HYB_GGA_XC_PBEH for PBE0). The residual forces of the minimum are reported in step 2.

7.3.2. 2. Hessian

FiniteDifferenceHessian differentiates analytic forces by central differences: every atom is displaced along \(x\), \(y\) and \(z\) in both directions (\(6N+1\) calculations with the reference; scheme='forward' needs \(3N+1\) but has a larger error). The default step of 0.005 Å reproduces analytic Hessians to about 1 cm-1 (see Validation).

from pyxtp.vibronic import FiniteDifferenceHessian

fd = FiniteDifferenceHessian(read('minimum.xyz'), factory,
                             workdir='hessian', step=0.005)
fd.run()                        # or fd.run(part=i, nparts=n)
modes = fd.normal_modes()
print(modes.summary())
modes.save('modes.npz')

or

xtp_vibronic hessian -x minimum.xyz -w hessian -t 8 -o options.xml
xtp_vibronic status -w hessian
xtp_vibronic modes -w hessian -m modes.npz
  • Every displacement runs in its own directory of workdir and is claimed with a lock file. A repeated run skips what is done, so an interrupted run resumes, and several processes can work on the same workdir at the same time, on one machine or on cluster nodes that share the file system. --part i --nparts n (run(part=i, nparts=n)) splits the jobs deterministically, e.g. for array jobs. The reference geometry is computed first, and the displaced SCFs start from its orbitals.

  • xtp_vibronic status --clear-locks removes the locks of jobs that crashed.

  • The workdir remembers geometry, step, scheme and atoms; a different setup is refused instead of mixing results.

  • xtp_vibronic modes reports the largest residual force at the reference geometry and the asymmetry of the finite-difference Hessian. Imaginary frequencies mean that the geometry is not a minimum; their modes are left out of the couplings with a warning.

  • indices (--atoms) displaces only some atoms (partial Hessian).

  • Any ASE calculator can be used through its own factory f(directory, displacement); an existing ORCA Hessian can be read with read_orca_hessian (xtp_vibronic spectrum --orca-hess).

7.3.3. 3. Gradients

DFT and GW-BSE are run once at the minimum, with the gradients that are needed, using dftgwbse options:

  • qp_gradients: QP levels for UPS and IPES, as a list of levels or ranges relative to HOMO and LUMO or as absolute MO indices, e.g. homo-9:lumo+9, homo,lumo or 30:40. The levels must lie in the QP window (gwbse.qpmin, gwbse.qpmax). No BSE is needed.

  • excited_state_forces: one BSE state for absorption or emission, e.g. s1 (needs dftpackage.xtpdft.compute_forces).

from pyxtp import xtp

calc = xtp(nthreads=8, excited_state='s1')
calc.set(**settings, **{'gwbse/gw/mode': 'evGW'})
calc.options.qp_gradients = 'homo-9:lumo+9'
atoms.calc = calc
atoms.get_potential_energy()        # writes the orb file

The gradients are stored in the orb file. For more levels or another excited state later, DFT and GW-BSE need not be repeated:

xtp_vibronic gradients molecule.orb --levels homo-15:lumo+15 -o options.xml
xtp_vibronic gradients molecule.orb --state s2 -o options.xml

This runs dftgwbse with tasks=gradients and must use the options of the original run (basis sets and functional are compared with the file; a different GW-BSE setup is caught by the QP and excitation energies rebuilt for the gradient). Levels whose gradients are stored already are not recomputed. From Python: pyxtp.vibronic.add_gradients.

DFT, GW and the three-index integrals are shared by all levels; every level then costs one coupled-perturbed Kohn-Sham (Z-vector) solution and one pass over the integral derivatives. The gradients are those of the perturbative QP energies; see Excited-State Gradients and Geometry Optimization for the methods covered (closed shell, RI, G0W0 or evGW with the plasmon-pole or exact self-energy, no embedding). Gradients from other sources can be used as well: read_gap_gradient reads orb files, ORCA-style .engrad files and arrays in eV/Å.

7.3.4. 4. Couplings and spectra

7.3.4.1. Single molecule

from pyxtp.vibronic import NormalModes, QPLevels, level_spectrum

modes = NormalModes.load('modes.npz')
levels = QPLevels.from_orb('molecule.orb')
for label, c in zip(levels.labels, levels.couplings(modes)):
    print(label, c.summary(3))

ups = level_spectrum(modes, levels, 'ionization', temperature=300,
                     sigma=0.02)
ups.save('ups.dat')

level_spectrum uses all occupied levels with gradients for electron-removal and ionization and all empty ones for electron-addition and electron-affinity; select restricts them, e.g. select=['homo-2', 'homo-1', 'homo']. Without site energies every band is placed at the molecular QP energy. The command line equivalent is

xtp_vibronic spectrum -m modes.npz -g molecule.orb -p ionization \
    --sigma 0.02 -o ups.dat

For one transition, e.g. the absorption to S1:

from pyxtp.vibronic import (couplings_from_gradient, read_gap_gradient,
                            vibronic_lineshape)

s1 = couplings_from_gradient(modes, read_gap_gradient('molecule.orb'))
shape = vibronic_lineshape(s1, 300, 'absorption', sigma=0.02)
# shape.energy is relative to the vertical excitation energy,
# shape.zero_zero is E_00 - E_vert

(xtp_vibronic lineshape -g molecule.orb -p absorption ...). couplings_from_geometry takes an optimized excited-state geometry instead of the gradient; the difference between the two routes measures anharmonicity and frequency changes.

7.3.4.2. Morphology

Site energies are given per level, on the energy axis of the process:

film = level_spectrum(modes, levels, 'electron-addition', temperature=300,
                      lambda_classical=0.01,
                      site_energies={'lumo': e_lumo, 'lumo+1': e_lumo1},
                      weights=np.exp(-depth / 2.0))
for label, note in film.notes.items():
    print(label, note)
xtp_vibronic spectrum -m modes.npz -g molecule.orb -p electron-addition \
    --lambda-classical 0.01 --sites sites.dat --columns lumo=3,lumo+1=4 \
    --where '1>4.0' --weight-column 7 -o ipes.dat

--columns assigns columns (0-based) of the site file to levels, --where selects rows (e.g. sites more than 4 nm below the surface, if column 1 holds the depth). Levels without site energies are shifted copies of the nearest level that has them; notes (and the header of the output file) say how every level was placed.

7.3.4.3. Parameters

temperature

K; enters the Bose occupations and the classical term

lambda_classical

classical (outer-sphere) reorganization energy, eV

sigma, gamma

Gaussian (standard deviation) and Lorentzian (half width) broadening of every vibronic line, eV

classical_below

modes below this energy are moved to the classical term (--classical-below in cm-1)

min_wavenumber

modes below this frequency (cm-1) are left out; their \(\lambda\) is reported

step

energy grid, eV

The instrument resolution is best applied when plotting (below), so that it can be changed without recomputing the spectrum.

7.3.5. 5. Figures

pyxtp.vibronic.plot (needs matplotlib) makes figures for publications: vector output, journal column widths (SINGLE_COLUMN 3.375 in, DOUBLE_COLUMN 7 in), and colours that are distinguishable with colour-vision deficiencies. Every level keeps its colour in all figures (HOMO and LUMO first, then by the distance from the gap).

from pyxtp.vibronic import Spectrum
from pyxtp.vibronic.plot import paper_style, plot_spectrum, plot_couplings

with paper_style():
    fig = plot_spectrum(Spectrum.load('ipes.dat'), broadening=0.3,
                        onset=True, experiment=(e_exp, i_exp))
    fig.savefig('ipes.pdf')
    print(fig.onset)                  # onset and first peak, eV
    plot_couplings(levels.couplings(modes)[0], title='LUMO').savefig('lumo.pdf')
xtp_vibronic plot -s ipes.dat --broadening 0.3 --onset \
    --experiment measured.dat -o ipes.pdf
xtp_vibronic lineshape -m modes.npz -g molecule.orb -p absorption \
    --couplings s1_modes.txt -o s1.dat
xtp_vibronic plot -c s1_modes.txt --title S1 -o s1_modes.pdf

plot_spectrum draws the contributions of the levels, the intrinsic total, the total with the instrument broadening (Gaussian, standard deviation) and an optional measured spectrum, scaled to the broadened total. With onset the tangent construction and the first peak are marked. plot_couplings shows \(\lambda_\nu\) (or \(S_\nu\)) per mode and the running sum.

7.4. Validation

  • Hessian: water, PBE0/def2-SVP, finite differences of xtp forces 1642/3866/3966 cm-1, analytic Hessian of PySCF 1643.5/3865.2/3964.9 cm-1.

  • QP gradients against finite differences of the QP energies with GW recomputed at displaced geometries (water, HOMO-2 to LUMO+2): relative errors below \(10^{-6}\) for G0W0 and evGW with the plasmon-pole and the exact self-energy.

  • Line shapes: Poisson progression at \(T=0\), Marcus limit at high temperature, normalization, mean and variance, detailed balance (unit tests in test_vibronic.py).

7.5. Limitations and good practice

  • Linear coupling only. The final state has the frequencies and modes of the initial state: no frequency changes, no Duschinsky rotation, and modes that are not displaced by symmetry do not appear even if their frequency changes (e.g. the out-of-plane mode of formaldehyde in S1). Herzberg-Teller effects are not included.

  • Consistency. Geometry, Hessian and gradients must come from the same functional and basis set, and the geometry must be a tight minimum: the gradient of the gap is only that of the final state if the ground-state gradient vanishes.

  • Soft modes. Low-frequency torsions have large Huang-Rhys factors but small \(\lambda\); in a solid they are hindered. classical_below (e.g. 100 cm-1) moves them into the classical term, min_wavenumber leaves them out.

  • Intensities. All levels and sites count equally; cross sections of photoemission and the oscillator strengths of other transitions are not included.

  • QP levels. The gradients are those of the perturbative QP energies (diagonal QP Hamiltonian). Levels with two QP solutions of similar weight (e.g. high virtual levels in evGW) should be kept out of the QP window. For degenerate levels the gradient of their mean energy is used.