This is an old revision of the document!
Table of Contents
How to run XAS LR-TDDFT calculations
This a a short tutorial on how to run near-edge X-ray absorption spectroscopy calculations using linear-response TDDFT. The method is implemented in CP2K under the XAS_TDP name. It relies on core-level specific approximations that enables efficient calculations of large and periodic systems. Both K- and L-edge are available. The details of the method can be found in Physical Chemistry Chemical Physics, 2021, DOI: 10.1039/D0CP06164F. Please cite this paper if you were to use the XAS_TDP method for work you publish. Note that XAS LR-TDDFT comes with a correction scheme, described further down this tutorial.
Brief theory recap
The method is based on 3 main core-specific approximations that boost the calculation efficiency. The first one is the core-valence separation. Due to large differences in energy and localization, core and valence states only weakly couple. Thus, when dealing with XAS, it is customary to simply ignore excitations from valence states.
The second approximation is the sudden approximation, in which the relaxation of electrons beyond the core region is neglected upon excitation of a core electron. Combined with the localized nature of core states, this allows to treat excitations one at a time rather than all at once. This is more efficient in the sense that diagonaling a series of small matrices scales better than diagonalizing a single much larger one.
Finally, a lot of 4-center 2-electron integrals have to be computed. Thanks to the core-valence separation and the sudden approximation, all required integrals involve the core state from which the exciation takes place. This allows for a core-specific resolution of the identity scheme (RI). For the Coulomb integrals:
\begin{equation} (pI|Jq) \approx \sum_{\mu, \nu} \ (pI|\mu) \ (\mu|\nu)^{-1} \ (\nu|Jq) \end{equation}
where $p,q$ represent atomic orbitals (Gaussian type orbitals/GTOs) and $I, J$ core orbtials. For non-zero integrals, $p,I$ and $q,J$ have to overlap. Since $I,J$ are localized core orbitals centered on the same atom, it is sufficient to take a RI basis only made GTOs centered on the excited atom. Note that in case of K-edge spectroscopy, there is only one core state to consider and $I=J$. For L-edge, $I,J$ span all three degenerate $2p$ states. This leads to particularly efficient integral evaluations.
For the exchange-correlation kernel, the RI scheme reads:
\begin{equation} (pI|f_{xc}|Jq) \approx \sum_{\kappa, \lambda, \mu, \nu, } \ (pI|\kappa) \ (\kappa|\lambda)^{-1} \ (\lambda|f_{xc}|\mu) \ (\mu|\nu)^{-1} (\nu|Jq) \end{equation}
where all integrals but $(\lambda|f_{xc}|\mu)$ are the same as for the Coulomb kernel above. Since the RI basis elements $\lambda, \mu$ are centered on the excited atoms, we only need the density in its vicinity. For this, we use a simple projection:
\begin{equation} \label{proj} \begin{aligned} n(\mathbf{r}) &=\sum_\sigma\sum_{pq} P^\sigma_{pq} \ \varphi_p(\mathbf{r}) \varphi_q(\mathbf{r})\\ %\pause &\approx \sum_\sigma \sum_{pq}\sum_{\mu\nu} P^\sigma_{pq} \ (pq\mu) \ S_{\mu\nu}^{-1} \ \chi_\nu(\mathbf{r})\\ &= \sum_\nu d_\nu \ \chi_\nu(\mathbf{r}), \end{aligned} \end{equation}
which turns the density into a linear combination of RI basis elements. This allows for easy and simple numerical integration of $(\lambda|f_{xc}|\mu)$. Note that the quality of the projection may suffer if there are (heavy) atoms close by since their core states may not be well described (GTOs are only sharp at their center). This can be addressed by either using pseudopotentials for the neighbors or adding their RI basis function for the projection.
For these approximations to work, the core states to be excited need to be identified among the Kohn-Sham oritals. They need to have a strong $1s$, $2s$ or $2p$ nature and be well localized. Not fullfilling these conditions will lead to wrong results.
The XAS_TDP input section
The parameters defining XAS LR-TDDFT calculations are found in the XAS_TDP subsection of DFT. Some external parameters also need to be set to specific values. In particular, RUN_TYPE should be set to ENERGY and DFT%QS%METHOD to GAPW. The combination of GAPW and all-electron basis sets allow for an accurate description of core states.
The most important keywords and subsections of XAS_TDP are:
DONOR_STATES: which define which core states need to be excited (and where to look for them)KERNEL: where the XC functional and exact exchange interaction (for hybrid TDDFT) are definedGRID: which defines the integration grids for the xc kernel $(\lambda|f_{xc}|\mu)$
The defaults value of all other keywords are in principle good enough.
Note that the first requirement for XAS LR-TDDFT is that the ground state calculation on which it is based is of good quality.
Simple examples
Illustrative examples usually tell more than long texts. Below, some typical input examples are displayed with explanations. They should cover most common use cases.
CO$_2$ molecule (K-edge)
This is a simple C and O K-edge calculation of the CO$_2$ molecule in the gas phase. The annotated input file is displayed below:
- CO2.inp
&GLOBAL PROJECT CO2 RUN_TYPE ENERGY &END GLOBAL &FORCE_EVAL &DFT BASIS_SET_FILE_NAME EMSL_BASIS_SETS POTENTIAL_FILE_NAME POTENTIAL AUTO_BASIS RI_XAS MEDIUM ! size of automatically generated RI basis &MGRID CUTOFF 500 REL_CUTOFF 40 NGRIDS 5 &END MGRID &QS METHOD GAPW ! It is necesary to use the GAPW method for &END QS ! accurate description of core states &POISSON PERIODIC NONE PSOLVER MT &END &SCF EPS_SCF 1.0E-8 MAX_SCF 30 &END SCF &XC &XC_FUNCTIONAL &LIBXC FUNCTIONAL HYB_GGA_XC_BHandHLYP &END LIBXC &END XC_FUNCTIONAL &HF FRACTION 0.5 ! BHandHLYP functional requires 50% exact exchange &END HF &END XC &XAS_TDP &DONOR_STATES DEFINE_EXCITED BY_INDEX ! We look for states by atom index: ATOM_LIST 1 2 ! we want to excite atoms 1 and 2 STATE_TYPES 1s 1s ! from their 1s core state. N_SEARCH 3 ! The 3 lowest energy MOs need to be searched (C1s, O1s, O1s) LOCALIZE ! States need to be actively localized because O atoms are &END DONOR_STATES ! equivalent under symmetry GRID C 250 500 ! Integration grid dimensions for C and O excited atoms GRID O 250 500 ! there are 250 angular points (Lebedev grid) and 500 ! radial points &KERNEL RI_REGION 2.0 ! Include RI basis elements from atoms within a 2.0 Ang ! sphere radius around the excited atom for the density projection &XC_FUNCTIONAL &LIBXC FUNCTIONAL HYB_GGA_XC_BHandHLYP &END LIBXC &END XC_FUNCTIONAL &EXACT_EXCHANGE FRACTION 0.5 ! Definition of the functional for the TDDFT kernel &END EXACT_EXCHANGE ! Here (and usually) taken to be the same as the ground state &END KERNEL &END XAS_TDP &END DFT &SUBSYS &CELL ABC 10.0 10.0 10.0 PERIODIC NONE &END CELL &COORD C 5.00 5.00 5.00 O 5.00 5.00 6.16 O 5.00 5.00 3.84 &END COORD &KIND C BASIS_SET 6-311G** ! Using all-electron basis sets and potential is necessary POTENTIAL ALL ! for the correct description of core states &END KIND &KIND O BASIS_SET 6-311G** POTENTIAL ALL &END KIND &END SUBSYS &END FORCE_EVAL
There are a few points of particular interest in the XAS_TDP section of this input file. When defining the DONOR_STATES subsection, we specify that excited atoms are defined by their index and proceed to list atoms 1 and 2. This is such that excitations take place from the Carbon 1s level and from one of the Oxygen 1s levels. Since both O atoms are equivalent under symmetry, there is no need to run the calculation for both of them (this calculation is small enough and can run on a single processor, this is mostly for illustration purposes).
In the KERNEL subsection, the RI_REGION keyword is set to 2.0. This is such that all atoms within a 2.0 Angstrom radius of the current atom provide RI basis functions for the projection of the density (see last equation of theory recap). This leads to a better quality description, especially for the Carbon atom which has 2 heavier Oxygens around. Note that if the O atoms were described with the use of a pseudopotential, there would be no need for RI_REGION.
In the resulting output file, there is a XAS_TDP section reporting the steps of the calculations. Especially important are the parts related to donor core state identification. For each excited atom/core level combination, a small report about Mulliken population analysis and overlap with pure donor level is printed. For the Carbon 1s:
# Start of calculations for donor state of type 1s for atom 1 of kind C
The following localized MO(s) have been associated with the donor state(s)
based on the overlap with the components of a minimal STO basis:
Spin MO index overlap(sum)
1 3 1.00113
The next best overlap for spin 1 is 0.00000 for MO with index 1
Mulliken population analysis retricted to the associated MO(s) yields:
Spin MO index charge
1 3 1.002
Both the overlap and the Mulliken charge should be as close to 1.0 as possible. This ensure that the molecular orbital selected is of the correct type (here projection on a C 1s Slater type orbital) and properly localized (there is a full electron associated to this MO, on this atom). If those numbers are lower, something went wrong with the core level identification. This is usually solved by increasing N_SEARCH or/and by usinge the LOCALIZE keyword in case some atoms are equivalent under symmetry.
Spectral information are given in a separate file named CO2.spectrum, where excitation energies are listed with corresponding oscillator strengths, for each excited core level.
Tetrahedral NaAlO$_2$ (K-edge, periodic)
This example is about crystalline sodium aluminate and illustrates how large periodic structures can be efficiently simulated.
- sodal.inp
&GLOBAL PROJECT sodal RUN_TYPE ENERGY PRINT_LEVEL LOW &END GLOBAL &FORCE_EVAL METHOD QS &DFT BASIS_SET_FILE_NAME BASIS_ADMM ! the pcseg-n and admm-n basis set families can be downloaded at https://www.basissetexchange.org BASIS_SET_FILE_NAME BASIS_PCSEG BASIS_SET_FILE_NAME BASIS_MOLOPT POTENTIAL_FILE_NAME POTENTIAL AUTO_BASIS RI_XAS MEDIUM &QS METHOD GAPW ! GAPW is necessary for core states &END QS &AUXILIARY_DENSITY_MATRIX_METHOD ! The ADMM methog greatly accelerated the ground state calculation ADMM_PURIFICATION_METHOD NONE ! This is the simplest ADMM scheme and has proven to work well &END AUXILIARY_DENSITY_MATRIX_METHOD &SCF MAX_SCF 30 EPS_SCF 1.0E-06 &OT MINIMIZER DIIS PRECONDITIONER FULL_ALL &END OT &OUTER_SCF MAX_SCF 6 EPS_SCF 1.0E-06 &END &END SCF &MGRID CUTOFF 400 REL_CUTOFF 40 NGRIDS 5 &END &XC &XC_FUNCTIONAL PBE ! This is the PBEh functional with 45% HFX &PBE ! Large fraction of HFX are ususally needed for XAS LR-TDDFT SCALE_X 0.55 &END &END XC_FUNCTIONAL &HF FRACTION 0.45 &INTERACTION_POTENTIAL POTENTIAL_TYPE TRUNCATED ! The tuncated Coulomb potential has to be used in PBCs CUTOFF_RADIUS 5.0 ! with a cutoff radius lower than half the cell size &END INTERACTION_POTENTIAL &SCREENING EPS_SCHWARZ 1.0E-6 ! Screening HFX integrals boosts performance &END SCREENING &END HF &END XC &XAS_TDP &DONOR_STATES DEFINE_EXCITED BY_KIND ! We define the excited atoms by kind, which is named Alx here KIND_LIST Alx ! There is only one Alx atom in the coordinates since all Al STATE_TYPES 1s ! atoms are equivalent under symmetry. The Alx atom is the only N_SEARCH 1 ! one decribed at all-electron level, which is why we use &END DONOR_STATES ! N_SEARCH = 1. There is also no need to LOCALIZE TAMM_DANCOFF ! TDA is turned on by default, but we make it explicit here GRID Alx 150 300 ENERGY_RANGE 20.0 ! This means that we onluy solve for excitation energies that are ! up to 20.0 eV above the first energy &OT_SOLVER MINIMIZER DIIS ! The iterative OT solver is typically more efficient than EPS_ITER 1.0E-4 ! full diagonalization for large systems &END OT_SOLVER &KERNEL &XC_FUNCTIONAL PBE &PBE SCALE_X 0.55 &END &END XC_FUNCTIONAL &EXACT_EXCHANGE OPERATOR TRUNCATED RANGE 5.0 SCALE 0.45 &END EXACT_EXCHANGE &END KERNEL &END XAS_TDP &END DFT &SUBSYS &CELL ABC 10.467947 10.651128 14.393541 &END CELL &TOPOLOGY COORD_FILE_NAME sodal.xyz COORD_FILE_FORMAT xyz &END TOPOLOGY &KIND O BASIS_SET DZVP-MOLOPT-SR-GTH BASIS_SET AUX_FIT FIT3 POTENTIAL GTH-PBE &END KIND &KIND Na ELEMENT Na BASIS_SET DZVP-MOLOPT-SR-GTH BASIS_SET AUX_FIT FIT3 POTENTIAL GTH-PBE &END &KIND Al BASIS_SET DZVP-MOLOPT-SR-GTH BASIS_SET AUX_FIT FIT3 POTENTIAL GTH-PBE &END &KIND Alx ! All atoms but the single Alx are described using pseudopotentials ELEMENT Al ! This greatly reduces the number of basis function and the cost of BASIS_SET pcseg-2 ! the calculation in general. AUX_FIT basis sets are for ADMM BASIS_SET AUX_FIT admm-2 POTENTIAL ALL &END &END SUBSYS &END FORCE_EVAL
There are many performance oriented keywords and subsection in the above input. Most importantly, only one atom is treated at the all-electron level (the one atom from which the excitation takes place), all other are described using pseudopotentials. Also quite important is the usage of the ADMM method. This allows for very efficient evaluation of the HFX energy in the ground state calculation. Finally, the OT iterative solver is used. Since only a handful of eigenvalues are calculated (those within 20.0 eV of the first excitation energy), this scales much better than a full digonalization.
This input file would generate a spectrum such as the one visible on figure 4 of the reference work. This is a much larger calculation than the first example though and would require a few hours on 20-30 processors (mostly to converge the SCF). In you are interested in reproducing this result, input, geometry and pcseg-2/admm-2 basis sets are available here.
TiCl$_4$ molecule (L-edge + SOC)
This example covers L-edge spectroscopy with the addition of spin-orbit coupling.
- TiCl4.inp
&GLOBAL PROJECT TiCl4 PRINT_LEVEL LOW RUN_TYPE ENERGY &END GLOBAL &FORCE_EVAL &DFT BASIS_SET_FILE_NAME BASIS_DEF2-TZVPD POTENTIAL_FILE_NAME POTENTIAL AUTO_BASIS RI_XAS LARGE &POISSON PERIODIC NONE PSOLVER MT &END POISSON &QS METHOD GAPW &END QS &MGRID CUTOFF 800 REL_CUTOFF 50 NGRIDS 5 &END &SCF EPS_SCF 1.0E-8 MAX_SCF 200 &MIXING METHOD BROYDEN_MIXING ALPHA 0.2 BETA 1.5 NBROYDEN 8 &END MIXING &END SCF &XC &XC_FUNCTIONAL &LIBXC FUNCTIONAL HYB_GGA_XC_B3LYP &END LIBXC &END XC_FUNCTIONAL &HF FRACTION 0.2 &END HF &END XC &XAS_TDP &DONOR_STATES DEFINE_EXCITED BY_KIND KIND_LIST Ti STATE_TYPES 2p ! 2p core state for L-edge &END DONOR_STATES ! No need to LOCALIZE since only one Ti atom TAMM_DANCOFF FALSE ! TDA is on by default, get full TDDFT like this DIPOLE_FORM LENGTH GRID Ti 500 1000 ! This is a fairly dense grid EXCITATIONS RCS_SINGLET ! For SOC calculations in closed-shell system, these 3 keywords EXCITATIONS RCS_TRIPLET ! are required. Singlet and triplet excitation are coupled together SOC ! with the SOC hamiltonian &KERNEL RI_REGION 5.0 ! To get the best possible density projection &XC_FUNCTIONAL &LIBXC FUNCTIONAL HYB_GGA_XC_B3LYP &END LIBXC &END XC_FUNCTIONAL &EXACT_EXCHANGE FRACTION 0.2 &END EXACT_EXCHANGE &END KERNEL &END XAS_TDP &END DFT &SUBSYS &KIND Cl BASIS_SET def2-TZVPD POTENTIAL ALL RADIAL_GRID 80 ! The GAPW grids are also used to evaluate the SOC operator LEBEDEV_GRID 120 ! it is good practice to use sligthly larger ones than the default &END KIND &KIND Ti BASIS_SET def2-TZVPD POTENTIAL ALL RADIAL_GRID 80 LEBEDEV_GRID 120 &END KIND &CELL ABC 10.0 10.0 10.0 PERIODIC NONE &END CELL &TOPOLOGY COORD_FILE_FORMAT XYZ COORD_FILE_NAME TiCl4.xyz &CENTER_COORDINATES &END CENTER_COORDINATES &END TOPOLOGY &END SUBSYS &END FORCE_EVAL
The structure of the input file is not very different from the CO$_2$ example. Notable differences are the DONOR_STATES subsection where 2p states are specified and the combinations of the EXCITATIONS and SOC keywords. Indeed, the way spin-orbit coupling is treated in XAS_TDP is by coupling together singlet and triplet excitation via the ZORA SOC Hamiltionian.
Note that this calculation is meant to be a benchmark calculation, hence the overall larger GRIDs and CUTOFFs. The result can be seen in the reference work, figure 1 a). The calculation takes about 10 minutes on 4 cores. All necessary files are available here.
In the output file, the donor state identification yields overlaps that are greater than one. This is due to the degenerate nature of 2p states. The candidate Kohn-Sham orbital is projected on 3 STOs for 2px, 2py and 2pz. To avoid cancelling contributions, the sum of the absolute overlap is taken.
# Start of calculations for donor state of type 2p for atom 1 of kind Ti
The following canonical MO(s) have been associated with the donor state(s)
based on the overlap with the components of a minimal STO basis:
Spin MO index overlap(sum)
1 7 1.36751
1 8 1.36751
1 9 0.99786
The next best overlap for spin 1 is 0.06653 for MO with index 27
Mulliken population analysis retricted to the associated MO(s) yields:
Spin MO index charge
1 7 1.000
1 8 1.000
1 9 1.000
FAQ
Which functional and basis sets to use ?
Hybrid functionals with high fraction of Hartree-Fock exchange are know to perform well for core spectroscopy. PBEh($\alpha=0.45$) and BHandHLYP have had success with this particular implementation. In periodic boundary conditions, the truncated Coulomb operator should be used (with truncation radius < half cell parameter).
For appropriate description of core states, all-electron basis sets should be used for the excited atom(s). MOLOPT basis sets and pseudopotentials can be used on all other atoms. There exist core specific basis such as pcX-n and cc-pCVXZ, but their usage is not necessary (based on basis set convergence studies on small molecules). Note that the pcseg-n basis sets are nice to use as they come with their own ADMM basis.
How do I make my calculation more accurate ?
The first necessity is to have a good ground state calculation. Thus, any change of paramter improving the SCF will reflect on the quality of the LR-TDDFT calculation.
Within XAS_TDP, a few parameters may play a role:
EPS_FILTERandEPS_PGFare used for screening. Lowering those will result in slower but more accurate calculations- Increasing the
GRIDdimensions will improve the quality of the numerical integration of the XC kernel. The upper limit for the number of angular points is 974. There is not upper limit for the radial point, but performance may suffer if too large. Usually, something likeGRID C 250 500is sufficient. - Increasing the
RI_REGIONin theKERNELsubsection will lead to a more accurate projection of the density on the RI basis. Basis functions centered on atoms within the region (defined by a sphere around the excited atom) are added for the projection. Increasing this parameter should come together with denserGRID.
By default, the RI basis used for the integral and the projection is autamatically generated. The quality of the RI basis can be changed via the AUTO_BASIS keyword in the DFT section. To improve from the default MEDIUM size, one can used: AUTO_BASIS RI_XAS LARGE/HUGE. Note that an external RI basis set can also be provided.
How do I make my calculation faster ?
All points mentioned above in the accuracy section can also be tweaked for performence. In general, lowering accuracy will lead to faster calculations.
For large systems, it is recommanded to use the OT iterative solver rather than the default full digonalization. This should improve the scaling of the method. See the NaAlO$_2$ example.
The Tamm-Dancoff approximation is well established and generally yields results as good as full TDDFT. It is moreover much cheaper than the latter. It is turned on by default, but you may want to make sure it is enabled.
The use of ADMM is highly recommanded for large systems, where the ground state HFX evaluation is the main bottleneck. It is also recommanded to use pseudopotentials on all atoms that are not excited as all-electron basis set tend to be large. If there exist no proper ADMM basis for the all-electron basis used for the excited atom, you may use the full basis as AUX_FIT. If the latter is very diffuse, it may be beneficial to remove the most diffuse elements.
The code is also both MPI and OMP parallelized. Using more core will, to a certain degree, speedup your calculations as well.
How do I plot a spectrum from the *.spectrum output file
For each donor state in the system, the *.spectrum file contains a list of excitation energies and corresponding oscillator strengths. This yields a stick spectrum which needs to be artificially broadened to match experiments. This is typically done using Gaussian of Lorentzian functions. Note that in case of spin-orbit calculation at the L-edge, results for the singlet, triplet and SOC excitations are given.
Remember that XAS LR-TDDFT produces an accurate spectrum, but it is usually wrongly positioned on the energy axis. Again, to match experiment, a rigid shift must be applied to the result.
My calculation yields a wrong/unphysical result, what do I do ?
Assuming that the ground state calculation is sound and well converged, there are two main causes for failure.
The proper core state was not found/identified. In the XAS_TDP part of the output file, look for the Mulliken population analysis and the overlap with a STO basis. Both quantities should close to 1.0. If they are not, there is a problem with the selected core state. You may want to increase the value of N_SEARCH in DONOR_STATES to scan additional Kohn-Sham orbitals. If there are multiple atoms that are equivalent under symmetry, make sure to use the LOCALIZE keyword of DONOR_STATES as well.
The numerical integration of the XC kernel $(\lambda|f_{xc}|\mu)$ is not accurate enough. In this case, you may want to increase the density of the integration grid with the GRID keyword, increase the RI_REGION and/or increase the quality of the genenerated RI basis set. Note that using an external RI basis set may also help as the basis generation scheme may fail (For example: def2-QZVP for Zn, use def2-QZVP-RIFIT in this case).
Finally, keep in mind that calculated spectra need to be rigidly shifted by some energy to match experiment.
First-principles correction scheme
As mentioned above, XAS LR-TDDFT results need to be rigidly shifted to match experiments. This is due to self-interaction error and the lack of orbital relaxation upon the creation of the core hole. An ab-initio correction scheme was developed to address these issues. Theory and benchmarks were published in The Journal of Chemical Physics 2021, DOI: 10.1063/5.0058124. Please cite this paper if you were to use this method.
Brief theory recap
XAS LR-TDDFT yield excitation energies as correction to ground state Kohn-Sham orbital energy differences, namely:
$$ \omega = \varepsilon_a - \varepsilon_I + \Delta_{xc}, $$
where $\varepsilon_a$ is the orbital energy of a virtual MO and $\varepsilon_I$ the energy of the donor core MO. Under Koopman's condition, these energies are interpreted as the electron affinity and and the ionization potential (IP). However, DFT is notoriously bad at prediction accurate absolute orbital eigenvalues. Therefore, and because $|\varepsilon_I| >> |\varepsilon_a|$, excitation energies are expected to be widely improved if the DFT energy $\varepsilon_I$ were to be replace by an accurate value of the IP.
The IP can be accurately calculated using the second-order electron propagator equation:
$$ IP_I = -\varepsilon_I - \frac{1}{2} \sum_{ajk}\frac{|\langle Ia||jk\rangle|}{-IP_I + \varepsilon_a -\varepsilon_j -\varepsilon_k} $$