Research Article

Chemistry—Methods [Link]/10.1002/cmtd.202400071

[Link]

Selecting the Optimal DFT Functionals for Reproducing the

UV-Vis Properties of Transition-Metal Photocatalysts
Péter Pál Fehér*[a] and András Stirling*[a, b]

We have applied our practice-oriented TDDFT benchmark poorer results while the range separated hybrids CAM B3LYP
strategy to assess the performance of pure, hybrid, range- and ωB97X D offer the most consistent performance for both
separated hybrid and double-hybrid functionals for reproducing datasets. Among the metals, iron proved to be most difficult for
the electronic spectra of a set of 137 transition-metal com- TDDFT: even M06 predicts singlet-singlet absorption spectra
plexes. Our results enable simple pre-screening of new and quintet ground states erroneously, especially for weak
transition-metal based photocatalysts. The reference data is ligand fields (i. e., only Fe N bonds). The functionals that do not
based on recently published measurements, from which we exhibit this behaviour are the pure (GGA, meta-GGA) func-
have established a new database. The database is named tionals, but the role of HF exchange in the spectral prediction is
TMPHOTCAT-137 to reflect the number of complexes (with Cu, equivocal; while functionals with high amounts of HF exchange
Ru, Ir, Fe, Au, Mo and W centres) included and the fact that all perform insufficiently, smaller amount of exact exchange yields
of them are either proven or potential photocatalysts. We have overall better performance (M06, B3LYP). We have also shown
found that the M06 functional is the best performer both in that inclusion of spin-orbit coupling for the heavier metals does
terms of practical accuracy and consistency for all metals except not improve the results. The updated spectrum optimizer code,
Fe. While B3LYP is similarly accurate for Ru and Ir complexes, for the TMPHOTCAT-137 database and the Jupyter Notebook used
the rest of the metals significant wavelength scaling is for analysis are available at [Link]/PeterF1234/spectrum_
necessary. Compared to our previous benchmark for organic optimizer.
photocatalysts, double-hybrid functionals exhibit considerably

Introduction source of energy. In this work we focus on complexes that

absorb in the visible range. The rational development of such
The importance of transition metal complexes (TMCs) in complexes benefits greatly from computational predictions and
chemistry is unquestionable.[1–3] This is mainly due to the therefore reliable theoretical methods that offer reasonable
versatility of their physical chemical properties which is accuracy are highly sought after.[7] It is widely accepted that
achieved through the fine tuning of their molecular structures. DFT offers the best compromise in terms of cost to perform-
Considering the metal centres, the remarkable diversity of their ance, however, there are considerable differences between the
electronic properties is exploited through different d valence individual functionals. This can be explored by benchmark
shell electron and spin configurations. The ligands provide both studies that reveal the strengths and weaknesses of each
stabilization and further opportunities for the fine tuning of functional. Benchmarks may focus on different aspects of
electronic structure on the metal and even allow certain performance such as the reproduction of overall lineshapes,
processes to take place on the ligands themselves.[4–5] Ligands vertical transition energies, absorption strength, the 0–0
that are on the forefront of current research are organic transition energy, etc.[8] Most of these properties are not directly
scaffolds with conjugated electronic structure as they allow for available from experiments (i. e., from the measured absorption
the design of complexes that absorb light,[6] a sustainable spectrum) and thus theoretical best estimates (TBE) are used.
TBEs, however, are based on high level wavefunction methods
that have severe limitations on the size of the dataset or on the
[a] Institute of Organic Chemistry, Research Centre for Natural Sciences, number of atoms in the molecules.[9–10] The use of experimental
Budapest, Hungary
[b] Department of Chemistry, Eszterházy Károly Catholic University, Eger,
spectra is therefore a better alternative for the larger TMCs.
Hungary Several benchmarks that use this approach for moderately sized
Correspondence: Péter Pál Fehér and András Stirling, Institute of Organic organic molecules are already available.[11–12] However, studies
Chemistry, Research Centre for Natural Sciences, Magyar tudósok körútja 2, for TMCs are considerably scarcer and to our knowledge there
Budapest 1117, Hungary. are no papers that extensively benchmark TDDFT performance
Email: [Link]@[Link] and [Link]@[Link]
for TMCs. One reason for this is that DFT generally performs
Supporting Information for this article is available on the WWW under worse for these systems,[13–14] while another is the lack of
[Link]
established databases. The generally poor performance is
© 2024 The Author(s). Chemistry - Methods published by Chemistry Europe
usually attributed to two main factors: predictions of unphysical
and Wiley-VCH GmbH. This is an open access article under the terms of the
Creative Commons Attribution License, which permits use, distribution and excitations[15] and neglect of spin-orbit coupling (SOC).[16–20] The
reproduction in any medium, provided the original work is properly cited. former originates from the incorrect description of charge-

Chemistry—Methods 2025, e202400071 (1 of 11) © 2024 The Author(s). Chemistry - Methods published by Chemistry Europe and Wiley-VCH GmbH
 Research Article
Chemistry—Methods [Link]/10.1002/cmtd.202400071

transfer (CT) states by the DFT functionals and leads to

additional states that are not present when the same molecule
is calculated with high level wavefunction methods. A deep
discussion of this error, however, is not the goal of the present
study as we have already explored this in our previous paper,[21]
and there are more specialized works in the literature.[22] We
only point out here that these additional states can be classified
as “ghost” and “spurious” states based on the associated
oscillator strengths. The “ghost” states have oscillator strength
close to zero and they do not contribute to the shape of the
UV-vis spectrum, whereas “spurious” states have nonzero
oscillator strengths and they may appear in the spectrum. As
our benchmark is based on the comparison of line shapes, only
the error from the “spurious” states contribute to the perform-
ance results. We have shown, however, that even these states
are not prominent for organic photocatalysts.[21] The largest
effect instead lies in the fact that for functionals of the first two
rungs of Jacob’s ladder in DFT more states need to be
calculated, reducing their advantage in terms of computational
cost.[21] One of the goals of this work to scrutinize how this
picture changes for transition metal complexes where metal-to-
ligand CT states are usually low in energy and have a significant
contribution to the visible absorption.
The other possible factor mentioned above is SOC which is
almost exclusively attributed to the presence of heavy
elements. It is a relativistic effect that, assuming a singlet
ground state, allows the otherwise spin-forbidden triplet
excitations to merge with singlet states and yields additional
mixed states that have nonzero oscillator strength. The result is Scheme 1. Outlines of the Ru complexes including 30 individual TMCs.

that peak intensity ratios can change,[10] or even additional well-

separated or shoulder peaks can appear next to the peaks from
pure singlet-singlet excitations.[16] Note that this is not to be algorithm and the ability to specify the interval in which the
confused with the similar effect provided by vibronic transi- computed and experimental spectra are compared. The aim of
tions. In terms of computational resources, SOC-TDDFT calcu- this study is to benchmark a representative set of DFT
lations are more demanding than regular TDDFT as 4N + 1 functionals with different levels of accuracy and computational
states (ground state, singlet excited states and the three spin cost for computing electronic spectra of a large set of transition
configurations of the triplets; N is the number of states to be metal complexes against experimental spectra. We wish to
calculated) must be considered and special all-electron basis obtain a general picture about the performance of the func-
sets optimized for relativistic calculations need to be employed. tionals in TDDFT calculations and also strive to provide insights
Therefore, we limited this part of the investigation to complexes that contribute to a more detailed understanding of the extent
that contain the heavier Ru and Ir centres as these are expected of the challenges the specific transition metals pose.
to exhibit the most pronounced SOC.
To alleviate the issue of the lack of reference data, we report
a new dataset that contains the UV-vis spectra of 137 TMCs. Computational Methods
Schemes 1–2 shows traditional and more recent Ru and Ir based
photocatalysts, while Schemes 3–4 includes TMCs that contain Benchmark methodology. Our benchmark approach was
the more common Cu and Fe centres together with some introduced in our previous paper.[21] Here, we only provide the
examples of Au, Mo and W complexes. We have obtained the outline: 1) the stick spectrum (pairs of vertical excitation
dataset by digitizing spectra from recent papers, focusing on energies Xi and oscillator strengths fi) of a molecule is computed
those that include photocatalytic activity and additional redox with TDDFT. 2) a sum of Gaussian broadening functions using a
potential data for further studies. We use this dataset to scaling (α) and a bandwidth (Δ) parameter is fitted as given by
benchmark 14 DFT functionals covering pure (GGA, meta-GGA), Equation 1:
hybrid and double-hybrid approximations with the inclusion of
X � �
SOC for selected functionals. For the evaluation of performance,
n
f ðx a 1 X i Þ2
sð xÞ ¼ piffiffiffiffiffi exp (1)
we have used our method developed for a smaller set of i¼1 D 2p 2D2
organic photocatalysts.[21] The corresponding Python code has
also been improved with added options for the optimization

Chemistry—Methods 2025, e202400071 (2 of 11) © 2024 The Author(s). Chemistry - Methods published by Chemistry Europe and Wiley-VCH GmbH
 Research Article
Chemistry—Methods [Link]/10.1002/cmtd.202400071

Scheme 2. Outlines of the Ir complexes including 29 individual TMCs.

Scheme 4. Outlines of the Fe, Au, Mo, and W complexes including 26

individual TMCs.

tal reference and the fitted spectrum using constraints 0.7 < α
< 1.3 and 0.1 eV < Δ < 0.45 eV. In practice the optimization is
performed on a grid defined by α and Δ. As seen in Eq. 1, for a
given spectrum all excitations are treated with the same α and
Δ. The code is available on Github.
The assumption behind our methodology (i. e. convoluting
the vertical excitation energies from the minimum of the
potential energy surface with a suitable Gaussian function) is
that all the effects not present in a simple stick spectrum (such
as vibronic effects) can be incorporated trough a uniform shift
and broadening. We consider this a reasonable approximation
that is necessary for a benchmark study of such magnitude
Scheme 3. Outlines of the Cu complexes including 52 individual TMCs. because properly taking into account these effects for the large
number of excited states would increase the computational
time with orders of magnitude. Full vibronic calculations are
nowadays viable options over simple TD calculations to obtain
where n is the number of calculated excitations and s(x) is the more detailed insights into the spectroscopic properties of
fitted spectrum as a function of x (energy). Note that Eq. 1 is molecules, including those that contain transition metal
calculated in the energy domain (which is the natural scale of centers.[23–26] For a small set of molecules, we have also
excitations), so a 1 is applied. 3) Then α and Δ are optimized to performed full vibronic calculations employing the CAM B3LYP
minimize the mean square error (MSE) between the experimen- and B3LYP functionals and taking into account the first 60

Chemistry—Methods 2025, e202400071 (3 of 11) © 2024 The Author(s). Chemistry - Methods published by Chemistry Europe and Wiley-VCH GmbH
 Research Article
Chemistry—Methods [Link]/10.1002/cmtd.202400071

electronic states using the ORCA package and compared theory SVP basis set throughout,[44] and modelled the solvent via the
with experiment (see SI for additional details and for the full implicit Solvation Model based on Density (SMD)[45] within the
results). The main conclusion is that in most cases these linear-response formalism. The def2-SVP basis applies effective
calculations lead to α values that are very similar to those core potential (ECP) for the metals below the first row.[46] We
provided by our Gaussian fitting procedure for vertical have also found that the upgrade to def2-TZVP or employing
excitations. This indicates that from a practical point of view the diffuse functions have minor impact on the final spectra
wavelength scaling required for each functional can be (Figure S1, Table S1) but increase the memory demand signifi-
obtained without the explicit inclusion of vibronic effects in the cantly and may introduce convergence issues. For technical
benchmark. reasons,[47] the double-hybrid calculations were performed as
As the absorption band maxima are usually red-shifted with single point TDDFT calculations on M06-L optimized geometries
respect to the vertical excitations due to various reasons, using the ORCA 5.0.4 software.[48–49] The Resolution of Identity
including the vibronic effects[27] one can expect that the (RI) approximation was applied for the Coulomb, exchange
optimized α is somewhat larger than 1 for ideal cases. These (COSX for HF exchange) and MP2 correlation integrals.[50] The
effects could be eliminated from the benchmark by considering Tamm-Dancoff approximation (TDA) was disabled in all calcu-
theoretical best estimates as references (comparing the vertical lations. As we have seen previously for organic molecules,
excitations). Then an α � 1 parameter would indicate excellent further memory savings can be achieved without impacting the
performance, but for metal compounds this route is not viable performance of spectrum prediction by reducing the number of
as theoretical best estimates are not yet available. Instead, we states calculated with TDDFT from 60 (used for pure and
define a � 10 % interval for α for identifying good performances hybrids) to 40 as double-hybrid functionals do not suffer from
from practical point of view, because the typical shifts of the “ghost” or “spurious” states.
experimental band maxima are well within this interval over the SOC calculations. The relativistic calculations were per-
majority of the visible spectrum (considering 0.1–0.2 eV devia- formed with the ORCA 5.0.4 software. We have used the Zero-
tion as the expected systematic error associated with the Order Regular Approximation (ZORA) and the corresponding
vibronic effects.[27–28]) all-electron basis set SARC-ZORA-TZVP for Ru and Ir together
The value of the optimized α parameter for a functional is with ZORA-DEF2-SVP for the rest of the atoms.[51–52] The RI
therefore a measure of how close the computed stick spectrum approximation was used in both the SCF and the SOC
is aligned with experimental peak positions, while Δ and the calculations. For SOC, the spin-orbit mean-field RI-SOMF(1X)
final MSE can be used as indicators of how well the functional is approximation was employed where RI is applied to the
able to reproduce the spectral line shape (e. g., the widths of Coulomb part.[53] We have tested two other implemented
individual peaks, the relative heights of two or more peaks, approximations (RI-SOMF(4X) and VEFF( 2X)), and found that
whether shoulders merge into the main peaks, etc.). The MSE they yield practically the same results (see SI for a comparison).
represents the functionals in a best-case situation (optimal α All SOC-TDDFT calculations were performed as single point
and Δ) and provides the best estimate for their overall perform- calculations on geometries optimized with the corresponding
ance. We also evaluate the functionals on the basis of how close functionals through the nonrelativistic approach described in
α is to unity, because this information is very useful for the the TDDFT section above. Solvation was also included via SMD.
practice where the prediction of peak positions based on the For pure functionals 60 singlets and 60 triplets were calculated
calculation of vertical excitations is commonly employed. We accounting for a total of 240 mixed excitations from the ground
note however that error cancellation may be behind favorable state. Note that it is important to keep the number of singlets
performances. the same as in regular TDDFT calculations because they
TDDFT calculations. The benchmark was performed using determine how low (in nanometres) the calculated spectrum
the following set of functionals: PBE,[29] B97-D3,[30] TPSS,[31] M06- will reach. For hybrid functionals, however, we reduced this
L,[32] M06,[33] M06-2X,[33] B3LYP,[34] CAM B3LYP,[35] ωB97X D,[36] number to 50 as the number of unphysical states is expected to
B2PLYP,[37] ωB2PLYP,[38] B2GP-PLYP,[39] SCS-PBE-QIDH[40] and SOS- be lower. This approach still leads to a very high memory cost
ωPBEPP86.[41] They represent the main types of DFT approxima- (up to 14GB per core) that is close to regular TDDFT for double-
tions and can be grouped into pure (GGA, meta-GGA), hybrid hybrids, however, the time it takes to finish is only about 1–
(global or range separated) and double-hybrid functionals. The 2 hours running on 8 physical cores of an AMD EPYC 7763 CPU.
pure and hybrid functionals are highly cost-efficient so their Note that SOC-TDDF for double-hybrids was unreasonably
results are based on full geometry optimizations together with expensive to perform.
frequency and TDDFT calculations. These calculations were The dataset. The spectra for the complexes shown in
performed using the Gaussian16 software package.[42] The Scheme 1 were obtained from recent literature via digitization
optimizations included the effect of the solvent cavity, except using the WebPlotDigitizer tool.[54] The initial geometries for the
for the Cu complexes where optimizations were done in gas optimizations described above were taken from the reported X-
phase (for three additional cases, see the Notes column in ray structures or xyz coordinates. If no such data were available,
supporting file TMPHOTCAT-137_notes.xlsx). Dispersion was the structures were built manually or derived from analogous
included via D3 with BJ damping if not implicitly included in complexes with X-ray or xyz data. For the more flexible
the functional.[43] For TDDFT, the calculation of the first 60 complexes, the CREST software was used to explore the
singlet vertical excited states was requested. We used the def2- conformational space.[55] CREST can also account for isomerism

Chemistry—Methods 2025, e202400071 (4 of 11) © 2024 The Author(s). Chemistry - Methods published by Chemistry Europe and Wiley-VCH GmbH
 Research Article
Chemistry—Methods [Link]/10.1002/cmtd.202400071

(i. e., fac/mer, cis/trans) as it allows bond rearrangement. notes can be found in the supporting file TMPHOTCAT-137_
However, in several cases the most stable conformer did not [Link].
correspond to the isomer inferred from NMR measurements by
the authors of the original papers. This was most common for Ir
complexes, for which it is known that not only the semi- Results and Discussion
empirical XTB but also DFT often provide wrong energy
ordering.[56] In these problematic cases we have used the isomer Benchmark using the complete dataset. First, we discuss the
reported in their original paper to perform the benchmark but overall performance of the functionals using all the TMCs
also calculated the other isomer for which the analysis is included in the dataset. The optimized values of parameters α
provided in the SI. Unfortunately, the paper from which the and Δ are listed in Table 1. For the spectrum of a typical TM-
largest subset of our Fe complex data was taken provides no photocatalyst calculated by a given functional we recommend
indication about fac or mer preference so the most stable mer the use of the values listed in Table 1. It can be seen in both
isomer of the complexes was used. More details about the Figure 1 and Figure 2 that M06 is clearly the best performer,
initial structures, the solvents used, references and additional followed by B3LYP. Regarding the quality of the estimations
displayed in Figure 1, the range separated hybrids ωB97X D
and CAM B3LYP also yield MSEs close to M06, however, they
require significant scaling through the α parameter. This short-
Table 1. Mean bandwidths (Δ), scaling factors (α) and MSE values averaged coming is apparent in Figure 2, which shows that the excitation
over all the complexes. The methods are ordered according to their mean wavelengths calculated by both ωB97X D and CAM B3LYP
MSE values. need to be scaled by more than 10 % in about 75 % of the
functional α Δ (eV) cases. On the other hand, double-hybrids B2GP-PLYP and SCS-
M06 0.97 (� 0.01) 0.27 (� 0.01)
B3LYP 0.95 (� 0.01) 0.28 (� 0.01)
ωB97X-D 1.13 (� 0.01) 0.30 (� 0.01)
CAM B3LYP 1.13 (� 0.01) 0.29 (� 0.01)
B2PLYP 0.94 (� 0.01) 0.29 (� 0.01)
B2GP-PLYP 1.01 (� 0.01) 0.31 (� 0.01)
SCS-PBE-QIDH 1.00 (� 0.01) 0.33 (� 0.01)
SOS-ωPBEPP86 1.07 (� 0.01) 0.32 (� 0.01)
ωB2PLYP 1.14 (� 0.01) 0.35 (� 0.01)
M06-L 0.85 (� 0.01) 0.26 (� 0.01)
TPSS 0.83 (� 0.01) 0.27 (� 0.01)
B97-D3 0.85 (� 0.01) 0.25 (� 0.01)
M06-2X 1.18 (� 0.01) 0.36 (� 0.01)
PBE 0.83 (� 0.01) 0.27 (� 0.01)

Figure 2. The distribution of the scale factor (α) and optimized bandwidth
(Δ) values and its breakdown into the performances of the individual
Figure 1. The average MSE of the functionals for the reproduction of functionals. The bar plots show the number of occurrences of each method
experimental line shapes using optimized fitting parameters. The average in the regions marked with colours: teal, � 10 % practical accuracy window;
was calculated on the complete dataset of 137 TMCs. The error bars dark orchid, optimal region; coral, wide region. The functionals are ordered
represent 95 % confidence intervals. according to their mean MSE errors (Figure 1).

Chemistry—Methods 2025, e202400071 (5 of 11) © 2024 The Author(s). Chemistry - Methods published by Chemistry Europe and Wiley-VCH GmbH
 Research Article
Chemistry—Methods [Link]/10.1002/cmtd.202400071

PBE-QIDH show promising MSEs in Figure 1, but they perform

similarly to B3LYP in the � 10 % test with about 50 % of cases
predicted well. However, this is still considerably behind the
80 % hit rate offered by M06.
One of the other two regions in Figure 2 indicate cases
where the optimal spectral fit requires the flattening of the line
shape (wide region, Δ > 0.35 eV). This behaviour is usually a
sign of poor performance and can be caused by the inaccurate
description of peak separation or by spurious states that are
“smeared out” by the spectrum fitting process. In our previous
work on organic photocatalysts (OPCs) we have observed that
almost exclusively pure DFT functionals contribute to this
region. For TMCs, however, Figure 2 indicates that this trend is
reversed: functionals with the largest amounts of HF exchange
(M06-2X, ωB97X D, CAM B3LYP and the double-hybrids) have
the largest contribution to the “wide” region while pure Figure 3. Stacked bar plots showing the distributions of α > 1.1; 1.1 � α � 0.9;
functionals yield the smallest (consistently less than 25 %). The and α < 0.9 for each functional, averaged over all molecules. The functionals
inclusion of HF exchange, however, is not necessarily detrimen- are ordered according to their mean MSE values.
tal as M06 and B3LYP perform just as well as pure functionals
do with a lower than 25 % hit rate in the wide region. This
suggests that functionals with small or zero HF exchange ratios M O bond enthalpies of transition-metal oxide diatomics: the
are more likely to capture the finer details of the spectra of performance of the M06-2X functional was significantly poorer
TMCs. Note that in addition to the HF exchange ratio, the than that of CAM B3LYP.[57]
performance of highly parametrized functionals such as M06 Another issue which can affect the performance of the
and M06-2X strongly depends on the datasets used to optimize meta-GGA functionals is the gauge variance of the kinetic
their internal parameters. The training set of both functionals energy density as implemented in Gaussian and ORCA.
includes excitation energies, however, M06-2X is parametrized Imposing gauge-invariance can improve of the performances of
only for nonmetals[33] that can explain its exceptionally poor some in TDDFT calculations, depending on the sensitivity of the
results for our TMC set. Range separation also has a positive exchange energy integrand to the kinetic energy density.[58] For
effect as ωB97X D featuring a high number of parameters example in case of TPSS functional the effect of the gauge
optimized on a non-transition-metal specific train set[36] still variance was shown to be small, whereas for M06-2X the effect
provides reasonable optimized spectra based on its mean MSE was predicted to be large.[58] However, in the present case the
in Figure 1. The almost identical performance of the 18- four meta-GGA functionals do not follow the expected pattern,
parameter ωB97X D and the 3-parameter CAM B3LYP also hence we conclude that gauge variance is not the primary
indicates that range separation has a strong influence on the reason of their performances.
spectrum prediction and that range separation is overall The third region in Figure 2 is a somewhat arbitrarily
beneficial: these functionals offer systematic errors that can be defined “optimal” region that relaxes the � 10 % requirement to
corrected by a properly chosen scale factor (α). These errors are � 20 % for α but has a limit for Δ (< 0.3 eV). These values are
also consistent in both the OPC and TMPC sets as we have based on the previous organic photocatalyst results where this
obtained the same α = 1.13 � 0.01 in both cases. Therefore, we region has been shown to contain spectra that are reasonably
can conclude that the scaled range separated hybrids are solid well-reproduced. Most functionals here have a hit rate around
all-around performers. The double-hybrids are more difficult to 50 %, which is surpassed by M06, B3LYP, CAM B3LYP and
assess as the range separated ωB2PLYP perform similar to somewhat surprisingly by B97-D3. The only functionals that are
ωB97X D and CAM B3LYP in the over- and underestimation considerably poor in this region are ωB2PLYP and M06-2X. Note
test (Figure 3), however the fitted spectra offer considerably that the performance of ωB2PLYP for TMCs here is in huge
worse MSE (Figure 1). The errors of the other double-hybrids contrast to what we have observed for the OPC molecules as its
are less systematic and thus further investigation is warranted hit rate fell from 80 % (OPC) to 20 % (TMC).
in their case. The under- and overestimation tendency of the functionals
We note that heavy parametrization can also be an obstacle shown in Figure 3 follow the well-established trend that pure
to good performance as comparison of the results of M06-2X functionals overestimate the absorption wavelengths, which is
and CAM B3LYP shows: both functionals are fitted to non- (over )compensated by the addition of HF exchange into the
metals and their performances were similar for organic functional. Only the double-hybrids B2PLYP, B2GP-PLYP and
photocatalysts.[21] Still, M06-2X with its 32 fitted parameters is SCS-PBE-QIDH deviate from this trend. Therefore, for most
less flexible here, whereas CAM B3LYP is less constrained to functionals the mean values of the optimized α and Δ
nonmetallic situations and can perform better for metal- parameters collected in Table S2 can be used to correct their
containing compounds. Similar trend could be observed for systematic errors and provide a reasonably good prediction for
ground state properties of transition-metal compounds, such as experimental spectra (the lower the MSE in Table S2 the better).

Chemistry—Methods 2025, e202400071 (6 of 11) © 2024 The Author(s). Chemistry - Methods published by Chemistry Europe and Wiley-VCH GmbH
 Research Article
Chemistry—Methods [Link]/10.1002/cmtd.202400071

The effect of spin-orbit coupling (SOC) and analysis of the

Ru/Ir dataset. The results in Figure 2 and their comparison with
the organic photocatalysts indicate that the presence of the
metal centre has a negative effect on the performance of the
functionals. This effect has a clear dependence on the amount
of HF exchange in the functionals as those with 20 and 27 %
(B3LYP, M06) ratios perform better than both those with no HF
exchange (pure DFT) and those with higher amounts of exact
exchange (M06-2X and range separated hybrids). The poor
performance of the double-hybrids also indicates that the
addition of MP2 correlation has no benefit either. A difference,
however, between OPCs and TMCs is that the heavier metals
can cause additional effects like SOC, an effect the computa-
tional protocol so far has not considered. As these calculations
are highly expensive, we have limited the discussion to Ru and
Ir complexes and used the mean-field RI-SOMF(1X) approxima-
tion to calculate spectra with the M06-L, M06 and ωB97X D
functionals on non-relativistic optimized geometries.
Figure 4 shows that the errors of the optimized spectra for
complexes in the Ru and Ir subsets follow the same trend as for
the complete dataset. M06-L has been improved by the
inclusion of SOC, while the other two functionals perform
slightly poorer. Figure 5 shows that the optimized parameters
have essentially the same distribution for M06 and M06-L while
for ωB97X D they are shifted away from the � 10 % and
“optimal” regions. Therefore, SOC is not the primary reason for
the performance decrease observed for TMCs.
Comparison of the complete dataset in Figure 2 with the
Ru/Ir dataset in Figure 5 reveals that the complexes of the
currently most popular metals used in photocatalysis are Figure 5. The distribution of the scale factor (α) and optimized bandwidth
described adequately by the M06 and B3LYP functionals as they (Δ) values and its breakdown into the performances of the individual
functionals for the merged Ru and Ir subsets that include 30 and 29
offer a nearly perfect hit rate for the � 10 % practical accuracy individual complexes, respectively. The bar plots show the number of
window with around 75 % of the predicted spectra falling into occurrences of each method in the regions marked with colours: teal, � 10 %
the “optimal” region. These results indicate a performance level practical accuracy window; dark orchid, optimal region; coral, wide region.
The functionals are ordered according to their mean MSE errors (Figure 1).
similar to that of the best methods obtained for the OPCs
(double-hybrids, DLPNO-CCSD). For comparison, these two

functionals have considerably inferior � 10 % performance

(~60 % for M06 and ~40 % for B3LYP) for the OPC dataset. This
points to error cancellations explaining the good performance
of these functionals for the Ru/Ir dataset. However, the narrow
confidence interval for M06 in Figure 4 indicates a high level of
consistency in the predictions.
Performance for the individual subsets. Comparison of the
results for the whole dataset in Figure 2 and for the Ru/Ir subset
in Figure 5 reveals huge differences for some functionals, B3LYP
being the most apparent case. Therefore, to reveal additional
insights about how DFT treats excitations for TMCs we have
also analysed every subset individually. Figure 6 shows the
distribution of the optimized scale factors for each functional
(except CAM B3LYP and SOS-ωPBEPP86 which are shown in
Figure S7) over each TMC subset. Note that the subsets differ in
sizes, i. e., they each contain a different number of individual
Figure 4. The average MSE of the functionals for the reproduction of TMCs which is reflected most apparently by the differences
experimental line shapes using optimized fitting parameters, where the between the areas under the KDE curves. The best performing
results of SOC effects are also included for 3 functionals. The average was
calculated on the merged Ru and Ir subsets that include 30 and 29 individual M06 functional shows exceptional practical accuracy and a high
complexes, respectively. The error bars represent 95 % confidence intervals.

Chemistry—Methods 2025, e202400071 (7 of 11) © 2024 The Author(s). Chemistry - Methods published by Chemistry Europe and Wiley-VCH GmbH
 Research Article
Chemistry—Methods [Link]/10.1002/cmtd.202400071

Figure 6. Distribution of the scale factor values for the different TMC subsets. The dashed lines indicate the kernel density estimate (KDE). The functionals are
ordered in a way that the first, second and third rows contain pure, hybrid and double-hybrid functionals, respectively. The CAM B3LYP and SOS-ωPBEPP86
functionals are not included to make the results more presentable. CAM B3LYP has a distribution almost identical to that of ωB97X D, whereas SOS-
ωPBEPP86 yields a distribution similar to that of ωB2PLYP; see Figure S7.

level of consistency for all metals except Fe where it has an whether these results are due to the possibility of low energy
essentially uniform distribution of scale factor parameters. triplet or quintet states. The Boltzmann-averaged spectra
The apparent failure for Fe complexes can be observed for calculated with the M06 and M06-L functionals in Figure S8
all other hybrids and for double-hybrids as well, but Figure 6 show considerable differences compared to the spectra that
indicates that pure functionals have no such issue. Based on contain only singlet-singlet excitations. In case of M06, this
inspection of the fitted spectra (see SI), pure DFT performs leads to better performance, while for M06-L worse results are
better due to the additional excitations at higher wavelengths obtained, see Figure S9. This can originate from the different
that are not present for hybrid DFT. These are usually population ratios: both functionals predict quintet ground
considered as spurious states and are dismissed as computa- states for TMCs with Fe N bonds (FeC series, Fe_bpy, Fe_tpy,
tional error, however, for Fe complexes (especially Fe_tpy, Fe_ Fe_dcpp, FeTP), but Table S3 indicates higher quintet-singlet
bpy, FeC_4, FeC_5) they improve the agreement with experi- gaps for M06 (100 % quintet population) compared to M06-L
ment. The detrimental effect of HF exchange is also apparent (non-negligible singlet population). Although the quintet spec-
for complexes with Fe C bonds (FeII_carbene, FeII_CNC) where tra for M06 offers better match with experiment, the results are
hybrid functionals with low amount of HF exchange (M06, evidently wrong as there is clear evidence that [Fe(bpy)3]2 +,
B3LYP) reproduce additional peaks that are not present for the [Fe(dcpp)2]2 + and [Fe(tpy)2]2 + have singlet ground states.[59]
rest of the hybrid functionals. Interestingly, however, even pure Note, however, that high-spin Fe(II) terpyridine complexes can
DFT fails to reproduce the states at higher wavelengths for be obtained via substitution at 6 and 6” positions.[60] These
some complexes, such as FeC_2 or FeC_3. We have checked modifications usually increase the steric bulk around the Fe(II)

Chemistry—Methods 2025, e202400071 (8 of 11) © 2024 The Author(s). Chemistry - Methods published by Chemistry Europe and Wiley-VCH GmbH
 Research Article
Chemistry—Methods [Link]/10.1002/cmtd.202400071

centre and thus weaken the ligand field allowing the distortion orbitals. Therefore, it is surprising that their performances for
of the octahedral geometry to stabilize the quintet state. On predicting the excitation spectra of the selected TMCs are
the other hand, introduction of electron-rich substituents or inferior to several functionals from the lower rungs. To explain
using carbene ligands (replacement of Fe N with Fe C bonds) this observation, we first note that their development required
strengthens the ligand field and provides the opposite fitting specific parameters, such as the weights of the HF
effect.[61–62] This is well-reproduced by M06 (Table S4) as the exchange and MP2 correlation (all double-hybrids), the range
Boltzmann spectra for the FeII_carbene and FeII_CNC series are separation parameter (ωB2PLYP) and the spin scaling parame-
essentially identical to the singlet-singlet spectra in these cases. ters (SCS-PBE-QIDH, SOS-ωPBEPP86). The optimization of these
As the evaluation of spin state energetics performance is parameters has been done dominantly on small main-group
outside of the scope of this benchmark, we only note here that compounds. The fitting of the exchange and correlation
the state ordering strongly depends on the amount of HF coefficients has been done against thermochemical data,
exchange and a common practice is to tune this parameter (i. e., whereas the other parameters (for ωB2PLYP, SCS-PBE-QIDH and
B3LYP*) to obtain satisfactory results for a given complex under SOS-ωPBEPP86) have been fitted against excitation energies of
study.[63–64] Our results obtained for M06 and M06-L reflect the small molecules of main-group elements including both valence
established trend where adding more HF exchange stabilizes and Rydberg excitations with nearly equal weights. Our
the high-spin ground states. For more details, see Ref. [48] and procedure focuses on the lowest ~4.5 eV region and favours
related papers that deal with the issue in a more focused excitations with large oscillator strengths of large complexes
manner. featuring transition-metals, therefore we cannot expect a
The Cu subset has the largest contribution to the database performance seen for small molecules composed of C, N, O and
with 53 individual TMCs. All these complexes contain Cu(I) H atoms.
centres and have a closed shell d10 ground state electron Another aspect which can contribute to the inferior
configuration with a low chance for observing low-lying metal- performance of the double-hybrids is the multireference
centred excited states.[68] Therefore, these complexes have character of many of the TMCs. We note that the static
simpler photophysics compared to Fe(II) complexes which is correlation effects arising from the multireference character can
reflected by the results in Figure 6. The overall well-performing affect all the functionals here. However, the CIS(D) terms built
functionals M06, B3LYP, ωB97X D and CAM B3LYP offer in the double-hybrids for excitations can amplify these effects
reliable results here as well. The only exception is that B3LYP thus weakening their performances. Note, that double-hybrids
requires a scaling for Cu complexes that is significantly different reproduce the excitation spectra of organic photocatalysts
than that for the other TMCs. M06-2X needs scaling that is remarkably well with this procedure which indicates that the
expected to reach well outside our reasonable limits (i. e., it has presence of TMs is the culprit of their weaker performances.
more than 30 % error); pure functionals exhibit a bimodal scale All in all, the inferior performance of double-hybrid func-
factor distribution, while double-hybrids fail to provide a tionals for TMCs can be attributed to the combined effects of
consistent performance. To analyse the behaviour of pure the missing physics properly addressing the static correlation
functionals, we have used B97-D3 results as it has the fewest effects and parts of the built-in ingredients optimized for
entries at the 30 % error limit (α reaching 0.7). The Cu set can be experimental values not directly relevant for the present
divided into 4 subsets based on the phosphine ligands: BINAP, protocol (such as Rydberg states or charge transfers with small
DPEPhos, XantPhos and dppf (see Scheme 3). The α distribution or zero oscillator strengths).
shown in Figure S11 indicates that the bimodality in Figure 6 is
caused by the difference between DPEPhos and dppf com-
plexes, for which the α values peak around 1.0 and 0.7, Conclusions
respectively. Note that dppf contains a ferrocene connecting
the two phosphorous atoms, so it would be reasonable to In this work, we have evaluated how the PBE, B97-D3, TPSS,
assume it to be the source of this large error. This, however, is M06-L, M06, M06-2X, B3LYP, CAM B3LYP, ωB97X D, B2PLYP,
in contrast with the findings for Fe complexes where pure ωB2PLYP, B2GP-PLYP, SCS-PBE-QIDH, and SOS-ωPBEPP86 DFT
functionals are the most reliable and also to the fact that hybrid functionals perform for the task of reproducing the experimen-
functionals have no such issue, see M06 in Figure S12. There- tal UV-vis spectra of 137 transition metal complexes containing
fore, the ferrocene group is not the reason behind the dppf Ru, Ir, Cu, Fe, Au, Mo and W centres. We have benchmarked the
results here. The other phosphines provide an essentially functionals using an updated version of our method that was
uniform distribution and therefore pure hybrids can be developed for organic photocatalysts and also compiled an
concluded to be unreliable for them. Interestingly, the distribu- original database from recent UV-vis measurements. We have
tion of the MSE fitting errors shown in Figure S11 reveals a found that the M06 functional performs outstandingly. The only
different behaviour as only the BINAP complexes yield notice- complexes it fails for are the weak ligand field Fe complexes,
ably poorer results. which may have quintet ground states. In these cases, the
Performance of the double-hybrid functionals. Double- consideration of a Boltzmann distribution improves the results
hybrid functionals have been developed to combine the best significantly, however, this is a result of error cancellation as
aspects of density functional and wavefunction theories. They neither the ground state multiplicity, nor the corresponding
belong to the fifth rung of Jacob’s ladder employing virtual spectra are predicted consistently. Considering the rest of the

Chemistry—Methods 2025, e202400071 (9 of 11) © 2024 The Author(s). Chemistry - Methods published by Chemistry Europe and Wiley-VCH GmbH
 Research Article
Chemistry—Methods [Link]/10.1002/cmtd.202400071

functionals, B3LYP performs almost as well as M06 for the Ru [5] R. F. J. Epping, D. Vesseur, M. Zhou, B. de Bruin, ACS Catal. 2023, 13,
5428–5448.
and Ir complexes but requires significant scaling for the rest of
[6] C. K. Prier, D. A. Rankic, D. W. C. MacMillan, Chem. Rev. 2013, 113, 5322–
the metals. The range separated ωB97X D and CAM B3LYP are 5363.
overall close to B3LYP, but they offer more consistent errors, [7] A. Nandy, C. Duan, M. G. Taylor, F. Liu, A. H. Steeves, H. J. Kulik, Chem.
i. e., systematic underestimation of the predicted wavelengths. Rev. 2021, 121, 9927–10000.
[8] A. D. Laurent, D. Jacquemin, Int. J. Quantum Chem. 2013, 113, 2019–
The double-hybrids show generally inferior performance, and 2039.
their use is not recommended for TMCs as far as reproducing [9] D. A. Wappett, L. Goerigk, J. Chem. Theory Comput. 2023, 19, 8365–8383.
experimentally observed electronic spectra is concerned. This is [10] D. Jacquemin, F. Kossoski, F. Gam, M. Boggio-Pasqua, P.-F. Loos, J.
Chem. Theory Comput 2023, 19, 8782–8800.
in sharp contrast with the results obtained for the organic [11] A. M. Grabarz, B. Ośmiałowski, Molecules 2021, 26, 7434.
photocatalysts, where they were the top performers. A similar [12] D. Jacquemin, I. Duchemin, X. Blase, J. Chem. Theory Comput. 2015, 11,
situation is observed for M06-2X, which indicates that high 5340–5359.
[13] C. Daniel, Coord. Chem. Rev. 2015, 282–283, 19–32.
amounts of HF exchange with or without MP2 correlation is [14] S. Mai, F. Plasser, J. Dorn, M. Fumanal, C. Daniel, L. González, Coord.
detrimental for the prediction of TMC spectra. The pure (GGA) Chem. Rev. 2018, 361, 74–97.
functionals offer relatively good performance for the Ru and Ir [15] F. Maschietto, M. Campetella, J. Sanz García, C. Adamo, I. Ciofini, J.
Chem. Phys. 2021, 154.
complexes but for the other metal they suffer from the large [16] C. Liao, J. M. Kasper, A. J. Jenkins, P. Yang, E. R. Batista, M. J. Frisch, X. Li,
number of unphysical states that can appear as extra peaks at JACS Au 2023, 3, 358–367.
high wavelengths or cause a lack of excitations (we have [17] S. Kotaru, P. Pokhilko, A. I. Krylov, J. Chem. Phys. 2022, 157.
[18] B. de Souza, G. Farias, F. Neese, R. Izsák, J. Chem. Theory Comput. 2019,
calculated 60 transitions for all functionals, except double- 15, 1896–1904.
hybrids) at lower wavelengths. Surprisingly, however, these [19] L. Freitag, L. González, Inorg. Chem. 2014, 53, 6415–6426.
functionals perform best for Fe complexes. The accurate [20] H. Brahim, C. Daniel, A. Rahmouni, Int. J. of Quantum Chem. 2011, 112,
2085–2097.
description of such complexes evidently requires multireference
[21] P. P. Fehér, Á. Madarász, A. Stirling, Chemistry-Methods 2023, 3.
calculations, but as our results show in some cases pure DFT or [22] J. M. Herbert, Phys. Chem. Chem. Phys. 2024, 26, 3755–3794.
hybrid DFT functionals for strong ligand field complexes can [23] F. Vazart, C. Latouche, J. Bloino, V. Barone, J. Phys. Chem. B 2014, 119,
7253–7257.
offer adequate predictions.
[24] H. Belaidi, S. Belaidi, C. Katan, C. Latouche, A. Boucekkine, J. Mol. Model.
2016, 22 265.
[25] F. Massuyeau, E. Faulques, C. Latouche, J. Chem. Theor. Comp. 2017, 13,
1748–1755.
Notes [26] C. Latouche, D. Skouteris, F. Palazzetti, V. Barone, J. Chem. Theor. Comp.
2015, 11, 3281–3289.
No competing financial interests have been declared. [27] S. Bai, R. Mansour, L. Stojanović, J. M. Toldo, M. Barbatti, J. Mol. Model.
2020, 26, 107.
[28] F. J. Avila Ferrer, J. Cerezo, E. Stendardo, R. Improta, F. Santoro, J. Chem.
Theory Comput. 2013, 9, 2072–2082.
Acknowledgments [29] J. P. Perdew, K. Burke, M. Ernzerhof, Phys. Rev. Lett. 1996, 77, 3865–3868.
[30] S. Grimme, S. Ehrlich, L. Goerigk, J. Comput. Chem. 2011, 32, 1456–1465.
[31] J. Tao, J. P. Perdew, V. N. Staroverov, G. E. Scuseria, Phys. Rev. Lett. 2003,
The authors acknowledge support from Grant K132236 of the 91.
NRDI Office, Hungary. We acknowledge KIFÜ for awarding us [32] Y. Zhao, D. G. Truhlar, J. Chem. Phys. 2006, 125, 194101.
access to HPC resource based in Hungary. [33] Y. Zhao, D. G. Truhlar, Theor. Chem. Acc. 2007, 120, 215–241.
[34] A. D. Becke, J. Chem. Phys. 1993, 98, 5648–5652.
[35] T. Yanai, D. P. Tew, N. C. Handy, Chem. Phys. Lett. 2004, 393, 51–57.
[36] J.-D. Chai, M. Head-Gordon, Phys. Chem. Chem. Phys. 2008, 10, 6615.
Conflict of Interests [37] S. Grimme, J. Chem. Phys. 2006, 124, 034108.
[38] M. Casanova-Páez, M. B. Dardis, L. Goerigk, J. Chem. Theory Comput.
2019, 15, 4735–4744.
The authors declare no conflict of interest. [39] A. Karton, A. Tarnopolsky, J.-F. Lamère, G. C. Schatz, J. M. L. Martin, J.
Phys. Chem. A 2008, 112, 12868–12886.
[40] É. Brémond, J. C. Sancho-García, Á. J. Pérez-Jiménez, C. Adamo, J. Chem.
Phys. 2014, 141, 031101.
Data Availability Statement [41] M. Casanova-Páez, L. Goerigk, J. Chem. Theory Comput. 2021, 17, 5165–
5186.
[42] Gaussian 16, Revision A.03, M. J. Frisch, G. W. Trucks, H. B. Schlegel, G. E.
The data that support the findings of this study are openly
Scuseria, M. A. Robb, J. R. Cheeseman, G. Scalmani, V. Barone, G. A.
available in Github at [Link]/PeterF1234/spectrum_opti- Petersson, H. Nakatsuji, X. Li, M. Caricato, A. V. Marenich, J. Bloino, B. G.
mizer, reference number 1. Janesko, R. Gomperts, B. Mennucci, H. P. Hratchian, J. V. Ortiz, A. F.
Izmaylov, J. L. Sonnenberg, D. Williams-Young, F. Ding, F. Lipparini, F.
Egidi, J. Goings, B. Peng, A. Petrone, T. Henderson, D. Ranasinghe, V. G.
Zakrzewski, J. Gao, N. Rega, G. Zheng, W. Liang, M. Hada, M. Ehara, K.
Toyota, R. Fukuda, J. Hasegawa, M. Ishida, T. Nakajima, Y. Honda, O.
[1] A. Biffis, P. Centomo, A. Del Zotto, M. Zecca, Chem. Rev. 2018, 118, Kitao, H. Nakai, T. Vreven, K. Throssell, J. A. Jr. Montgomery, J. E. Peralta,
2249–2295. F. Ogliaro, M. J. Bearpark, J. J. Heyd, E. N. Brothers, K. N. Kudin, V. N.
[2] R. E. Rodríguez-Lugo, M. Trincado, M. Vogt, F. Tewes, G. Santiso- Staroverov, T. A. Keith, R. Kobayashi, J. Normand, K. Raghavachari, A. P.
Quinones, H. Grützmacher, Nat. Chem. 2013, 5, 342–347. Rendell, J. C. Burant, S. S. Iyengar, J. Tomasi, M. Cossi, J. M. Millam, M.
[3] G. S. Nambafu, A. M. Hollas, S. Zhang, P. S. Rice, D. Boglaienko, J. L. Klene, C. Adamo, R. Cammi, J. W. Ochterski, R. L. Martin, K. Morokuma,
Fulton, M. Li, Q. Huang, Y. Zhu, D. M. Reed, V. L. Sprenkle, G. Li, Nat. O. Farkas, J. B. Foresman, J. Gaussian, Inc. Wallingford CT, 2016.
Commun. 2024, 15. [43] S. Grimme, S. Ehrlich, L. Goerigk, J. Comput. Chem. 2011, 32, 1456–1465.
[4] V. Lyaskovskyy, B. de Bruin, ACS Catal. 2012, 2, 270–279. [44] F. Weigend, R. Ahlrichs, Phys. Chem. Chem. Phys. 2005, 7, 3297.

Chemistry—Methods 2025, e202400071 (10 of 11) © 2024 The Author(s). Chemistry - Methods published by Chemistry Europe and Wiley-VCH GmbH
 Research Article
Chemistry—Methods [Link]/10.1002/cmtd.202400071

[45] A. V. Marenich, C. J. Cramer, D. G. Truhlar, J. Phys. Chem. B 2009, 113, [60] E. C. Constable, G. Baum, E. Bill, R. Dyson, R. van Eldik, D. Fenske, S.
6378–6396. Kaderli, D. Morris, A. Neubrand, M. Neuburger, D. R. Smith, K. Wieghardt,
[46] D. Andrae, U. Häußermann, M. Dolg, H. Stoll, H. Preuß, Theor. Chim. Acta M. Zehnder, A. D. Zuberbühler, Chem. Eur. J. 1999, 5, 498–508.
1990, 77, 123–141. [61] S. Suhr, N. Schröter, M. Kleoff, N. Neuman, D. Hunger, R. Walter, C.
[47] Unfortunately, our HPC infrastructure is not well suited for jobs that Lücke, F. Stein, S. Demeshko, H. Liu, H.-U. Reissig, J. van Slageren, B.
require lots of memory (1 GB/core, so CPU cores must be over- Sarkar, Inorg. Chem. 2023, 62, 6375–6386.
provisioned and waste resources), which has been the bottleneck of the [62] Y. Liu, T. Harlang, S. E. Canton, P. Chábera, K. Suárez-Alcántara, A.
present study. Fleckhaus, D. A. Vithanage, E. Göransson, A. Corani, R. Lomoth, V.
[48] F. Neese, Wiley Interdiscip. Rev. Comput. Mol. Sci. 2011, 2, 73–78. Sundström, K. Wärnmark, Chem. Commun. 2013, 49, 6412.
[49] F. Neese, Wiley Interdiscip. Rev. Comput. Mol. Sci. 2022, 12, e1606. [63] E. I. Ioannidis, H. J. Kulik, J. Chem. Phys. 2015, 143.
[50] F. Neese, J. Comput. Chem. 2003, 24, 1740–1747. [64] M. Pápai, G. Vankó, C. de Graaf, T. Rozgonyi, J. Chem. Theory Comput.
[51] J. D. Rolfes, F. Neese, D. A. Pantazis, J. Comput. Chem. 2020, 41, 1842– 2012, 9, 509–519.
1849. [65] D. C. Ashley, Coord. Chem. Rev. 2017, 337, 97–111.
[52] D. A. Pantazis, X.-Y. Chen, C. R. Landis, F. Neese, J. Chem. Theory Comput. [66] E. I. Ioannidis, H. J. Kulik, J. Phys. Chem. A 2017, 121, 874–884.
2008, 4, 908–919. [67] S. Song, M.-C. Kim, E. Sim, A. Benali, O. Heinonen, K. Burke, J. Chem.
[53] F. Neese, J. Chem. Phys. 2005, 122. Theory Comput. 2018, 14, 2304–2311.
[54] [Link] last accessed: 2024. Q1. [68] J. Beaudelot, S. Oger, S. Peruško, T.-A. Phan, T. Teunens, C. Moucheron,
[55] P. Pracht, F. Bohle, S. Grimme, Phys. Chem. Chem. Phys. 2020, 22, 7169– G. Evano, Chem. Rev. 2022, 122, 16365–16609.
7192.
[56] K. H. Hopmann, Organometallics 2016, 35, 3795–3807.
[57] K. A. Moltved, K. P. Kepp, ChemPhysChem 2019, 20, 3210–3220.
[58] R. Grotjahn, F. Furche, M. Kaupp, J. Chem. Phys. 2022, 157, 111102.
[59] L. L. Jamula, A. M. Brown, D. Guo, J. K. McCusker, Inorg. Chem. 2013, 53, Manuscript received: November 24, 2024
15–17. Version of record online: ■■■, ■■■■

Chemistry—Methods 2025, e202400071 (11 of 11) © 2024 The Author(s). Chemistry - Methods published by Chemistry Europe and Wiley-VCH GmbH