Download as pdf or txt
Download as pdf or txt
You are on page 1of 11

PCCP

View Article Online


PAPER View Journal | View Issue

Machine learning approaches to understand and


predict rate constants for organic processes in
Published on 14 January 2021. Downloaded by University of Toledo on 5/16/2021 6:06:50 PM.

Cite this: Phys. Chem. Chem. Phys.,


2021, 23, 2742 mixtures containing ionic liquids†
Tamar L. Greaves, *a Karin S. Schaffarczyk McHale,b Raphael F. Burkart-Radke,b
Jason B. Harper *b and Tu C. Le*a

The ability to tailor the constituent ions in ionic liquids (ILs) is highly advantageous as it provides access
to solvents with a range of physicochemical properties. However, this benefit also leads to large
compositional spaces that need to be explored to optimise systems, often involving time consuming
experimental work. The use of machine learning methods is an effective way to gain insight based on
existing data, to develop structure–property relationships and to allow the prediction of ionic liquid
properties. Here we have applied machine learning models to experimentally determined rate constants
of a representative organic process (the reaction of pyridine with benzyl bromide) in IL–acetonitrile
mixtures. Multiple linear regression (MLREM) and artificial neural networks (BRANNLP) were both able to
model the data well. The MLREM model was able to identify the structural features on the cations and
anions that had the greatest effect on the rate constant. Secondly, predictive MLREM and BRANNLP
models were developed from the full initial set of rate constant data. From these models, a large number
of predictions (49000) of rate constant were made for mixtures of different ionic liquids, at different
proportions of ionic liquid and molecular solvent, at different temperatures. A selection of these predic-
Received 10th August 2020, tions were tested experimentally, including through the preparation of novel ionic liquids, with overall
Accepted 14th January 2021 good agreement between the predicted and experimental data. This study highlights the benefits of
DOI: 10.1039/d0cp04227g using machine learning methods on kinetic data in ionic liquid mixtures to enable the development
of rigorous structure–property relationships across multiple variables simultaneously, and to predict
rsc.li/pccp properties of new ILs and experimental conditions.

Introduction consuming and often difficult to prepare, characterise and trial


meaningfully large sets of ILs. This compositional space rapidly
Ionic liquids (ILs) are salts that are liquid below B100 1C.1–4 increases for ILs in most applications, where they are typically
Simple ILs consist of one type of cation and one type of anion mixed with a molecular solvent (where the nature of that
and, even with these simple systems there are a vast number solvent and the proportion of each solvent in the reaction
of possible combinations.2,3,5,6 The benefit of these many mixture are important), have solutes present (once again, the
combinations is that the properties of ILs can be modified nature and concentration are important), and the process can
through changing their chemical structure.7 However, it is time be carried out under different conditions (such as tempera-
ture). There are a few approaches that have been used in the
IL field to address this issue, including the use of high
a
College of Science Engineering and Health, RMIT University, Melbourne, VIC throughput techniques to experimentally screen libraries of
3001, Australia. E-mail: tamar.greaves@rmit.edu.au, tu.le@rmit.edu.au
b
ILs and IL containing systems,8–17 and computational methods
School of Chemistry, University of New South Wales, Sydney, NSW 2052,
Australia. E-mail: j.harper@unsw.edu.au
to predict physicochemical properties.18–21 With the expansion
† Electronic supplementary information (ESI) available: Summary of the ionic of data available on various IL physicochemical properties,
liquids used in the training models; quantitative values of each descriptor used machine learning methods have been used to build predictive
for the ionic liquid constituents; comparison of the predicted versus experimental models for such properties; these properties include
values for k2 and log(k2) for other models considered; analysis of MLREM model
density,22–28 viscosity,25,26,29–31 surface tension of neat ILs,32,33
using 18 descriptors; general experimental methods; complete synthetic details
for the ionic liquids 1–17, composition of stock solutions and measured rate
surface tension of binary26,30,34–36 or tertiary IL mixtures,26,37–40
constant data for mixtures containing ionic liquids 1–17; comparison of predicted toxicity,41,42 and gas solubilities in neat ILs or IL containing
and experimental data for ionic liquids 1–17. See DOI: 10.1039/d0cp04227g solvents.43–51

2742 | Phys. Chem. Chem. Phys., 2021, 23, 2742--2752 This journal is © the Owner Societies 2021
View Article Online

Paper PCCP

The key benefits of using machine learning methods are the solvents, where much literature is available, with the focus on
ability to extract complex trends for multivariate systems, to predicting reaction products.70–74 Similarly, machine learning has
develop robust predictions, and to gain insight into the sys- been used to optimise specific reactions, where the reaction
tems, allowing the design of new ones to achieve a desired conditions (including the catalyst, solvent, reagent and tempera-
outcome. There are many different machine learning methods ture) were included as inputs. As an extension, some groups have
such as support vector machine,52 relevance vector machine,53 developed integrated high throughput synthesis robots that can
and random forest,54 but none of these methods has been found interact directly with the machine learning models, feeding in
to always outperform others.55–59 The most effective method has new information, and learning from the previous data.74,75 How-
been found to be dependent on the dataset. We have previously ever, it has been noted that, while these models and high
reported that the surface tension and liquid nanostructure of throughput approaches are very beneficial, there are still limita-
Published on 14 January 2021. Downloaded by University of Toledo on 5/16/2021 6:06:50 PM.

ternary mixtures of non-stoichiometric protic ionic liquids with tions. These limitations included the need for some human
water can be modelled well by multiple linear regression (MLR) supervision,74 and the importance of nuanced descriptors that
and artificial neural networks (ANN).60 The MLR models have the can represent more complex reaction features, such as the three
benefit of providing quantitative structure–property relationships dimensional conformations of the molecules.76 In work closely
that can be used for designing new systems. In contrast, the ANN related to this study, the reaction rate of a similar substitution
provided better agreement with both the test and training set data process was explored through integrating quantum mechanical
but could not allow the quantitative comparison of different (QM) calculations with computer-aided molecular design (CAMD)
factors on the property being modelled. to identify solvents that might lead to faster reaction rates.77 Their
In this paper we have employed machine learning methods algorithm was able to successfully identify nitromethane as a
on the use of ILs as co-solvents in a substitution reaction. solvent that would result in a 40% rate constant increase com-
Specifically, we have used second order rate constant (k2) data pared to acetonitrile, which was the solvent that gave the greatest
for the substitution reaction between pyridine and benzyl rate constant initially.78,79
bromide (Scheme 1) that has been systematically collected by However, while successful, these methods require high quality
Harper et al.61–69 In this series of papers, they have quantified computational input. While these integrated QM approaches have
the effect on k2 of IL–acetonitrile mixtures, exploring how the advantage of requiring less experimental data, they do need to
varying the cation structure, the anion structure, the proportion be able to model effectively the solvent mixture (and hence the
of IL, and the temperature affects the reaction outcome. Most resultant solvent effects). The use of IL–molecular solvents signifi-
of the ILs trialled led to an increase in the rate constant, which cantly increases the complexity of the system and appropriate
was identified as being due to an entropic gain from loss of IL models are still in development.80,81 As such, solvent effects are
structure around the precursors on moving to the transition explored in this paper using available experimental data, and ML
state; this benefit offset a smaller enthalpic cost. Variation of methods on such experimental data has the potential to provide
the nature of the cation showed that accessibility to the charged greater understanding of the microscopic origin of the rate
centre was important and resulted in rate constant increases constant change, to predict reaction outcomes, and perhaps even
relative to acetonitrile, whereas the reaction had a lower k2 than to lead to IL solvent design.
in neat acetonitrile if the added IL had a sterically hindered In this paper we have used 160 pieces of rate constant data
charge centre.61,64,66 The importance of the charged centre was for the reaction outlined in Scheme 1, in different ionic liquid–
further demonstrated by increased rate constants in cases with acetonitrile mixtures to develop MLR and ANN models‡. The
electron withdrawing group on the cation.66,68 Consequently, MLR model was used to enable complex structure–property
there is significant quantitative data available for this reaction relationships to be developed. The ANN model was used for
in mixtures of ionic liquids and acetonitrile that are suitable for prediction of rate constants for the reaction in ionic liquid–
analysis using machine learning methods. Herein, we investigated acetonitrile mixtures where the ionic liquids were based on
whether further trends in the rate constant data can be identified, each possible combination of cations and anions used in this
and if robust models can be developed that can be used to predict study, along with other variables including temperature and
rate constants for this reaction in other ionic liquid mixtures. proportion of ionic liquid in the reaction mixture. From this
While we are not aware of any previous research into using extensive list, a representative set of 17 reactions (with condi-
machine learning methods to analyse the outcomes of organic tions that spanned the range used in initial evaluations) were
reactions in ILs, the suitability of such methods has clearly been then used to validate the model.
shown for analysing data on organic reactions in molecular

Experimental
Rate constant data on the reaction in Scheme 1 has been
reported by Harper et al.61–69 A total of 160 data points

Scheme 1 The bimolecular nucleophilic substitution reaction between ‡ The effect of variations in the electrophile and nucleophile have also been
pyridine and benzyl bromide, which has been examined extensively in considered,61,66,67,95,96 but these data have not been included in this study as
mixtures containing an ionic liquid.61–69 there are limited data on the addition of a number of other variables.

This journal is © the Owner Societies 2021 Phys. Chem. Chem. Phys., 2021, 23, 2742--2752 | 2743
View Article Online

PCCP Paper

Table 1 Descriptors used in these models, their description and what is described by that descriptor. Cation descriptors are from Dragon82

Descriptor Description Portion described Initiala MLREMb BRANNLPc


Intercept n/a | | ‘
ILmolefraction Mole fraction of ionic liquid in the reaction mixture n/a | | |
InverseT Inverse temperature n/a | | |
AnionTotalNsingle Number of sp3 hybridised nitrogen Anion | ‘ ‘
AnionTotalC Number of carbon atoms Anion | ‘ |
AnionTotalS Number of sulphur atoms Anion | ‘ ‘
AnionTotalO Number of oxygen atoms Anion | ‘ ‘
AnionTotalF Number of fluorine atoms Anion ‘ ‘ |
AnionTotalB Number of boron atoms Anion ‘ | |
Published on 14 January 2021. Downloaded by University of Toledo on 5/16/2021 6:06:50 PM.

AnionTotalNTriple Number of sp hybridised nitrogen atoms Anion ‘ | |


AnionTotalP Number of phosphorus atoms Anion ‘ ‘ |
nDB Number of double bonds Cation ‘ ‘ |
nAB Number of aromatic systems (ring size) Cation | ‘ |
RBN Number of rotatable bonds Cation | | |
RBF Rotatable bond fraction Cation | ‘ |
ARR Aromatic ratio in the cation Cation | ‘ |
nCIC Number of rings (cyclomatic number) Cation | | |
nCIR Number of circuits Cation | | ‘
nCp Number of primary sp3 hybridised carbon atoms Cation | | |
nCs Number of secondary sp3 hybridised carbon atoms Cation | | ‘
nCrs Number of secondary sp3 hybridised carbon atoms Cation | | |
in ring structures
nCar Number of sp2 hybridised carbon atoms in Cation | ‘ |
aromaticsystems
nO Number of oxygen atoms Cation | | |
nCL Number of chlorine atoms Cation | ‘ ‘
N-073 Number of sp3 hybridised nitrogen atoms attached to Cation | | |
anaromatic ring
N-075 Number of sp2 hybridised nitrogen atoms Cation | | |
N-079 Number of nitrogen centres with a positive charge Cation | | ‘
nHAcc Number of acceptor atoms for hydrogen bonds (N, O, F) Cation | | |
Hy Hydrophilic factor Cation | | |
Ui Unsaturation index Cation | ‘ |
AMR Ghose–Crippen molar refractivity Cation | | ‘
MLOGP Moriguchi octanol–water partition coefficient (log P) Cation | | |
TPSA Total polar surface area Cation | ‘ ‘
a
Removed descriptors were underrepresented in the data sets. b Removed descriptors had no effect on the standard error of prediction in the
model using MLREM. c Removed descriptors had zero weighting in the prediction model using BRANNLP.

(covering 36 different ILs, a range of proportions of ionic liquid properties such as log P, unsaturation index, hydrophilic factor,
in the IL–acetonitrile mixtures and a range of temperatures; for molar refractivity and polar surface area. Descriptors that had
a summary and link to the Figshare repository see ESI†). The the same values for different ILs were excluded. The quantita-
data were randomly partitioned three independent times – 80% tive value of each descriptor is provided in Tables S2 and S3 of
for training (128 points in the set) and 20% for testing (32 the ESI† for the cations and anions, respectively. Descriptors
points in the set). For each partitioning, modelling and statis- that were highly correlated, with a correlation coefficient of
tical analysis were done independently. greater than 0.98, or did not have much variation across the
The ILs and reaction conditions were represented by 33 data set (only changed for one or two samples) were removed by
descriptors, which are summarised and described in Table 1. the BioModeller package85–87 that was used for constructing the
Most descriptors are only for either the cation or the anion, and machine learning models; descriptors with correlation between
this is specified in the ‘portion described’ column. An addi- 0.9 and 0.98 are not automatically removed, but subject to
tional intercept term for fitting was included for the Multiple human screening to avoid the exclusion of unrelated descrip-
Linear Regression with Expectation Maximization (MLREM) tors. The correlation coefficient between the secondary sp3
models. The ILmolefraction and InverseT described the mole hybridised carbon atoms (nCs) and the number of rotatable
fraction of the IL in the ionic liquid reaction mixture and the bonds (RBN) had a correlation coefficient of 0.96, however
inverse of the temperature, respectively. The inverse of the these were both retained due to the chemically different
temperature was used since there is a linear relationship with features they represent. The correlation matrix of the initial
ln(k).83 The ILs were described by the remaining 30 descriptors, pool of descriptors is provided in Table S4 of the ESI.†
which were taken from the Dragon descriptor data base,82,84 Sparse feature selection algorithms including Multiple
with eight descriptors for the anions, and 22 for the cations. Linear Regression with Expectation Maximization (MLREM)88
These Dragon descriptors include constitutional indices, func- and Bayesian Regularized Artificial Neural Networks with
tional group counts, atom-centered fragments and molecular Laplacian prior (BRANNLP)89 were employed to further prune

2744 | Phys. Chem. Chem. Phys., 2021, 23, 2742--2752 This journal is © the Owner Societies 2021
View Article Online

Paper PCCP

out irrelevant descriptors. Unlike the basic multiple linear The reaction mixtures for the kinetic studies were prepared
regression (MLR) and Bayesian Regularized Artificial Neural under pseudo first order conditions such that they contained at
Networks with Gaussian Prior (BRANNGP), which make use of a least a 10-fold excess of pyridine relative to benzyl bromide.
Gaussian prior and minimize the sum of the squares of errors Reaction progress was monitored to 495% completion using
1
between the measured and predicted values, MLREM and H NMR spectroscopy, noting depletion of the signal corres-
BRANNLP use Laplacian prior to automatically set the weights ponding to the benzylic protons of benzyl bromide at d ca.
of irrelevant descriptors to zero and minimize the modulus 4.0. The exceptions were the kinetic data determined in mix-
between the measured and predicted values. A detailed expla- tures containing either (i) the ionic liquid [bmim][PF6] 2, where
nation of these algorithms can be found elsewhere.85–90 These the reaction was monitored as above but only to ca. 10%
BRANN based methods are difficult to overtrain since they have completion and analysed using initial rates methodology; or
Published on 14 January 2021. Downloaded by University of Toledo on 5/16/2021 6:06:50 PM.

an objective Bayesian criterion for stopping training, and are (ii) the ionic liquid [bmpi][PF6] 11, where reaction progress was
unlikely to overfit since they train on a number of effective monitored (again to 495% completion) using 1H NMR spectro-
network weights and remove those that are not relevant. The scopy following formation of the signal corresponding to
neural networks had three layers, input, hidden and output. the benzylic protons of the product of the reaction shown in
The number of nodes in the input layer was equal to the Scheme 1 at d ca. 5.6.
number of descriptors, while the hidden layer had 3 nodes
and the output layer had one single node corresponding to the
rate constant of the organic reactions. The rate constant, k2, Results and discussion
was initially used as the output in the models, however it was
quickly identified that log(k2) was more appropriate for the The modelling process was conducted on the rate constant data
MLREM models, which is consistent with ln(k) being propor- for the reaction shown in Scheme 1 that included data for
tional to the activation energy.83 Consequently, log(k2) was used 36 different ILs (33 variants of the cation and four variants for
as the output for the MLREM models, while k2 and log(k2) were the anion of the ILs, see Fig. S1, ESI†). It is important to note
used for the neural network models. We randomly partitioned that not all ILs have the same number of data points (based on
the data set into a training and a test set which consisted of available literature) and that the chemical space covered by the
80% and 20% of the data, respectively. This process was done ions considered is limited. The latter is a function of the
three times to ensure generalized results. Note that no valida- structural requirements for the components to ensure a liquid
tion set was used because the Bayesian regularization provides phase, but also the practicalities of preparing the ionic liquids.
a criterion for stopping training and therefore eliminates the Whilst this feature should not preclude analysis, it should be
need for a validation set.85,86 The quality of the models was borne in mind in subsequent discussions.
assessed using the square of the correlation coefficient between The 160 data points were randomly partitioned into 80%
the predicted and measured values (R2) as well as the standard training and 20% test sets. This process was done three times
error of estimation (SEE) and prediction (SEP) for the training to remove bias and improve the generalisation of the model,
set and test sets. Good models have R2 of close to 1.0 together with all three sets giving comparable results. The MLREM,
with low SEE and SEP. BRANNGP and BRANNLP models were run independently on
each training set, and then tested on the corresponding test set.
Preparation of ionic liquids and determination of The descriptors used for each model are provided in Table 1.
rate constants The BRANNGP had a similar performance to the BRANNLP
Pyridine and benzyl bromide were distilled91 and stored over (see ESI†). In the algorithms, the MLREM model minimises the
activated molecular sieves (3 Å) at 4 1C prior to use. Acetonitrile sum of the squares of error, where the BRANNLP minimises the
was distilled91 and stored over activated molecular sieves (3 Å) magnitude of the absolute errors.
at room temperature under nitrogen. All other chemicals were The comparison of the predicted versus experimental values
purified through literature methods91 and used immediately. for k2 and log(k2) for MLREM are shown in Fig. S2 (ESI†) for one
The ionic liquids used were generally prepared through the of the three training/test sets. There is clearly a poor correlation
N-alkylation of the appropriate precursor, followed by between predicted and experimental values when k2 is used as
anion metathesis (complete synthetic procedures are given in the output (Fig. S2a, ESI†), which is vastly improved by using
the ESI†). log(k2) (Fig. S2b, ESI†). The squared correlation coefficients (R2)
All kinetic measurements were obtained by monitoring for the use of k2 and log(k2) for MLREM were 0.77 and 0.96,
reaction progress using 1H NMR spectroscopy with either a respectively. This large difference in the correlation quality is
Bruker Avance III 400 (400 MHz, 1H), a Bruker Avance III 500 consistent with log(k2) having a more linear relationship
(500 MHz, 1H) or a Bruker Avance III 600 (600 MHz, 1H) with the descriptors than k2. Consequently, log(k2) has been
spectrometer equipped with either a BBO, BBFO or TBI probe. used as the output for all MLREM models. The predicted versus
The results were shown to be reproducible regardless of the experimental values for log(k2) are provided in Fig. 1a for the
NMR spectrometer and probe used. The temperature of the MLREM models, and in Fig. 1b and c for log(k2) and k2 for the
NMR spectrometer was calibrated using a thermocouple con- BRANNLP models. The comparison plots for the BRANNGP
taining ethanol. models are provided in Fig. S3 of the ESI.† The predicted versus

This journal is © the Owner Societies 2021 Phys. Chem. Chem. Phys., 2021, 23, 2742--2752 | 2745
View Article Online

PCCP Paper

Table 2 The bimolecular rate constants for the reaction shown in


Scheme 1 at 295 K in reaction statistical results for the MLREM, BRANNLP
and BRANNGP models, averaged over three training and testing sets

Training Test
Output Model Effective weights R2 SEE Q2 SEP
k2 MLREM 29.0 0.77 0.101 0.64 0.135
BRANNLP 19.7 0.90 0.055 0.88 0.067
BRANNGP 31.2 0.92 0.046 0.90 0.056

log(k2) MLREM 28.0 0.96 0.053 0.91 0.063


Published on 14 January 2021. Downloaded by University of Toledo on 5/16/2021 6:06:50 PM.

BRANNLP 31.0 0.96 0.038 0.92 0.056


BRANNGP 32.0 0.97 0.035 0.94 0.086

experimental values are shown on the same training/test set for


all three models for both k2 and log(k2).
The statistical results for the MLR, BRANNLP and BRANNGP
models are given in Table 2 for both k2 and log(k2) outputs.
These are the average of the statistics for the three training and
test sets. The effective weights give a measure of the complexity
of the model. The squared correlation coefficients (R2) and
descriptors removed were for the total number of fluorine,
boron or phosphorus atoms, and without these present, the
standard error of estimation (SEE) were obtained for the train-
ing set of data, and similarly the squared correlation (Q2) and
standard error of prediction (SEP) were obtained on the test set
of data. Consistent with Fig. S1a (ESI†), the MLREM model
using k2 as the output had poor performance, while the MLREM
with log(k2) had a high R2 and Q2, showing the model has good
predictive capability. The BRANNLP had good R2 and Q2 using
both k2 and log(k2) as the output, but noticeably better perfor-
mance using the latter for all measures. The BRANNGP had
comparable statistics to BRANNLP.
While the best model based on the statistics is the BRANNLP
model with log(k2) as the output, the MLREM model also has
excellent statistics, and an additional advantage in that it can
be used to identify structure–property relationships through
visualising the contribution from each descriptor. Of the origi-
nal set of 33 descriptors, there were 5 removed, as noted in
Table 1; one cation descriptor and four anion descriptors.
While these descriptors are relevant to represent the full set
of ILs, the random partitioning of the data into training and
test sets led to some of the training sets not having the range of
either cations or anions that required these descriptors, and
hence the algorithms removed them. This removal resulted in a
total of four descriptors for the anions and 21 for the cations. It
should be noted that the four anion tetrafluoroborate and
hexafluorophosphate anions were equivalent in the models.
The scaled weightings for each of the remaining 28 descrip-
Fig. 1 (a) Predicted log(k2) versus experimental log(k2) for the reaction tors for the MLREM model with log(k2) as the output are shown
shown in Scheme 1 as analysed by MLREM, (b) predicted k2 versus in Fig. 2. The input descriptors with a positive contribution are
experimental k2 for the reaction shown in Scheme 1 as analysed by expected to lead to larger rate constants. It is evident that the
BRANNLP, and (c) predicted log(k2) versus experimental log(k2) for the
largest contribution is positive, and from the nCs (total number
reaction shown in Scheme 1 as analysed by BRANNLP. The training set is
shown by the circles and the test set by the triangles. of secondary sp3 hybridised carbon atoms in the cation). The
largest negative contributions were from the RBN (number of
rotatable bonds in the cation) and AMR (molar refractivity

2746 | Phys. Chem. Chem. Phys., 2021, 23, 2742--2752 This journal is © the Owner Societies 2021
View Article Online

Paper PCCP

descriptors are more robust for this set of ILs and reactions.
Interestingly, these descriptors do not simply correlate with
those with negligible coefficients for the MLREM model (Fig. 2),
and hence while they may not have variation within each
training set, some do appear to have a significant contribution.

Prediction of rate constants


Additional MLREM, BRANNLP and BRANNGP models were
generated, and used to predict the rate constants of the
Menschutkin reaction (Scheme 1) in IL–acetonitrile solvents.
Published on 14 January 2021. Downloaded by University of Toledo on 5/16/2021 6:06:50 PM.

The data was not partitioned as training and test data to


maximise the data used as inputs for the models, and the 33
input descriptors (shown in Table 1) were used as inputs. This
methodology enabled the tetrafluoroborate and hexafluoropho-
Fig. 2 MLREM coefficients for each of the 27 descriptors (intercept not sphate anion to be treated as distinct species, though there was
included) used to assess the reaction shown in Scheme 1, using log(k2) as still minimal data present for them. The output log(k2) was
the output, averaged for the three training sets. used since it had led to the best models in the previous section.
For the MLREM model the best results based on the standard
error of estimation were obtained when the model had 19
descriptors (including the intercept). The resulting coefficients
for the cation calculated based on group contributions). are shown in Fig. S4 (ESI†), and it is evident that while there are
The reaction conditions showed that the IL mole fraction had differences compared to those generated using the data parti-
a weak positive contribution, and the temperature a positive tioned into training sets, overall, they are similar. The descrip-
contribution (since the inverse temperature had a negative tors that were removed by the algorithm are identified in
contribution). Table 1. Unfortunately, the inclusion of only two anion descrip-
It is worth briefly commenting on those prominent descrip- tors resulted in the model treating hexafluorophosphate
tors. The reaction conditions are consistent with what would be and bis(trifluoromethanesulfonyl)imide anions as identical.
expected based on the understanding of temperature effects on The cations were uniquely identified by the 14 cation descrip-
reaction, but also observed dependencies of the reaction out- tors with non-zero coefficients. For the BRANNLP model, the
come for the process shown in Scheme 1 on the proportion of coefficients for some descriptors were zero, and hence not
the ionic liquid in the reaction mixture.68 Likewise readily used in the model, and these are again identified in Table 1.
explained, the number of rotatable bonds in the carbon atom In contrast to the MLREM model there were five descriptors
might be considered to indicate conformational flexibility and retained for the anions with non-zero coefficients, which
hence difficulty for the identified key interaction between the enabled the four anions to be uniquely described. Similarly,
cation and pyridine to occur (cf. charge accessibility64,68). Molar the cations were uniquely identified by the 16 cation descrip-
refractivity is a measure of the polarizability of the molecule; tors with non-zero coefficients. Consequently, it was expected
a negative correlation here is consistent with the need for that the BRANNLP would be more reliable in predicting rate
localised charge in order for the key interaction to occur.64,68 constant data.
The origin of the positive parameter is less clear; why having A total of 9504 predictions of log(k2) were made for the
secondary sp3 hybridised carbon atoms would enhance reactiv- reaction shown in Scheme 1 in mixtures containing ionic
ity is not immediately apparent. liquids made up of all the possible binary combinations of
The same set of 28 descriptors were used for the BRANNLP the 33 cations and four anions used in the training set, at mole
and BRANNGP models, again with five removed as they were fractions of ionic liquid in acetonitrile in the range 0.1–0.9 and
not valid for the training sets. The ANN models do not enable at a temperature in the range 260–330 K. This data set included
contributions to be extracted for each descriptor. However, predictions for 132 ILs, which included 36 ILs from the original
across the three, separate training/test sets it was noted that set and 96 new ILs based on the other possible cation–anion
some of the descriptors had a zero weight. This zero weight was combinations. The predicted log(k2) values were then converted
due to the limited data available for some cation and anion to k2. Both models resulted in predicted rate constant data in
features, and hence for some data partitioning certain descrip- the range ca. 10 5–10 1 mol L 1 s 1; that is, over four orders of
tors had minimal variation within the training set. Specifically, magnitude.
for the BRANNLP model with the log(k2) output, there were zero
weights for Hy, nO, N-075, nAB, and AMR for the first indepen- Validation of the model
dent training set, AnionTotalC, Ui, nO, nHAcc and nAB for the The predicted rate constant data for the reaction shown in
second set and Hy, N-075 and AMR for the third set. To some Scheme 1 included the original 36 ILs used, along with predic-
extent, this indicates that these descriptors are more limited in tions for 96 new ILs made up of the constituent anions from the
their variation across the ILs, and implies that the other original set of ILs. A series of 17 ILs (Fig. 3) was chosen,

This journal is © the Owner Societies 2021 Phys. Chem. Chem. Phys., 2021, 23, 2742--2752 | 2747
View Article Online

PCCP Paper
Published on 14 January 2021. Downloaded by University of Toledo on 5/16/2021 6:06:50 PM.

Fig. 3 The ionic liquids 1–17 prepared to evaluate the models developed. For details of nomenclature, see ESI.†

including 11 from these 96 new ILs;§ this set was carefully some small differences between the experimental conditions
selected to cover the range of ions used in the training set. The and the outputs of the models, but these are expected to have
rate constant for the reaction shown in Scheme 1 was deter- little effect on the comparisons made.
mined in mixtures of one of these ionic liquids and acetonitrile, The agreement between the predicted and experimental
with a range of IL mole fractions and temperatures used to test values for this new data set was highly dependent on which
the predictability of the model. Note that the set includes two ions were present in the ionic liquid and varied between
repeats for ionic liquids 9 and 16. The former salt was used to extremely good and quite poor. A summary is provided in Table
see the effect for a case over two different proportions of ionic S7 (ESI†), including the newly determined k2 values, the pre-
liquid in the reaction mixture, while the latter was used to dicted k2 values from MLREM and BRANNLP methods, and
check whether changing the temperature had a significant the percentage difference. A representation of the efficacy of
effect on a poor predictability. The proportions of ionic liquid the models is provided in Fig. 4. It is evident from Table S7
in the reaction mixtures and temperatures used were all close (ESI†) and Fig. 4 that generally the predictions were close to the
to those included in the models, with the exact IL mole fraction, experimental data, and that both models performed well.
temperature and rate constants provided in the ESI.† There are It should be noted that the training set there were only three
ionic liquids that did not contain the bis(trifluoromethanesulfonyl)
imide anion, and these three consisted of the 1-butyl-3-
§ Where considered here, those ionic liquids that made up the initial training set
were considered under different conditions (particularly at different composi-
methylimidazolium cation paired with either dicyanimide,
tions in the reaction mixture and different temperatures) to provide data not hexafluorophosphate or tetrafluoroborate. Consequently, the
available in the training set. influence on the training of these three anions was limited. The

2748 | Phys. Chem. Chem. Phys., 2021, 23, 2742--2752 This journal is © the Owner Societies 2021
View Article Online

Paper PCCP

Overall there is excellent agreement considering the broad


range of cations and anions included in the validation set, the
relatively limited training set and the wide range of rate
constants considered. It is evident that both the MLREM and
BRANNLP have made meaningful predictions of the expected
rate constants for ILs which were not included in the original
study. It is apparent that neither the MLREM nor the BRANNLP
model performed significantly better than the other. While there
were a few outliers, the majority of predictions were well within an
order of magnitude, which is good for organic reactions.
Published on 14 January 2021. Downloaded by University of Toledo on 5/16/2021 6:06:50 PM.

The difficulty in predicting rate constants lies in the depen-


dence on multiple variables, and for this paper these are the
reaction temperature, the nature of the ionic liquid and
the proportion of the ionic liquid in the reaction mixture. While
the use of systematic studies enables some structure–property
relationships to be determined, they are not able to extract the
complex multivariate trends that can be obtained using
machine learning methods. Here we have addressed this lim-
itation in the predictive tools available, and shown that given a
reasonable set of initial data that meaningful predictions can
be made, noting specific issues with certain ion types that may
be a function of the descriptors used and the training data
available. A key advantage of the approach used in this paper is
that it involved experimental data and an assumption free
position for the selection of descriptors. This approach meant
that the models developed were purely based on the experi-
mental data, with no intervention in the selection of descrip-
tors, and no assumptions required for computational models.
Due to the multitude of possible cations and anions which
can be present in ILs, there is a very broad compositional space
which needs to be meaningfully represented. Here our
approach has been to use DRAGON descriptors to describe
the chemical structures of the cations and anions, where a set
Fig. 4 Comparison of the log(k2) values determined experimentally com- of 33 descriptors was able to uniquely characterise all the
pared to those predicted using the (a) MLREM and (b) BRANNLP models. structures. A key advantage of this approach is that it can be
Numbering corresponds to the ionic liquids. The line shown is y = x.
easily extended to any set of ionic liquids through the inclusion
of more descriptors. In addition, the methodology is not
dependent on any physical properties of the ILs, such as
validation set of ILs was deliberately selected to consist of a viscosity and melting point, though these could be included
broad range of anions and cations. as inputs.
From Fig. 4 it is apparent that the majority of the predictions It should be noted that the work described is based on a
had good agreement with the experimental results, and across reaction that is extremely well-described in a range of ionic
all of the ionic liquids considered, using either MLREM or liquids; as such, there is systematic data covering different
BRANNLP. Two ionic liquids, [bmmo][N(CN)2] 14 and [TOA][BF4] conditions (temperature, solvent composition) and different
16, are the exceptions, with both of these ionic liquids having cation and anion types. We would like to emphasise that high
predicted values substantially lower than the experimental values. quality models require consistent sets of data, with data cover-
While the origin of the discrepancy, and the direction of such, is ing the compositional space. The cases where the model
not clear, for these ILs there were limited relevant training data. presented has (relatively) limited predictability correspond to
Not only were the cations for both these ionic liquids only those systems with (relatively) less data in the training set.
paired with the bis(trifluoromethanesulfonyl)imide anion in the Importantly, while the reaction itself is straightforward
training set, but the dicyanimide and tetrafluoroborate anions (a bimolecular substitution involving a neutral amine nucleo-
were only paired with the 1-butyl-3-methylimidazolium cation in phile), the understanding gained from this model might be
the training set. readily extended to series where less information is available
It is worth noting, however, that in contrast, the predictions but the reaction is known to behave similarly in ionic
for other ionic liquids containing these anions (including salts liquids.92–94 That is, there is the potential for broader applic-
6, 8, 10 and 13) all matched the experimental data well. ability, even into systems where less data is available.

This journal is © the Owner Societies 2021 Phys. Chem. Chem. Phys., 2021, 23, 2742--2752 | 2749
View Article Online

PCCP Paper

Conclusions 7 R. R. Hawker, R. S. Haines and J. B. Harper, in Targets in


Heterocyclic Systems, ed. R. Noto, 2014, vol. 18, pp. 141–213.
In this study we have developed multiple linear regression 8 D. Yalcin, C. J. Drummond and T. L. Greaves, Phys. Chem.
models (MLREM) and neural networks (BRANNLP) for the rate Chem. Phys., 2019, 21, 6810.
constant, k2, of the reaction between benzyl bromide and 9 T. L. Greaves, K. Ha, B. W. Muir, S. C. Howard,
pyridine. The reaction has been conducted in mixtures of A. Weerawardena, N. Kirby and C. J. Drummond, Phys.
ionic liquid and acetonitrile at different temperatures, with Chem. Chem. Phys., 2015, 17, 2357–2365.
different ionic liquids, and in different proportions of the two 10 Z. Findrik, G. Megyeri, L. Gubicza, K. Belafi-Bako, N. Nemestothy
co-solvents. An original set of 160 rate constant data from 36 and M. Sudar, J. Cleaner Prod., 2016, 112, 1106–1111.
ILs was used as the inputs. Structure–property relationships 11 A. Zhu, L. Li, C. Zhang, Y. Shen, M. Tang, L. Bai, C. Du,
Published on 14 January 2021. Downloaded by University of Toledo on 5/16/2021 6:06:50 PM.

were obtained through using the MLREM model, and both S. Zhang and J. Wang, Green Chem., 2019, 21, 307–313.
MLREM and BRANNLP showed good agreement for training 12 S. P. F. Costa, B. S. F. Martins, P. Pinto and M. Saraiva,
and test sets. Additionally, predictive models were developed J. Hazard. Mater., 2016, 309, 165–172.
using the full data set, and used to obtain rate constant 13 Z. X. Wang, Y. Mu, Y. L. Wang, Q. M. Bing, T. Su and
predictions for 96 new ILs along with the previous 36 ILs, J. Y. Liu, Solid State Sci., 2017, 64, 76–83.
covering all possible cation and anion combinations of the 14 M. T. Rahman, T. Fukuyama, I. Ryu, K. Suzuki,
original ions used. A set of 17 ionic liquids, 11 not included in K. Yonemura, P. F. Hughes and K. Nokihara, Tetrahedron
the initial training set, were selected to ensure that they covered Lett., 2006, 47, 2703–2706.
the range of ions used in the training set, and prepared where 15 G. Mafra, A. A. Viera, J. Merib, J. L. Anderson and
necessary. The rate constant for the reaction was determined in E. Carasek, Anal. Chim. Acta, 2019, 1063, 159–166.
mixtures containing them – at a particular temperature and at a 16 D. Godfrey, J. H. Bannock, O. Kuzmina, T. Welton and
particular solvent composition with acetonitrile. Overall there T. Albrecht, Green Chem., 2016, 18, 1930–1937.
was good agreement for most of these experimental data with 17 M. Zavrel, D. Bross, M. Funke, J. Buchs and A. C. Spiess,
the predictions from both MLREM and BRANNLP analyses, Bioresour. Technol., 2009, 100, 2580–2587.
noting the limitations in describing ionic liquids containing 18 W. Beckner and J. Pfaendtner, J. Chem. Inf. Model., 2019, 59,
some components. This work has shown that the use of 2617–2625.
machine learning is highly useful for predicting the rate con- 19 K. Karu, A. Ruzanov, H. Ers, V. Ivanistsev, I. Lage-Estebanez
stants of organic reactions in ILs, given sufficient experimental and J. M. G. de la Vega, Computation, 2016, 4, 25.
data to build the models. 20 D. L. Peng, J. A. Zhang, H. Y. Cheng, L. F. Chen and Z. W. Qi,
Chem. Eng. Sci., 2017, 159, 58–68.
21 S. Tan, S. B. Acevedo and E. I. Izgorodina, J. Chem. Phys.,
Conflicts of interest 2017, 146, 064108.
There are no conflicts to declare. 22 A. Barati-Harooni, A. Najafi-Marghmaleki, M. Arabloo and
A. H. Mohammadi, J. Mol. Liq., 2016, 224, 954–964.
23 A. Barati-Harooni, A. Najafi-Marghmaleki and A. H.
Acknowledgements Mohammadi, J. Mol. Liq., 2017, 231, 462–473.
24 M. R. Fatehi, S. Raeissi and D. Mowla, J. Supercrit. Fluids,
KSSM and JBH acknowledge financial support from the Aus-
2014, 95, 60–67.
tralian Research Council Discovery Project Funding Scheme
25 K. Golzar, S. Amjad-Iranagh and H. Modarress, Ind. Eng.
(Project DP180103682). RFBR is grateful to the School of Chem-
Chem. Res., 2014, 53, 7247–7262.
istry, University of New South Wales for a Summer Vacation
26 K. Golzar, H. Modarress and S. Amjad-Iranagh, Int.
Research Scholarship. All data is available at the following
J. Greenhouse Gas Control, 2016, 53, 187–197.
Figshare repository: 10.25439/rmt.12665651.
27 Y. Huang, X. Zhang, Y. Zhao, S. Zeng, H. Dong and S. Zhang,
Phys. Chem. Chem. Phys., 2015, 17, 26918–26929.
Notes and references 28 K. Paduszynski, Ind. Eng. Chem. Res., 2019, 58, 5322–5338.
29 M. R. Fatehi, S. Raeissi and D. Mowla, J. Mol. Liq., 2017, 227,
1 T. Welton, Chem. Rev., 1999, 99, 2071–2083. 309–317.
2 T. Welton, Biophys. Rev., 2018, 10, 691–706. 30 R. Haghbakhsh and S. Raeissi, J. Mol. Liq., 2015, 211, 948–956.
3 M. Hirao, H. Sugimoto and H. Ohno, J. Electrochem. Soc., 31 X. J. Kang, Z. J. Zhao, J. G. Qian and R. M. Afzal, Ind. Eng.
2000, 147, 4168–4172. Chem. Res., 2017, 56, 11344–11351.
4 J. S. Wilkes and M. J. Zaworotko, Chem. Commun., 1992, 32 M. Lashkarbolooki, Sep. Sci. Technol., 2017, 52, 1454–1467.
965–967. 33 J. A. Lazzus, F. Cuturrufo, G. Pulgar-Villarroel, I. Salfate and
5 T. L. Greaves and C. J. Drummond, Chem. Rev., 2008, 108, P. Vega, Ind. Eng. Chem. Res., 2017, 56, 6869–6886.
206–237. 34 S. Atashrouz, H. Mirshekar, A. Hemmati-Sarapardeh,
6 T. L. Greaves and C. J. Drummond, Chem. Rev., 2015, 115, M. K. Moraveji and B. Nasernejad, Korean J. Chem. Eng.,
11379–11448. 2017, 34, 425–439.

2750 | Phys. Chem. Chem. Phys., 2021, 23, 2742--2752 This journal is © the Owner Societies 2021
View Article Online

Paper PCCP

35 P. Diaz-Rodriguez, J. C. Cancilla, G. Matute and 64 E. E. L. Tanner, H. M. Yau, R. R. Hawker, A. K. Croft and


J. S. Torrecilla, J. Ind. Eng. Chem., 2015, 21, 1350–1353. J. B. Harper, Org. Biomol. Chem., 2013, 11, 6170–6175.
36 A. Najafi-Marghmaleki, M. R. Khosravi-Nikou and A. Barati- 65 S. T. Keaveney, D. V. Francis, W. N. Cao, R. S. Haines and
Harooni, J. Mol. Liq., 2016, 220, 232–237. J. B. Harper, Aust. J. Chem., 2015, 68, 31–35.
37 P. Diaz-Rodriguez, J. C. Cancilla, K. Wierzchos and 66 R. R. Hawker, J. Panchompoo, L. Aldous and J. B. Harper,
J. S. Torrecilla, Sens. Actuators, B, 2015, 206, 139–145. ChemPlusChem, 2016, 81, 574–583.
38 J. C. Cancilla, P. Diaz-Rodriguez, G. Matute and J. S. Torrecilla, 67 K. S. Schaffarczyk McHale, R. R. Hawker and J. B. Harper,
Phys. Chem. Chem. Phys., 2015, 17, 4533–4537. New J. Chem., 2016, 40, 7437–7444.
39 M. Hosseinzadeh and A. Hemmati-Sarapardeh, J. Mol. Liq., 68 R. R. Hawker, R. S. Haines and J. B. Harper, Chem. Com-
2014, 200, 340–348. mun., 2018, 54, 2296–2299.
Published on 14 January 2021. Downloaded by University of Toledo on 5/16/2021 6:06:50 PM.

40 M. Mesbah, E. Soroush and M. R. Kakroudi, J. Mol. Liq., 69 A. Schindl, R. R. Hawker, K. S. S. McHale, K. T.-C. Liu,
2017, 225, 778–787. D. C. Morris, A. Y. Hsieh, A. Gilbert, S. W. Prescott,
41 L. D. Cao, P. Zhu, Y. S. Zhao and J. H. Zhao, J. Hazard. R. S. Haines, A. K. Croft, J. B. Harper and C. M. Jäger, Phys.
Mater., 2018, 352, 17–26. Chem. Chem. Phys., 2020, 22, 23009–23018.
42 S. Y. Ma, M. Lv, X. Y. Zhang, H. L. Zhai and W. J. Lv, 70 W. Beker, E. P. Gajewska, T. Badowski and B. A. Grzybowski,
Chemom. Intell. Lab. Syst., 2015, 144, 138–147. Angew. Chem., Int. Ed., 2019, 58, 4515–4519.
43 H. R. Amedi, A. Baghban and M. A. Ahmadi, J. Mol. Liq., 71 C. W. Coley, R. Barzilay, T. S. Jaakkola, W. H. Green and
2016, 216, 411–422. K. F. Jensen, ACS Cent. Sci., 2017, 3, 434–443.
44 A. Baghban, A. H. Mohammadi and M. S. Taleghani, Int. 72 G. Fic and G. Nowak, Chemom. Intell. Lab. Syst., 2005, 75,
J. Greenhouse Gas Control, 2017, 58, 19–41. 137–148.
45 A. R. Bahmani, F. Sabzi and M. Bahmani, J. Mol. Liq., 2015, 73 D. Fooshee, A. Mood, E. Gutman, M. Tavakoli, G. Urban,
211, 395–400. F. Liu, N. Huynh, D. V. Vranken and P. Baldi, Mol. Syst. Des.
46 M. Fattahi, H. Abedini, A. Baghban and M. A. Anbaz, Pet. Eng., 2018, 3, 442–452.
Sci. Technol., 2017, 35, 1117–1123. 74 B. Maryasin, P. Marquetand and N. Maulide, Angew. Chem.,
47 C. A. Faundez, E. N. Fierro and J. O. Valderrama, J. Environ. Int. Ed., 2018, 57, 6978–6980.
Chem. Eng., 2016, 4, 211–218. 75 J. M. Granda, L. Donina, V. Dragone, D. L. Long and
48 M. M. Ghiasi and A. H. Mohammadi, J. Mol. Liq., 2017, 242, L. Cronin, Nature, 2018, 559, 377–381.
594–605. 76 G. Skoraczynski, P. Dittwald, B. Miasojedow, S. Szymkuc,
49 M. E. Hamzehie and H. Najibi, J. Nat. Gas Sci. Eng., 2016, 29, E. P. Gajewska, B. A. Grzybowski and A. Gambin, Sci. Rep.,
252–263. 2017, 7, 3582.
50 M. Mesbah, S. Shahsavari, E. Soroush, N. Rahaei and 77 H. Streubing, Z. Ganase, P. G. Karamertzanis, E. Siougkrou,
M. Rezakazemi, J. CO2 Util., 2018, 25, 99–107. P. Haycock, P. M. Piccione, A. Armstrong, A. Galindo and
51 M. A. Sedghamiz, A. Rasoolzadeh and M. R. Rahimpour, C. S. Adjiman, Nat. Chem., 2013, 5, 952–957.
J. CO2 Util., 2015, 9, 39–47. 78 N. D. Austin, N. V. Sahinids, I. A. Konstantinov and
52 C. Cortes and V. Vapnik, Mach. Learn., 1995, 20, 273. D. W. Trahan, AIChE J., 2018, 64, 104–122.
53 M. E. Tipping, J. Mach. Learn. Res., 2001, 1, 211–214. 79 C. Gertig, L. Kröger, L. Fleitmann, J. Scheffczyk, A. Bardow and
54 T. K. Ho, Proceedings of the 3rd International Conference K. Leonhard, Ind. Eng. Chem. Res., 2019, 58, 22835–22846.
on Document Analysis and Recognition, Montreal, QC, 80 E. I. Izgorodina, Z. L. Seeger, D. L. A. Scarborough and
14–16 August, 1995, pp. 278–282. S. Y. S. Tan, Chem. Rev., 2017, 117, 6696–6754.
55 T. Le, C. Epa, F. R. Burden and D. A. Winkler, Chem. Rev., 81 K. Goloviznina, J. N. Canongia Lopes, M. Costa Gomes and
2012, 112, 2889–2919. A. A. H. Pádua, J. Chem. Theory Comput., 2019, 15,
56 J. Schmidt, M. R. G. Marques, S. Botti and M. A. L. Marques, 5858–5871.
npj Comput. Mater., 2019, 5, 83. 82 S. R. L. Talete, Dragon for Windows, 2007.
57 Y. Liu, T. Zhao, W. Ju and S. Shi, J. Materiomics, 2017, 3, 159–177. 83 P. W. Atkins, J. D. Paula and J. Keeler, Atkins’ Physical
58 J. Wei, X. Chiu, S.-Y. Sun, K. Xu, H.-X. Deng, J. Chen, Z. Wei Chemistry, Oxford University Press, Oxford, 11th edn, 2017.
and M. Lei, InfoMat, 2019, 1, 338–358. 84 R. Todeschini, V. Consonni, R. Mannhold, H. Kubinyi and
59 D. C. Elston, Z. Boukouvalals, M. S. Butrico, M. D. Fuge and H. Timmerman, Handbook of Molecular Descriptors, Wiley
P. W. Chung, Sci. Rep., 2018, 8, 9059. VCH, Weinheim, 2000.
60 D. Yalcin, T. C. Le, C. J. Drummond and T. L. Greaves, 85 F. R. Burden and D. A. Winkler, J. Chem. Inf. Comput. Sci.,
J. Phys. Chem. B, 2019, 123, 4085–4097. 1999, 39, 236–242.
61 H. M. Yau, A. G. Howe, J. M. Hook, A. K. Croft and 86 F. R. Burden and D. A. Winkler, J. Med. Chem., 1999, 42,
J. B. Harper, Org. Biomol. Chem., 2009, 7, 3572–3575. 3183–3187.
62 H. M. Yau, S. J. Chan, S. R. D. George, J. M. Hook, A. K. Croft 87 D. A. Winkler and F. R. Burden, Mol. Simul., 1999, 24,
and J. B. Harper, Molecules, 2009, 14, 2521–2534. 243–258.
63 H. M. Yau, A. K. Croft and J. B. Harper, Faraday Discuss., 88 M. A. T. Figueiredo, IEEE Transactions on Pattern Analysis
2012, 154, 365–371. and Machine Intelligence, 2003, 25, 1150–1159.

This journal is © the Owner Societies 2021 Phys. Chem. Chem. Phys., 2021, 23, 2742--2752 | 2751
View Article Online

PCCP Paper

89 F. R. Burden and D. A. Winkler, QSAR Comb. Sci., 2009, 28, 93 S. T. Keaveney, R. S. Haines and J. B. Harper, Org. Biomol.
645–653. Chem., 2015, 13, 3771–3780.
90 F. R. Burden and D. A. Winkler, QSAR Comb. Sci., 2009, 28, 94 S. T. Keaveney, R. S. Haines and J. B. Harper, Org. Biomol.
1092–1097. Chem., 2015, 13, 8925–8936.
91 W. L. F. Armarego and C. L. L. Chai, Purification of Labora- 95 S. T. Keaveney, B. P. White, R. S. Haines and J. B. Harper,
tory Chemicals, Butterworth-Heinemann, Oxford, 2013. Org. Biomol. Chem., 2016, 14, 2572–2580.
92 S. T. Keaveney, K. S. Schaffarczyk McHale, R. S. Haines and 96 S. T. Keaveney and J. B. Harper, RSC Adv., 2013, 3,
J. B. Harper, Org. Biomol. Chem., 2014, 12, 7092–7099. 15698–15704.
Published on 14 January 2021. Downloaded by University of Toledo on 5/16/2021 6:06:50 PM.

2752 | Phys. Chem. Chem. Phys., 2021, 23, 2742--2752 This journal is © the Owner Societies 2021

You might also like