Angewandte

Research Articles Chemie

www.angewandte.org

How to cite: Angew. Chem. Int. Ed. 2024, 63, e202311526

Flow Chemistry doi.org/10.1002/anie.202311526

Revisiting the Paradigm of Reaction Optimization in Flow with a

Priori Computational Reaction Intelligence
Pauline Bianchi and Jean-Christophe M. Monbaliu*
Dedicated to Professor Klavs F. Jensen on the occasion of his 70th birthday

their imagination has extended, feeding new ways to access

Abstract: The use of micro/meso-fluidic reactors has molecular architectures and cover emerging needs. Count-
resulted in both new scenarios for chemistry and new less iterations and disruptive advances however still con-
requirements for chemists. Through flow chemistry, verge to a very pragmatic conclusion: developing new
large-scale reactions can be performed in drastically synthetic routes is resource-, time- and waste-intensive. A
reduced reactor sizes and reaction times. This obvious fundamental shift in how chemists approach their work is
advantage comes with the concomitant challenge of re- necessary.
designing long-established batch processes to fit these Flow technology is one of the most impactful paradigm
new conditions. The reliance on experimental trial-and- changes in the modern history of organic chemistry, offering
error to perform this translation frequently makes flow superior process control, reduced waste generation, and
chemistry unaffordable, thwarting initial aspirations to enhanced safety from R&D to production scale.[1–3] These
revolutionize chemistry. By combining computational advantages, which are obtained through the use of micro/
chemistry and machine learning, we have developed a mesofluidic reactors (MFRs) can be further enhanced with
model that provides predictive power tailored specifi- the integration of artificial intelligence,[4,5] process analytical
cally to flow reactions. We show its applications to technology (PAT) tools,[6] and automation.[7,8] This makes
translate batch to flow, to provide mechanistic insight, to flow technology particularly adapted to robust and versatile
contribute reagent descriptors, and to synthesize a production systems,[9,10] and yet, breaking away from
library of novel compounds in excellent yields after centuries of macroscopic batch synthesis has remained
executing a single set of conditions. difficult.[11] There are mainly three aspects that keep making
reactions challenging to perform with MFRs: slow kinetics
which imposes onerous reactor size requirements, viscous
liquids which are difficult to pump, or solid particles which
Introduction can clog the reactor channel. The last two issues have been
addressed by designing MFR configurations with specific
Organic chemistry has accompanied and sustained the technical features. Slow kinetics, on the other hand,
development of fuels, therapies, and advanced materials represent a particular challenge considering that flow
throughout human history. Several technological revolutions reactions typically require a timeframe below 15 min to be
later, the tools and overall context have changed: while scalable. Traditional chemistry has not been developed to
robots and automated systems are being increasingly afford high conversions within such a narrow window nor
implemented in labs and on production lines, chemists does it indeed have the tools to reach them (Figure 1A).[11]
remain pivotal assets behind innovation. The horizon of Flow reactions are typically accelerated by using process
intensification techniques, such as high reaction concentra-
tions and/or increased temperatures.[12] The development of
[*] P. Bianchi, Prof. Dr. J.-C. M. Monbaliu
Center for Integrated Technology and Organic Synthesis (CiTOS), such intensified conditions has, until now, depended on trial-
MolSys Research Unit, University of Liège and-error chemistry, occasionally aided by a design-of-
B6a, Room 3/19, Allée du Six Août 13, 4000 Liège (SartTilman) experiments approach or machine learning (ML) such as
(Belgium) Bayesian Optimization models,[13] to increase the efficiency
E-mail: [email protected] of the exploration. Predictive models based on mining data
Homepage: https://www.citos.uliege.be from databases such as Reaxys, while successful when
Prof. Dr. J.-C. M. Monbaliu applied to batch protocols, have limited use for flow
WEL Research Institute chemistry due to the relatively small number of reactions
Avenue Pasteur 6, 1300 Wavre (Belgium)
that have been reported in MFRs.[14] Other types of
© 2023 The Authors. Angewandte Chemie International Edition predictive ML models based on Structure-Reactivity Rela-
published by Wiley-VCH GmbH. This is an open access article under
tionships (SReARs) to relate chemical structure to a desired
the terms of the Creative Commons Attribution Non-Commercial
NoDerivs License, which permits use and distribution in any med- output (e.g., function or enantiomeric excess) have been
ium, provided the original work is properly cited, the use is non- developed,[15] yet only in the context of batch reactions.
commercial and no modifications or adaptations are made. Despite the great insight they provide, the dependence of

Angew. Chem. Int. Ed. 2024, 63, e202311526 (1 of 9) © 2023 The Authors. Angewandte Chemie International Edition published by Wiley-VCH GmbH
 15213773, 2024, 5, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/anie.202311526 by University Of Massachusetts Dartmouth, Wiley Online Library on [03/12/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
Angewandte
Research Articles Chemie

Figure 1. Overview of the interdisciplinary approach for revisiting the paradigm of reaction optimization. (A) One of the main challenges in
translating from batch to flow conditions: batch chemistry was developed relying on extended reaction times whereas flow requires process
intensification. (B) Our rationale for the adoption of quantum chemistry and machine learning to access rational and intelligent flow processes.
(C) Case study: electrophilic amination between nitrosoarenes 1 and silyl enol ethers 2 toward hydroxylaminated ketones 3. Compounds 3 are
valuable aminated building blocks. (D) Representative examples of APIs potentially synthesizable through electrophilic aminations with
nitrosoarenes.

these models on experimental data represents a greater chemical reactions, a kinetic constant (k) can then be related
challenge for a young field such as flow chemistry where to the rate and concentration of the reaction. In other words,
very little knowledge exists to even guide initial experi- if ΔG‡ can be accessed with high accuracy, Erying’s equation
ments. If we are to catch up with centuries of accumulated becomes an invaluable gateway to predict a reaction time, at
batch chemistry knowledge, the reliance of predictive tools any given temperature and concentration.
on experimental data needs to be lessened.
While searching for less experiment-dependent ways to �
kkB T DG6¼

predict feasibility in flow, we became interested in the k¼ e RT

(1)
h
mathematical relationship described between kinetics (k),
temperature (T) and the activation barrier (ΔG‡) of a These observations led us to develop a quantum
reaction afforded by Erying’s equation [1] (with k = reaction chemistry-guided assistant, a prototype of which was
rate constant, k = transmission coefficient, H = Planck’s con- successfully applied to the development on a large scale of
stant (J s), kB = Boltzmann constant (J K 1), T = temperature an intensified flow synthesis of a contraceptive
(K), R = gas constant (cal K 1 mol 1), ΔG‡ = activation bar- pharmaceutical.[16] This prototype used DFT to provide an
rier (cal mol 1)). Considering the rate laws of non-zero order ideal set of reagents, residence times, and temperatures to

Angew. Chem. Int. Ed. 2024, 63, e202311526 (2 of 9) © 2023 The Authors. Angewandte Chemie International Edition published by Wiley-VCH GmbH
 15213773, 2024, 5, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/anie.202311526 by University Of Massachusetts Dartmouth, Wiley Online Library on [03/12/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
Angewandte
Research Articles Chemie

reach the target conversion, therefore minimizing waste, SReARs were explored.[15,26] Fourteen different reagent
spending, and exposure to highly active hormonal substan- properties for each of 25 nitrosoarenes 1 (Figure 2A) and 25
ces. Despite its effectiveness, the method had two main silyl enol ethers 2 (Figure 2A) were computed and analyzed
drawbacks: suboptimal accuracy of the estimated conditions to identify feature intercorrelations. The size of the data set
and taxing computational requirements to access high and its inherent features were selected based on previous
quality ΔG‡ information. Therefore, a more practical works.[27–29] This resulted in the selection of two properties
solution was sought (Figure 1B). to be used as features for a ML training set, specifically the
The use of a relative logarithm (log(k/k0)), comparing a nitrogen charge and the B5 Sterimol steric parameter for
reaction of interest (k) to the kinetics of a reference reaction nitrosoarenes 1, and the C = C stretching and the B1 Sterimol
(k0), was proposed as a first step to increase the accuracy of steric parameter for silyl enol ethers 2 (Figure 2B). The ΔG‡
our estimates. Aside from providing data normalization and values for the reactions between these same 25 nitro-
variance stabilization, log(k/k0) can provide immediate soarenes 1 (Figure 2A) with 2 a and for the 25 silyl enol
intuitive insight about how fast or slow a reaction is, ethers 2 with 1 a were computed to calculate log(k/k0). We
compared to the model reaction. Reactions involving nitro- then searched for a ML model to relate log(k/k0) to the
soarene electrophiles 1 (Figure 1C) were selected to test this selected features. Model generalization was assessed with k-
approach. The choice was based on their potent nitrogen- fold cross-validation (k = 5). A total of four ML models
transfer capacities which unfortunately remain underex- were evaluated, namely, Least Absolute Shrinkage and
ploited due to their capricious reactivity.[17–19] Developing a Selection Operator (LASSO), Ridge, K-Nearest Neighbors
protocol for the application of 1 would provide a precious (KNN) and Random Forest (RF) regressions (Supporting
tool for the preparation of pharmaceutically relevant Information, see Sections 6.1. and 6.2.). Among the four, the
aminated derivatives (Figure 1D). Ridge regression afforded the clearest correlation for both
the nitrosoarene 1 (R2train = 0.93) and the silyl enol ether 2
training sets (R2train = 0.93) (Figure 2C).
Results and Discussion Having trained our system to determine log(k/k0) using
only in silico reagent properties, an experimental test
A Roadmap to Flow procedure was performed. The trained ML-model was used
to determine optimal temperatures to experimentally access
The reaction between nitrosobenzene (1 a) and unsubsti- ΔG‡ for the reactions between five novel nitrosoarenes (1 c–
tuted silyl enol ether (2 a) results in the highly selective g) with 2 a, and between five silyl enol ethers (2 b–f) with 1 a
formation of the corresponding N-addition product (3 a) and (Figure 2C). This was done by classifying reactions into 5
was chosen as the model reaction with kinetic constant k0.[20] categories according to their computationally derived log(k/
The analogous reaction between 2-nitrosotoluene (1 b) and k0) value. Each of the 5 categories suggested a temperature
2 a was used to define a first relative constant log(k1b/k0). and residence time range at which relevant kinetic informa-
Extensive computational and experimental probing of these tion could be obtained (Figure 2C, table). Category 5
two reactions was performed to show the robust accuracy of reactions, for which no reactivity was predicted using MFRs,
this metric (Supporting Information, see Section 7.1.). were also recorded as negative data.[30,31] This experimental
Current records for precise ΔG‡ determination feature a guidance is particularly relevant to design kinetics experi-
mean absolute error (MAE) between 0.5 and 1 kcal mol 1 ments in MFRs where, contrary to batch procedures, time is
which translates to a ~ 5-fold change in the predicted defined by the MFR setup, meaning kinetics cannot be
reaction time.[21–23] This means that an estimated residence performed simply by monitoring the reaction over time, as it
time of 5 min could take place anywhere between 1 to is typically done for batch kinetics.
25 min, and therefore that the margin of error exceeds the The experiments were performed using an automated
narrow time window for optimal flow conditions (i.e., flow setup inspired by Jensen et al.[32] Our setup incorpo-
< 15 min). Twenty computational methods to determine rated real-time monitoring of temperature, pressure and
ΔG‡ were assessed, from which a combination of Minnesota conversion in four compact units: (1) a dosing unit including
functionals[24] with corrective factors for concentration and feed solutions and syringe pumps; (2) a reactor unit
quasi-harmonic oscillators,[25] provided values that were combining a proportional-integral-derivative (PID) temper-
within 0.3 kcal mol 1 of experimental ΔG‡ values, reducing ature controller and a stainless steel coil reactor; (3) a
the uncertainty for predicted time to ~ 1.5-fold (Supporting downstream analytical unit featuring an in-line IR spectrom-
Information, see Section 5.3.). The method was translated eter and (4) a DIY control unit relying on LabVIEWTM
into a DIY LabVIEWTM software (SnapPy) to automatically (enFLOW) to coordinate all auxiliaries and execute instruc-
extract corrected ΔG‡ and kinetic data from output files tions from an Excel sheet.
generated by a Quantum Mechanics software. The data from the test set was plotted into the Ridge
To be applicable to a library of compounds, the time regressions which provided a clear correlation for the
required to predict log(k/k0) needed to be reduced from reactions varying the nitrosoarene (1 c–g) (R2test = 0.96) and
hours to a few minutes. The need for a method to determine the reactions varying the silyl enol ethers (2 b–f) (R2test
log(k/k0) without having to compute ΔG‡ was therefore = 0.96, Figure 2B). The ΔG‡ values extracted from our ML
evident. Instead, easily computed reagent properties (e.g., model afforded a MAE of � 0.2 kcal mol 1 and a root-mean-
electronic, steric, or vibrational) to find log(k/k0) through square error (RMSE) of � 0.3 kcal mol 1 in comparison to

Angew. Chem. Int. Ed. 2024, 63, e202311526 (3 of 9) © 2023 The Authors. Angewandte Chemie International Edition published by Wiley-VCH GmbH
 15213773, 2024, 5, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/anie.202311526 by University Of Massachusetts Dartmouth, Wiley Online Library on [03/12/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
Angewandte
Research Articles Chemie

Figure 2. Development of the QM/ML predictive model. (A) Presentation of the various sets used. Colors represent the category for the kinetics in
flow (green: cat. 1, blue: cat. 2, orange: cat. 3; purple: cat. 4). (B) Ridge regressions to describe nitrosoarenes 1 (left, drawn orbital: LUMO) or silyl
enol ethers 2 (right, drawn orbital: HOMO) for both training and test sets. (C) Overview of the experimental testing with the presentation of
discriminating categories based on log(k/k0). Each category suggested optimum reaction conditions for the kinetics studies.

the experimental values (Supporting Information, see Sec- Extension to Mayr’s Database
tion 7.2.).
This ML model was used to contribute to the reaction
prediction scale developed by Mayr which predicts a
reaction rate using experimentally derived reactivity param-

Angew. Chem. Int. Ed. 2024, 63, e202311526 (4 of 9) © 2023 The Authors. Angewandte Chemie International Edition published by Wiley-VCH GmbH
 15213773, 2024, 5, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/anie.202311526 by University Of Massachusetts Dartmouth, Wiley Online Library on [03/12/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
Angewandte
Research Articles Chemie

eters for the electrophile (E) and the nucleophile (sN and N) To add nitrosoarene electrophiles 1 to the database, the
according to equation [2].[33–35] reactivity parameters (sN and N) associated with our
reference nucleophile (2 a) needed to be obtained in
logðk20 � C Þ ¼ sN ðE þ NÞ (2) acetonitrile. This was performed by determining the kinetics
of the reactions between silyl enol ether 2 a and four
azodicarboxylates (4, Figure 3A) for which the reactivity

Figure 3. Applications of this work in Physical Organic Chemistry. (A) Electrophilicity Mayr indexes determined for nitrosoarenes (1), as compared
to azodicarboxylates (4). (B) Linear free-energy relationships for nitrosoarenes (1, left) and silyl enol ethers (2, right) based on the electrophilic
substituent constants σp + from Brown.[40] Purple values highlight predicted σp + from unreported substituents. (C) Molecular electrostatic potential
(MEP) surfaces of various nitrosoarenes (1) and silyl enol ethers (2) ranked based on the electrophilic substituent constants σp + from Brown[40] (in
red: negative potential, in blue: positive potential).

Angew. Chem. Int. Ed. 2024, 63, e202311526 (5 of 9) © 2023 The Authors. Angewandte Chemie International Edition published by Wiley-VCH GmbH
 15213773, 2024, 5, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/anie.202311526 by University Of Massachusetts Dartmouth, Wiley Online Library on [03/12/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
Angewandte
Research Articles Chemie

parameter E was reported.[36] The nucleophilicity parameters conversion at a given concentration and temperature for a
for 2 a allowed to calculate the E parameter of each combination of reagents 1,2. We have expanded the use of
nitrosoarene 1 a–g using our kinetics data. The E parameters log(k/k0)-based categories developed for the testing of the
obtained were in the range of 8 to 11, comparable to the ML model (Figure 2C, table) to rank reactions based on
values of azodicarboxylates (Supporting Information, see their efficiency in a MFR. We propose to use the log(k/k0)
Section 7.3.). value to sort reactions in five categories based on the
According to Mayr’s categories, their reactions with silyl predicted residence time. Firstly, because categories depend
enol ethers such as 2 a are correctly classified as occurring on reaction conditions, a temperature and concentration
under thermal activation,[37,38] except for 1 g which reacts need to be defined before a reaction can be sorted. From
spontaneously (Figure 3A). Aside from contributing the there, reactions can be considered as follows: Category 1
experimentally obtained E parameters for a new family of (residence time < 1 min) for reactions with outstanding
electrophiles to Mayr’s database, we offer a ML model that potential in MFR; category 2 (residence time of 1–5 min)
could be used in the future to provide estimates for indicates a high potential in MFR; category 3 reactions
substituted nitrosoarenes of choice.[39] (residence time of 5–15 min) are considered to have medium
potential; and category 4 (residence time of 15 min) have a
low potential to be applied in MFR. Category 5 reactions
Application toward LFERs require more than 15 min under the chosen conditions and
are considered unsuitable for flow. Considering the proper-
Using the ML model, Linear Free-Energy Relationships ties of our system, 150 °C and 0.50 M (or 0.25 when required
(LFER) analysis of the reaction mechanism could be for solubility) have been chosen as reference conditions to
performed without any additional experiments. The com- classify our reactions (Figure 4B).
puted log(k/k0) values for 15 nitrosoarenes 1 and silyl enol Using the ML model, the log(k/k0) for the reactions of
ethers 2 substituted at the para position were plotted against three novel nitrosoarenes 1 h–j with 2 a, and for the reactions
the corresponding electrophilic substituent constant (σp +) of three new silyl enol ethers (2 g–i) with 1 a were calculated.
reported by Brown (Figure 3B).[40] These values, along with the log(k/k0) of the reagents
When the molecular probe is borne by nitrosoarene 1, studied thus far (1 a–g and 2 a–f), were used to assign a
the rate of the electrophilic amination is accelerated with category and to calculate the residence time needed to
increasingly powerful electron withdrawing groups (EWG). obtain a target conversion (tconv.) between 85 or 95 % at
This reaction acceleration, indicated by a positive reaction 150 °C (5 bars). These instructions were then transferred to
constant (ρ), reflects the incoming electron flow at the the automated flow platform (Figure 4A), which successfully
transition state. A ρ value of + 3.6 highlights the high synthesized a total of 18 new compounds (Figure 4C) after
sensitivity of the reactivity of nitrosoarenes to electronic testing only one set of conditions per reaction (Supporting
modulation. On the other hand, when the molecular probe Information, see Section 7.4.). Noticeably, these conditions
is located on silyl enol ether 2, two trends are observed. were non-generic, each of them being tailored to fit the
Initially, EWGs moderately deter the reaction due to the reactivity of each different compound.
decreasing electron density of the nucleophile (ρ =-0.59). The predicted conditions led to experimental conver-
However, when strong EWGs are used (i.e. 4-CF3, 4-CN, 4- sions (expconv) that matched the targeted conversions with
NO2), the trend inverts and a ρ value of + 3.33 is obtained. high fidelity (MAE ~ 3 % and RMSE ~ 4 %), with the
Such a drastic change in behavior reflects the ambivalent exception of compounds 3 c and 3 m. For the latter, an azoxy
nature of both nitrosoarenes 1 and silyl enol ethers 2.[17,41] side-product 5 c,m (Figure 4A) was isolated from the reactor
The incorporation of strong EWGs in 2 results in an inverse effluent, which indicated a hydride transfer from the
electronic demand by which 2 behaves as the electrophile aminated product 3 c,m to the dimer 1 c,m as a side-
and 1 as the nucleophile. To further confirm this finding, the reaction.[43] Increasing the excess of silyl enol ether to 3
molecular electrostatic potential (MEP) has been computed, equivalents reinstated the expected conversion.
clearly showing a change in the electrophilic and nucleo- When high selectivity was obtained, the pure product
philic centers (Figure 3C). The broader dispersion of the could be isolated after evaporation of the effluent under
scatter plot of 2 emphasizes a higher sensitivity towards reduced pressure. For reactions requiring purification, the
steric hindrance, rather than the electronic factors described isolation of the product was sometimes problematic due to
by Brown’s electrophilic constants. This observation agrees its instability. Nine aminated products were isolated in
with the main descriptor identified for the ML model, νc = c, excellent yields (3 a,d–h, 86–91 %) and seven in moderate
which gathers both electronic and steric information.[42] The isolated yields (3 b,c,i,k,l,o,q, 60–79 %), while products 3 m
ML model can also be used to predict unreported σp +, as (23 %) and 3 r (42 %) were isolated in lower amounts as a
illustrated for compounds 1 f,j and 2 e (Figure 3B). result of TMS deprotection or degradation into benzoic
acid.[44] Predictions were not impacted by these side-
reactions since they occurred after the electrophilic amina-
Predicting optimum reaction conditions tion (Figure 4A).
Because substituent effects are additive, our model is not
A particularly powerful application of our ML model is the limited to predicting reactions using necessarily either 1 a or
capacity to predict the time needed to reach a target 2 a. When considering a reaction between both a new

Angew. Chem. Int. Ed. 2024, 63, e202311526 (6 of 9) © 2023 The Authors. Angewandte Chemie International Edition published by Wiley-VCH GmbH
 15213773, 2024, 5, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/anie.202311526 by University Of Massachusetts Dartmouth, Wiley Online Library on [03/12/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
Angewandte
Research Articles Chemie

Figure 4. Applications of this work in synthetic organic chemistry. (A) Overview of the flow setup for the synthesis of silylated α-hydroxylaminated
ketones 3 and description of the side-products. (B) Illustration of a predictive model to perform the reactions at 0.5 M to obtain 95 % conversion.
Black lines represent combinations of temperature and log(k/k0) to reach 95 % conversion within 1, 5 or 15 min of residence time. (C) Products
synthesized using conditions predicted by the ML model. Conversions and yields were determined by off-line HPLC analyses using calibration
curves at the maximal absorption of each species (DAD detector). Predicted conversions: [tconv.]—Experimental conversions: expconv.—Yields:
(yield). Values of log(k/k0) are written next to the molecule. Colors represent the category (green: cat. 1—outstanding, blue: cat. 2—high, orange:
cat. 3—medium; purple: cat. 4—low).

nitrosoarene 1 k and new silyl enol ether 2 j (Figure 4C, 3 s), same model reaction k0 (i.e., log(k1k/k0) + log(k2j/k0)).
the relative rate constant log(k1k2j/k0) can be obtained by Nitrosoarene 1 k is predicted to be an outstanding electro-
adding the predicted values for 1 k and of 2 j compared to a phile, with a log(k1k/k0) value of + 2.0 (category 1). Silyl enol

Angew. Chem. Int. Ed. 2024, 63, e202311526 (7 of 9) © 2023 The Authors. Angewandte Chemie International Edition published by Wiley-VCH GmbH
 15213773, 2024, 5, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/anie.202311526 by University Of Massachusetts Dartmouth, Wiley Online Library on [03/12/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
Angewandte
Research Articles Chemie

ether 2 j is a poorly reactive nucleophile, with a predicted meant to serve, in a way, as a chemical translator from batch
log(k1k/k0) of 1.3 (category 4). Their combination predicted to flow. We expect that expediting the development of flow
a category 2 reaction with a log(k(1k2j)/k0) of 0.7, indicating an chemistry will have further repercussions in industrial
overall high potential in flow. The predicted conversion reshoring efforts, by providing safer and greener methods in
(tconv: 95 %) was successfully obtained experimentally accordance with modern regulations.
(expconv: 96 %). If more than one process needs to be evaluated, e.g.,
when constructing chemical libraries, the ML model
becomes a door towards rational synthesis development.
Conclusion Because the ML model provides reactivity predictions as
well as mechanistic insight solely through computational
When chemists optimize reactions in MFRs, they typically data, it reduces the need for experimentation. This is
either feed upon preliminary batch experiments (prior art or showcased by the successful use of categories based purely
trials), trials and errors or guided trials (DoE) in available on quantum mechanics and machine learning to predict
flow systems. This optimization process is time-, resource- optimal conditions to perform the testing. Although test sets
consuming, and produces ample waste. This is particularly are an important part of ML methods, the strength of our
concerning when the availability of resources is scarce, when approach lies in the construction of a model with predictive
toxic/high activity or unstable compounds are involved. We power based purely on computational data. In other words,
have developed a different approach. A ML–DFT model reactions and conditions for which no previous knowledge
was constructed for the expedient computation of an in exists, can be preliminarily modeled without experimenta-
silico kinetic metric, log(k/k0), which predicts with high tion. The ML model becomes therefore a scout for the vast
fidelity experimental data (MAE � 0.2 kcal mol 1 and and new chemical space that flow chemistry provides.
RMSE � 0.3 kcal mol 1). This tool was successfully applied
to nitrosoarene electrophilic aminations for which no prior
experimental data exists. In addition to drastically accelerat-
Acknowledgements
ing optimization, this tool also offered a powerful gateway
to incorporate data into known Physical Organic Chemistry We thank Prof. Raphaël Robiette (Université catholique de
databases (Mayr’s reactivity scale and LFER). The model Louvain, Belgium) for the stimulating discussions on
here reported is based on the properties of the reagents and computational chemistry; Prof. Kristof Van Hecke (Ghent
therefore is limited to exothermic reactions in which the University, Belgium) for the X-ray diffraction on samples 3 a
transition state is closer in energy and structure to the and 1 g; Laurent Collard (Université catholique de Louvain,
reagents. However, the construction of an analogous model Belgium) for the HRMS analyses on compounds 1 g, 1 h, 1 i,
applied to endothermic reactions could be envisaged, as well 2 e, 2 f, 2 h, and 3 a–s; Michaël Schmitz (CiTOS, University of
as models accommodating other factors excluded here such Liège, Belgium) and Dr. Martin Cattoen (CiTOS, University
as catalyzed processes or non-homogeneous conditions. of Liège, Belgium) for elements of the automated system
Complementary tools to analyze competitive pathways and design and the flow platform. Dr. Diana Silva (CiTOS,
to suggest corrective actions should also be developed in the University of Liège, Belgium), Dr. Yi-Hsuan Tsai (CiTOS,
future. University of Liège, Belgium) are acknowledged for helpful
The computation power required to develop this type of discussions related to the preparation of the manuscript and
model is available to many chemistry labs, facilitating its for the proofreading of the manuscript. We also thank
widespread adoption. To assist any person wishing to Sébastien Schoenmaeckers (ISLV, University of Liège,
explore this approach, we have included a step-by-step Belgium) for the proofreading of the manuscript. This work
tutorial in our Supporting Information (see Supporting is supported by the Walloon Region as part of the funding
Information, Scheme S1). Indeed, the growing accessibility for the FRFS-WEL-T strategic axis. The authors acknowl-
of quantum mechanics and machine learning makes it edge the WEL Research Institute (grant WEL-T-CR-2023A
increasingly hard to justify a guesstimate approach to - 05, “Smart Flow Systems”), the University of Liège
synthesis. This is particularly crucial for flow syntheses (“Crédit d’opportunité stratégique du Conseil universitaire
where the expense of re-writing centuries of batch protocols de la Recherche et de la Valorisation” under Grant No.
threatens to outweigh the potential benefits of flow. It is in OPP_CURV_22-44) and the F.R.S.-FNRS (Incentive grant
great deal the lack of rational ways of performing this for scientific research MIS F453020F, J.C.M.M.; PhD
translation that has deferred the transformation that flow Fellowship ASP 1.A.054.21F, P.B.) for funding. Computa-
technology was once hailed to bring to the chemical world. tional resources were provided by the “Consortium des
Aside from the pragmatic purpose of reaction optimiza- Équipements de Calcul Intensif” (CÉCI), funded by the
tion, the ease these tools provide to advance physical “Fonds de la Recherche Scientifique de Belgique” (F.R.S.-
organic chemistry descriptors is an invitation to increase the FNRS) under Grant No. 2.5020.11a and by the Walloon
depth of analysis of chemical developments. The DFT Region.
methods we identified also go beyond the application shown
here, being sufficiently accurate to immediately provide an
intensified temperature protocol that would suggest if and
how a known batch reaction can be applied in flow. It is

Angew. Chem. Int. Ed. 2024, 63, e202311526 (8 of 9) © 2023 The Authors. Angewandte Chemie International Edition published by Wiley-VCH GmbH
 15213773, 2024, 5, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/anie.202311526 by University Of Massachusetts Dartmouth, Wiley Online Library on [03/12/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
Angewandte
Research Articles Chemie

Conflict of Interest [16] P. Bianchi, A. Dubart, M. Moors, D. Cornut, G. Duhirwe, J.

Ampurdanés, J.-C. M. Monbaliu, React. Chem. Eng. 2023, 8,
The authors declare no conflict of interest. 1565–1575.
[17] S. Dana, I. Ramakrishna, M. Baidya, Synthesis 2017, 49, 3281–
3290.
[18] V. E. H. Kassin, R. Morodo, T. Toupy, I. Jacquemin, K.
Data Availability Statement Van Hecke, R. Robiette, J.-C. M. Monbaliu, Green Chem.
2021, 23, 2336–2351.
The Manuscript is accompanied with thorough Supporting [19] P. Bianchi, J.-C. M. Monbaliu, Org. Chem. Front. 2022, 9, 223–
Materials, which include computational data and methods as 264.
[20] T. Sasaki, Y. Ishibashi, M. Ohno, Chem. Lett. 1983, 12, 863–
well as detailed experimental protocols and analytical data.
866.
Data and scripts for the ML study are available as [21] K. Jorner, T. Brinck, P. O. Norrby, D. Buttar, Chem. Sci. 2021,
supplementary files, as well as flow data (kinetics procedures 12, 1163–1175.
and predicted residence times). [22] K. Spiekermann, L. Pattanaik, W. H. Green, Sci. Data 2022, 9,
Unprocessed data for structural identification and purity 417.
(NMR FID, HRMS, XRD structures, HPLC chromato- [23] E. H. E. Farrar, M. N. Grayson, Chem. Sci. 2022, 13, 7594–
grams) for all new compounds are uploaded as additional 7603.
[24] N. Mardirossian, M. Head-Gordon, Mol. Phys. 2017, 115,
Supporting Information on Dryad (DOI: 10.5061/dry-
2315–2372.
ad.3ffbg79q3). Two software packages (SnapPy and en- [25] I. Funes-Ardoiz, R. S. Paton, zendo repository 2018, https://doi.
FLOW) were developed during this work and are accessible org/10.5281/ZENODO.1435820.
on Zenodo repository (DOI: 10.5281/zenodo.8116089 for [26] J. M. Crawford, C. Kingston, F. D. Toste, M. S. Sigman, Acc.
SnapPy and for 10.5281/zenodo.8116110 enFLOW). Chem. Res. 2021, 54, 3136–3148.
[27] V. Dhayalan, S. C. Gadekar, Z. Alassad, A. Milo, Nat. Chem.
2019, 11, 543–551.
Keywords: Amination · Automation · Continuous Flow · [28] S. K. Nistanaki, C. G. Williams, B. Wigman, J. J. Wong, B. C.
Density Functional Calculations · Machine Learning Haas, S. Popov, J. Werth, M. S. Sigman, K. N. Houk, H. M.
Nelson, Science 2022, 378, 1085–1091.
[29] H. Shalit Peleg, A. Milo, Angew. Chem. Int. Ed. 2023, 62,
[1] A. Adamo, R. L. Beingessner, M. Behnam, J. Chen, T. F. e202219070.
Jamison, K. F. Jensen, J.-C. M. Monbaliu, A. S. Myerson, E. M. [30] F. Strieth-Kalthoff, F. Sandfort, M. Kühnemund, F. R. Schäfer,
Revalor, D. R. Snead, T. Stelzer, N. Weeranoppanant, S. Y. H. Kuchen, F. Glorius, Angew. Chem. Int. Ed. 2022, 61,
Wong, P. Zhang, Science 2016, 352, 61–67. e202204647.
[2] A. Bonner, A. Loftus, A. C. Padgham, M. Baumann, Org. [31] M. P. Maloney, C. W. Coley, S. Genheden, N. Carson, P.
Biomol. Chem. 2021, 19, 7737–7753. Helquist, P.-O. Norrby, O. Wiest, J. Org. Chem. 2023, 88,
[3] T. Noel, L. Capaldo, Z. Wen, Chem. Sci. 2023, 14, 4230–4247. 5239–5241.
[4] M. I. Jeraal, S. Sung, A. A. Lapkin, Chem. Methods 2021, 1, [32] T. Hart, V. L. Schultz, D. Thomas, T. Kulesza, K. F. Jensen,
71–77. Org. Process Res. Dev. 2020, 24, 2105–2112.
[5] O. J. Kershaw, A. D. Clayton, J. A. Manson, A. Barthelme, J. [33] H. Mayr, T. Bug, M. F. Gotta, N. Hering, B. Irrgang, B. Janker,
Pavey, P. Peach, J. Mustakis, R. M. Howard, T. W. Chamber- B. Kempf, R. Loos, A. R. Ofial, G. Remennikov, H. Schimmel,
lain, N. J. Warren, R. A. Bourne, Chem. Eng. J. 2023, 451, J. Am. Chem. Soc. 2001, 123, 9500–9512.
138443. [34] H. Mayr, A. R. Ofial, Pure Appl. Chem. 2005, 77, 1807–1821.
[6] P. Sagmeister, R. Lebl, I. Castillo, J. Rehrl, J. Kruisz, M. Sipek, [35] H. Mayr, A. R. Ofial, J. Phys. Org. Chem. 2008, 21, 584–595.
M. Horn, S. Sacher, D. Cantillo, J. D. Williams, C. O. Kappe, [36] T. Kanzian, H. Mayr, Chem. Eur. J. 2010, 16, 11670–11677.
Angew. Chem. Int. Ed. 2021, 60, 8139–8148. [37] R. Lucius, R. Loos, H. Mayr, Angew. Chem. Int. Ed. 2002, 41,
[7] C. P. Breen, A. M. K. Nambiar, T. F. Jamison, K. F. Jensen, 91–95.
Trends Chem. 2021, 3, 373–386. [38] A. I. Leonov, D. S. Timofeeva, A. R. Ofial, H. Mayr, Synthesis
[8] S. Chatterjee, M. Guidi, P. H. Seeberger, K. Gilmore, Nature 2019, 51, 1157–1170.
2020, 579, 379–384. [39] M. Vahl, J. Proppe, Phys. Chem. Chem. Phys. 2023, 25, 2717–
[9] A. M. K. Nambiar, C. P. Breen, T. Hart, T. Kulesza, T. F. 2728.
Jamison, K. F. Jensen, ACS Cent. Sci. 2022, 8, 825–836. [40] H. C. Brown, Y. Okamoto, J. Am. Chem. Soc. 1958, 80, 4979–
[10] A. A. Volk, R. W. Epps, D. T. Yonemoto, B. S. Masters, F. N. 4987.
Castellano, K. G. Reyes, M. Abolhasani, Nat. Commun. 2023, [41] V. García-Vázquez, A. Carretero Cerdán, A. Sanz-Marco, E.
14, 1403. Gómez-Bengoa, B. Martín-Matute, Chem. Eur. J. 2022, 28,
[11] C. Holtze, R. Boehling, Curr. Opin. Chem. Eng. 2022, 36, e202201000.
100798. [42] A. Milo, E. N. Bess, M. S. Sigman, Nature 2014, 507, 210–214.
[12] V. Hessel, D. Kralisch, N. Kockmann, T. Noël, Q. Wang, [43] C. Volpe, S. Meninno, A. Roselli, M. Mancinelli, A. Mazzanti,
ChemSusChem 2013, 6, 746–789. A. Lattanzi, Adv. Synth. Catal. 2020, 362, 5457–5466.
[13] P. Mueller, A. D. Clayton, J. Manson, S. Riley, O. S. May, N. [44] M. Ohno, M. Ido, S. Eguchi, J. Chem. Soc. Chem. Commun.
Govan, S. Notman, S. V. Ley, T. W. Chamberlain, R. A. 1988, 1530–1531.
Bourne, React. Chem. Eng. 2022, 7, 987–993.
[14] P. P. Plehiers, C. W. Coley, H. Gao, F. H. Vermeire, M. R.
Dobbelaere, C. V. Stevens, K. M. Van Geem, W. H. Green,
Front. Chem. Eng. 2020, 2, 5. Manuscript received: August 8, 2023
[15] C. B. Santiago, J. Y. Guo, M. S. Sigman, Chem. Sci. 2018, 9, Accepted manuscript online: October 24, 2023
2398–2412. Version of record online: December 22, 2023

Angew. Chem. Int. Ed. 2024, 63, e202311526 (9 of 9) © 2023 The Authors. Angewandte Chemie International Edition published by Wiley-VCH GmbH