Gelation properties of various long chain amidoamines: Prediction of solvent gelation via machine learning using Hansen solubility parameters

Article abstract of DOI:10.1016/j.molliq.2020.112587

Four new amphiphilic long chain amidoamine derivatives displaying different structure variations are synthesized and tested in 27 liquids and compared to the study of two similar molecules already reported in the literature. In many cases, these compounds can act as low molecular weight gelators to form a three-dimensional network in organic liquids or water, which can be confirmed by FE-SEM observations and rheology measurements. For each sample, XRD diffraction of the corresponding xerogel and FT-IR analysis of native supramolecular gels reveal that they can self-assemble into lamellar-like aggregates or in pseudo-cubic structures, depending on the alkyl chain length and the steric hindrance of the polar head. The number of amide bonds and their positions inside gelator structures are determinant for the nature of the packing. For each gelator, we perform a series of gelation tests in each of the solvents and show that Hansen parameters, which are known characteristics of each liquid, can be used to successfully predict their gelation properties via machine learning in the vast majority of liquids at a concentration of 4 wt%.

Full text of DOI:10.1016/j.molliq.2020.112587

Journal Pre-proof
Gelation properties of various long chain amidoamines:
Prediction of solvent gelation via machine learning using Hansen
solubility parameters
Frederic Delbecq, Guillaume Adenier, Yuki Ogue, Takeshi Kawai
PII:
S0167-7322(19)35196-7
DOI:
https://doi.org/10.1016/j.molliq.2020.112587
MOLLIQ 112587
Reference:
To appear in:
Journal of Molecular Liquids
Received date:
Revised date:
Accepted date:
16 September 2019
12 January 2020
26 January 2020
Please cite this article as: F. Delbecq, G. Adenier, Y. Ogue, et al., Gelation properties
of various long chain amidoamines: Prediction of solvent gelation via machine learning
using Hansen solubility parameters, Journal of Molecular Liquids(2018), https://doi.org/
10.1016/j.molliq.2020.112587
This is a PDF file of an article that has undergone enhancements after acceptance, such
as the addition of a cover page and metadata, and formatting for readability, but it is
not yet the definitive version of record. This version will undergo additional copyediting,
typesetting and review before it is published in its final form, but we are providing this
version to give early visibility of the article. Please note that, during the production
process, errors may be discovered which could affect the content, and all legal disclaimers
that apply to the journal pertain.
© 2018 Published by Elsevier.

Journal Pre-proof
Gelation Properties of Various Long Chain Amido Amines:
Prediction of Solvent Gelation via Machine Learning using Hansen
Solubility Parameters
Frederic Delbecqa*, Guillaume Adenierb, Yuki Oguec, Takeshi Kawaic
a
Ecole Supérieure de Chimie Organique et Minérale (ESCOM), EA TIMR 4297 UTC/ESCOM, 1 allée du
Réseau Jean-Marie Buckmaster, F-60200 Compiègne, France. *E-mail : f.delbecq@escom.fr
b 2 rue des Pommerelles, F-60200 Compiègne, France.
c
Department of Industrial chemistry, Tokyo University of Science, 1-3 Kagurazaka, Shinjuku, Tokyo
162-8614, Japan
ABSTRACT: Four new amphiphilic long chain amidoamine derivatives displaying different
structure variations are synthesized and tested in 27 liquids and compared to the study of
two similar molecules already reported in the literature. In many cases, these compounds
can act as low molecular weight gelators to form a three-dimensional network in organic
liquids or water, which can be confirmed by FE-SEM observations and rheology
measurements. For each sample, XRD diffraction of the corresponding xerogel and FT-IR
analysis of native supramolecular gels reveal that they can self-assemble into lamelar-like
aggregates or in pseudo-cubic structures, depending on the alkyl chain length and the sterric
hindrance of the polar head. The number of amide bonds and their positions inside gelator
structures are determinant for the nature of the packing. For each gelator, we perform a
series of gelation tests in each of the solvents and show that Hansen parameters, which are
known characteristics of each liquid, can be used to successfully predict their gelation
properties via machine learning in the vast majority of liquids at a concentration of 4 wt %.
Key-words: Amphiphilic compounds, supramolecular gelators, Hansen solubility parameters,
gelation prediction models.
1. Introduction
During the last decades, the interest for the production of low molecular weight gelators has
increased rapidly. These gelators are molecules that can self-assemble in liquids to form
fibers of sheet-like aggregates when subjected to unidirectional non-covalent forces such as
hydrogen bonds, - stacking or Van der Waals hydrophobic interactions [1]. The resulting
supramolecular fibers get entangled into a 3D-network that entraps the solvent (organic
liquid or water) within its interstitial spaces. It is important to remark that the gelation
phenomenon often occurs in an intermediate state between solubility and insolubility in a
given solvent. In fact, independently of the relative hydrophilic-lipophilic balance (HLB)
values of a good gelator, many conditions and factors are responsible for its gelation ability,
such as the nature of the solvent, the possible supersaturation of the liquid, which often
cause aggregation-type variations, the kinetics of the gelation phenomenon, and of course
the molecular structure of the gelator [2].

Journal Pre-proof
Among gelators, modified small lipopeptides [3,4] or long amidoamines displaying a large
variety of different polar heads have been studied for various applications such as drug
delivery systems, lithium-ion batteries, cosmetics, or oil-spill treatment. Gelators are
sometimes intended as templates for the generation and stabilization of nanosized materials
of noble metals [5-11]. For industrial purposes, it is essential to be able to understand and
predict the gelation abilities of gelator candidates in liquids of particular interest for specific
applications.
In this work, we investigate the impact of small and localized modifications of two lead
compounds 1 and 2 recognized as long amidoamine gelators [5,7], which have good relative
gelation capability on the packing nature of formed aggregates in standard liquids such as
toluene or water. The structure of all molecules is summarized in Fig. 1.
Fig. 1. Structure of organogelators 1 to 6.
Looking at the structure of the new compounds, that is, gelators 4-6, the most striking
features are the following. The elongation of the fatty tail by the replacement of the current
C18 alkyl chain with a longer alkyl chain of 22 atoms of carbon for gelator 4, the conversion
of terminal primary amine groups of gelator 1 into two urea moieties for gelator 5 and the
introduction of a methyl group on both sides of the polar head of gelator 2, closed to the OH
extremities in gelator 6, but also the presence of an additional amide bond inserted in the
structure of gelator 3.
After performing gelation tests on gelators 1-6 and confirmed the real gel nature of
selected sample by rheology, we observe the morphology of each gel sample using field
emission scanning microscopy (FE-SEM). Studying the respective air-dried xerogel of toluene
gel samples provides information on the strength of the intermolecular interaction by
Fourier Transform Infrared Spectroscopy (FT-IR). X-ray diffraction (XRD) of xerogels is
expected to provide enough indications on the possible packing nature formation. To the
best of our knowledge, no comprehensive and predictive study has been performed to date
on families of gelators with long fatty alkyl chains except in the works of Bouteiller et al [12]
or Rogers [22], where he tried to predict the gelation behavior in edible oils of 34
compounds grafted with at least one long alkyl chain, using the Hansen Solubility
Parameters via a predictive model.
The second part of this study is inspired by an article by Yan et al [13] and by other recent
works [14-24] that have demonstrated how 3-D spherical Hansen plots can be used to
predict variations in gelation capabilities inside a family of compounds. Hansen solubility
parameters (HSPs) have already been employed to predict the solubility of different material

Journal Pre-proof
in liquids such as polymers [25], small organic compounds [26-30], organometallic
complexes [31] and fatty-based biodiesels [32] through a machine learning approach. These
parameters provide a remarkable tool to build a comprehensive understanding of our series
of gelators and to predict their gelation ability before being tested. Given that HSPs are
known for all solvents, in particular for those reported in this work, it is thus possible to
quantify their potential to interact with gelators using their dispersion (훿푑), dipole-dipole or
polar (훿푝), and hydrogen bonding (훿퐻) interactions.
Three types of region of interest can be schematically identified in HSP space: solubility (S),
gelation (G) and insolubility (I) regions. They are empirically determined by gelation tests
consisting of mixing a small amount of the gelator being considered with various solvents.
The result of each gelation test (S, G, or I) with a particular solvent is marked in HSP space
and placed according to the HSP coordinates of the solvent. The set of all gelation tests thus
reveal regions in HSP space with similar gelation results. The results of a gelation test with
an unknown solvent can thus be determined according its location inside a previously
identified region in HSP space. These regions are typically represented as spheres, which can
be concentric or not.
However, in many cases, this spherical approach is too rigid to represent complex gelating
features, in particular for high polar solvents such as polyols or diols which are often found
outside the main gelation sphere. A question thus remains unsolved. Is it possible to give a
complete prediction of the gelation ability of a family of compounds without synthesizing
each of these derivatives? A good matching between real tests and simulation could provide
an answer to this question.
2. Experiment
2.1. Materials
All chemicals that we used were reagent grades obtained from Aldrich, Acros Chemical or
TCI Japan. Commercially available reagents were used without further purification, except in
the case of methyl acrylate furnished by Kanto Chemicals, which was purified by distillation
under reduced pressure in a nitrogen atmosphere, and in the case of octadecylamine, which
was recrystallized two times starting from hexane.
2.2. Organic synthesis
Gelators 1 and 2 were synthesized using common procedures reported in the literature [5, 7].
Gelator 3 was prepared using the same protocol as reported in a previous paper [8].
NMR spectra were recorded in deutered solvents such as DMSO-d6 or CDCl3 with a Bruker
1
400 Ultrashield Spectrometer operating at 400 MHz. (See Figure S1 of the ESI for H NMR
charts of compounds 3-6). Chemical shifts () are quoted in ppm and are referenced to TMS
(tetramethylsilane) as an internal standard. Coupling constants (J) are quoted in Hz, common
splitting patterns and their abbreviations were s (singulet), d (doublet), t (triplet), q (quartet),
and m (multiplet). Mass spectra were recorded on Bruker Daltonics micro TOF Focus mass
spectrometer.

Journal Pre-proof
For the gelators 1 and 2, their purities were confirmed by 1H NMR and mass spectroscopy.
2.2.1. Synthesis
of
N-(2-amino-ethyl)-3-{[2-(2-amino-ethylcarbamoyl)-ethyl]-
behenylamino} propionamide 4
To 50 mL of toluene were added 5.5 g (16.17 mmol) of behenic acid. Then, to the resulted
dispersion was added dropwised 10.26 g (80.78 mmol) of oxalyl chloride and the complete
mixture was stirred refluxing for two hours under argon atmosphere. The solvent and
unreacted reagents were both removed under reduced pressure to produce colored oil
which was introduced in 50 mL of THF without further purification. In the next step, 20 mL of
25 % NH4OH aqueous solution was added to the previous clear yellow solution and a white
solid started to precipitate which was successively filtrated on Buchner, washed with a slight
volume of water and THF. (mass = 5.43 g ; yield = 99 %) The recovered benehamide powder
was dried in a dessicator and remained insoluble in common organic solvents. FT-IR: 3396
cm-1 (NH) ; 2849-2916 cm-1 (CH2) ; 1647 and 1526 cm-1 (amide). MS : Calcd (M+H+) 339 found
339.23.
In the next stage, 2.66 g (7.84 mmol) of the previous behenamide was dispersed in 40 mL of
dry THF, followed by addition of 0.60 g (15.68 mmol) of Lithium aluminium hydride (AlLiH4).
Thus, the resulted dispersion was allowed to stir refluxing for 1 day. After returned to the
room temperature, the solution was poured in water and extracted with several volumes of
CHCl3 and diethyl ether. All organic layers was separated and combined before the drying on
MgSO4. The final organic solution was evaporated in vacuum to produce a white solid. (mass
= 1.1 g ; yield = 43 %). 1H NMR (CDCl3) :  0.85 (t, 3H, J = 1.62 Hz) ; 1.1 (m, 40 H, CH2) ; 3.50 (t,
2H, J = 6.3 Hz, CH2NH2). MS : Calcd (M+H+) 325 found 325.13.
By using the synthesis protocol reported for 1, the diester intermediates was obtained as
follow: 1.1 g (3.38 mmol) of the previously synthesized amine was introduced in 30 mL of
MeOH, followed by the addition of 1.45 g (16.86 mmol) of methylacrylate. The resulted
solution was stirred at 60°C for 4 hours, a slightly amount of white solid precipitated which
was later filtrated. Both solvent and unreacted acrylate was removed by distillation in
normal pressure to produce clear oil that recrystallized at room temperature. The generated
diester was introduced without further purification in 30 mL of MeOH containing 7.0 g (0.11
mol) of ethylene diamine. The resulted solution was heated up at 70°C for one day, all
liquids were removed under vacuum to give an oil which crystallized into a white cream solid.
(mass : 1.02 g ; yield = 55 % for two steps)
1H NMR (CDCl3):  0.88 (t, 3H, J = 1.44 Hz) ; 1.25 (br, 36 H, CH2) ; 1.42 (br, 4H, CH2CH3,
CH2CH2CH2N) ; 2.37-2.82 (m, 14 H, CH2CH2CO, CH2N, NCH2CH2, CH2NH2) ; 3.30 (m, 4H, NHCH2).
MS : Calcd (M+H+) 553 found 553.17.
2.2.2. Synthesis of [(octadecyliminobis)bis(ethylenecarbonylaminoethylene)]-1,1’diurea
5
The diamine 1 (1.0 g, 2.01 mmol) and 0.60 g (4.42 mmol) of commercial phenylcarbamate
(97 % Aldrich) were both dissolved in 40 mL of THF. The resulted solution was stirred at 60°C
and 0.44 g (4.42 mmol) of trietylamine (TEA) was added before carry on the stirring for 4

Journal Pre-proof
hours. Once the heating was stopped, the solvent was removed in vacuum and the crude
product was recristallized from a mixture of MeOH and toluene to produce a yellow solid.
(mass = 0.85 g ; yield = 72 %).
1H NMR (CDCl3) :  0.88 (t, 3H, J = 7.0 Hz, CH2CH3) ; 1.25-1.63 (Br, 24H, CH2), 2-16-2.73 (m,
14H, CH2CH2CONH, NCH2CH2CO, CH2CH2N and NCH2CH2CO) ; 3.27 (m, 8H, CH2CH2NHCONH2
and NHCH2CH2). MS : Calcd (M+H+) 583.2 found 573.17.
2.2.3. Synthesis of N-(2-hydroxy-1-methylethyl)-3-{[2-(2-hydroxy-1-methylethylcarbamoyl)
-ethyl]octadecylamino}propionamide 6
The diester precursor or more exactly the 3-[(2-methoxycarbonyl-ethyl)-
(octadecylamino]propionic acid methyl ester (3.0 g, 6.80 mmol) was introduced in 15 mL of
MeOH, then 1.03 g (13.80 mmol) of (S-)-(+)-2-aminopropanol was added and the resulted
mixture was stirred for one week at 60°C. The methanol was removed by evaporation and
the resulted solid was triturated in acetone and recovered by filtration on buchner before
drying in vacuum. (mass = 2.8 g ; yield = 78 %).
1H NMR (CDCl3) :  0.88 (t, 3H, J = 7.2 Hz, CH3) ; 1.16 (d, 6H, CHCH3) ; 1.25-1.42 (br, 32 H,
CH2) ; 2.36-2.73 (m, 10 H, CH2CH2CO, CH2N and NCH2CH2CO) ; 3.40-3.68 (2m, 4H, CH2OH) ;
4.96 (t, 2H, NHCH(CH3)CH2). MS : Calcd (M+H+) 527 found 528.45.
2.3. Gelation tests
Gelators 1-6 taken at different concentrations were mixed in a test tube with a precise
volume of a selected solvent and heated up until a clear transparent solution was obtained.
The resulting solution was left to cool down at room temperature. After a few minutes, if the
initially viscous or liquid solution became a gel-like material, the sample was turned upside
down for confirmation. The sample was considered as a gel if the sample remained perfectly
immobilized. In other cases, it was recognized as soluble, insoluble or simply a gelatinous
precipitate.
2.4. Characterizations of gels
Rhelogy measurements were realized using a Physica HCR 301 (Anton Paar) rotational
rheometer equipped with a disk and a cone forming the measuring system suited for high
viscosity samples. The rheometer is also equipped with a Peltier temperature control system
that can regulate the temperature in a range from - 40 to 200°C. The measurements were
monitored by the Rheoplus software. All experiments were performed at 20°C and the
viscosity was checked for frequencies ranging from 0.1 to 10 Hz. The viscoelasticity of
selected samples was characterized in terms of the elastic modulus (G’) and the loss
modulus (G’’). X-ray diffraction (XRD) measurements were performed with a Rigaku Ultima
IV diffractometer using an air-dried sample of 2 wt % supramolecular gels. An FT-IR was
conducted using a spectrometer (Thermo-Fisher Scientific Nicolet 6700 FT-IR) equipped with
an MCT detector with a resolution at 4 cm-1. A liquid cell with ZnSe windows was employed
for organogels, and a Harrick GATR attenuated total reflection (ATR) attachment was used
for hydrogels. A demountable liquid cell with a CaF2 window was employed for solvent
solution FT-IR measurements. Then, one thousands scans were recorded on average for

Journal Pre-proof
each sample spectrum. Scanning electron microscopy was carried out using a Jeol JCM-5700
microscope. The xerogels were subjected to Au-Pd Sputtering (Hitachi E-1010 sputter coater)
before starting the experiments. Differential Scanning Calorimetry (DSC) was conducted on a
DSC 8MC Mettler-Toledo using aluminium pans. Scans were conducted under nitrogen with
a heating rate of 10°C/min in the temperature range of 40-250°C.
3. Results and discussion
3.1. Evaluation of gelation ability for all gelators (1-6)
The gelation ability of gelators 1-6 was examined for a large panel of organic liquids, as well
as for water. (Some gel samples are given as example in the ESI, Fig. S2) The gelation tests
were calibrated as follows. For each sample a mass of 0.04 g was introduced in 1.0 mL of
solvent and heated up until the material was entirely dissolved. When the sample remained
insoluble in the liquid, it was recorded as such. When dissolved, the resulting transparent or
clear yellow solution was left to cool down at room temperature. It then turned into a more
viscous or turbid solution until a gel started to appear, within a short period of time.
Depending on the gelator, the formed gel could be stable for up to three months without
syneresis, or could simply precipitate within a few hours. In order to confirm the gelation
ability of each gelator 1-6, the corresponding 4 wt % gel samples in cyclohexane were
employed for testing their rheological properties. All samples were observed as strong
elastic gels throughout the entire frequency range with higher G’ values always over 10000.
Here, G’ values were systematically higher than G’’ and always parallel with the respect to
the angular frequency. The gelation ability of gelators 1-6 was tested in a large variety of
solvents, as reported in Table 1. It should be mentioned that the gelation ability of gelators 1,
2 and 3 has already been partially described in the literature [5,7, 8].
Table 1. Gelation ability for gelators 1-6 in water and various organic liquids introduced in a
liquid at 4 wt % with their respective Hansen solubility parameters values (MPa1/2).
훿푑
훿푝
훿퐻
1
G
I
2
G
S
G
S
S
S
S
G
I
S
G
S
I
S
S
I
3
G
I
G
S
S
I
S
G
G
S
G
I
4
I
G
I
5
G
I
G
I
6
G
S
G
S
I
S
S
G
G
S
G
S
I
Toluene
18.0
17.3
18.4
19.0
19.6
19.4
18.0
16.8
14.9
17.8
17.8
16.8
14.4
15.5
15.3
15.3
17.4
18.4
1.4
2.8
0
8.8
8.6
5.1
5.1
0
0
3.1
0
5.7
2.9
10.4
6.1
18.0
13.7
16.4
2.0
7.4
2.0
5.9
3.7
10.2
12.9
0.2
0
5.7
0.6
8.0
5.1
7.0
4.1
6.1
11.3
10.2
2-Methylfuran
Benzene
Pyridine
Acetophenone
Aniline
m-Cresol
Cyclohexane
n-Hexane
Chloroform
CCl4
G
G
S
S
S
G
G
I
G
S
G
I
S
G
G
S
G
G
S
G
S
G
I
I
S
S
G
I
G
G
G
I
I
I
I
G
G
THF
Dibutylether
Acetone
MIBK
Acetonitrile
DMF
I
I
S
I
I
I
I
I
S
S
S
I
I
I
G
G
G
S
S
DMSO
I
I

Journal Pre-proof
Ethylacetate
Ethyllevulinate
Propylene carbonate
Methanol
Butanol
Ethylene glycol
Glycerol
15.8
16.5
20.0
15.1
16.0
17.0
17.4
17.8
15.5
5.3
7.8
7.2
6.8
4.1
22.3
15.8
26.0
29.3
1.0
G
I
I
S
S
I
G
G
S
I
I
I
S
S
G
G
I
I
I
I
S
S
I
G
I
G
G
G
I
S
I
G
I
I
I
I
I
S
I
S
G
I
I
I
I
S
S
I
G
i
18.0
12.3
5.7
11.0
12.1
0.4
Triethylamine
Water
16.0
42.3
G
S
I
G
MIBK = Methylisobutyl ketone; THF = Tetrahydrofuran; DMF = Dimethyl formamide; DMSO
=Dimethyl Sulfoxide ; S : soluble ; G : gel or opaque gel; I : insoluble.
3.2. Machine learning classification of HSP space
Predicting the formation of gels, such as displayed in Table 1, is a daunting task [13]. Hansen
solubility parameters (HSP) are widely used to help classify how specific compounds react
when immerged into solvents [14], and more specifically to attempt to predict the formation
of gels [15].
As we will see, for gelators 1-6 the regions defining all three relevant classes (soluble, gel,
insoluble) in HSP space, usually established by the calculation of a specific interaction radius
(Ra) [32], are generally too complex to be able to be sketched with a simple spherical
approach, be it concentric or not [16,17]. In order to provide useful insight into gel
formation, we therefore chose to turn to machine learning approaches to classify HSP space
for each gelator, whenever possible. This approach has already been used successfully to
predict gel formation [18].
For this purpose, several machine learning methods have been compared and we have
evaluated their performances not only by measuring their predictive abilities but also by
assessing their interpretative usefulness. A common drawback of machine learning
approaches is indeed that although they can be excellent at predicting classes of previously
unseen input states, they are essentially black boxes providing results without explanation
as to why a specific input state is predicted to belong to a certain class rather than to
another. We thus require that the model should provide visualizations that favor human
interpretation.
3.3. Predictive performances of various machine learning methods
We implement the machine learning classification with a code written in Wolfram
Mathematica, with which we compare the performances of several methods: nearest
neighbors (k-NN), Naïve Bayes, Random Forest, Logistic Regression, Neural Network,
Support Vector Machine (SVM), Decision Tree, and Gradient Boosted Tree.
As can be seen in Table 1, the gelation properties of each gelator have been tested for each
of the 27 potential solvents. The input data are the Hansen solubility parameters (δd, δp and
δH) in Table 1, and the output is the corresponding result of the gelation test in the column
of the gelator being considered.

Journal Pre-proof
for each algorithm 퐴 ∈ {SVM,RF,NN,…}
initialize the number of correct predictions 푛correct
for each solvent 푘:
(
)
discard solvent data (푿푘, 푦푘) from dataset { 푿푖, 푦푖 , ∀푖}
{(
)
}
train new classifier 푐푘 on remaining data 푿푖, 푦푖 , ∀푖 ≠ 푘
predict class 푦̃푘 of discarded solvent 푘 as 푦̃ = 푐푘(푿푘)
푘
if 푦̃ = 푦푘 increment number of correct predictions 푛correct
푘
calculate accuracy as 푛correct/푁푠표푙푣푒푛푡푠
Fig. 2. Pseudocode for LOOC algorithm comparison, where the HSP (휹풅, 휹풑, 휹푯 ) of each
solvent 풊 are taken as input 푿풊 for the classifiers and the 2-class response (Gel or non-Gel) of
the gelator is taken as output 풚풊.
To mitigate the risk of over-fitting, we evaluate the predictive performances of each method
using a leave-one-out cross-validation (LOOC) procedure, with a two-class distinction (gel or
non-gel). It means that we take out one solvent from this list and train the machine learning
algorithm with the remaining 26 solvents. Once the classifier has been trained with the set
of 26 solvents, we test whether it can correctly predict the class of the solvent that was left
out. We repeat this procedure with all possible 26 to 1 partitions and count how many times
the method correctly predicted the class of the solvent that was left out. The pseudocode
for this procedure is described in Fig. 2. The predictive performance displayed in Fig. 3 is the
ratio of the number of correct predictions over the number of partitions (that is, 27). We use
this procedure for all gelators and for all classification methods.
1
0.9
0.8
kNN
0.7
NaiveBayes
0.6
RandomForest
0.5
LogisticRegression
NeuralNetwork
SVM
DecisionTree
GradientBoostedTrees
0.4
0.3
0.2
0.1
0
1
2
3
4
5
6
Gelators
Fig. 3. Predictive performances of several machine learning methods, for all gelators 1-6. The
performances are measured with a leave-one-out cross-validation (LOOC) method, with a
two-class distinction (gel or non-gel). The performance of a fully random classifier would be
equal to 0.5.
The performances of the different methods vary with each gelator (see Fig. 3). Since we
consider only two output classes (gel or non-gel), the predictive performances should be

Journal Pre-proof
compared to that of a fully random classifier, that is, one that would attribute classes at
random with a probability 1/2. As can be seen in Fig. 3, the performances of the
classification methods are only consistently better than a random classifier for gelator 1, 2, 3
and 6.
The predictive performance is particularly poor for gelator 4. It can be attributed to the fact
that the classes for this gelator are not grouped as clusters but rather intricately intertwined,
with a probability to find a closest neighbor of the same class of only 0.37, very close to that
of a random probability of 1/3.
The performances of the different methods appear to be very similar. For a specific gelator,
the fact that one particular method performs better than others may very well be a mere
coincidence. Choosing a different method for each gelator to improve performances might
very well be equivalent to over-fitting. Similarly, adjusting the machine learning settings to
maximize the classification accuracy and predictive performance for each gelator might very
well lead to an overestimation of the performances of the classifiers.
3.4. Interpretation of results using Support Vector Machine method
Since all methods perform more or less equally well as predictive tools, we therefore chose
the method that provides the most natural looking and easily interpretable visualizations. In
this respect, the best choice appears to be that of Support Vector Machine (SVM). Other
methods tend to either look blocky with straight lines that look rather contrived, or blurred
with regions that are not easy to distinguish or interpret.
Since we want to be able to visualize not only the gelating region (G) but also the soluble (S)
and insoluble (I) regions in HSP space, we train the SVM as a 3-class problem. We have
measured the performance of the SVM classifier with LOOC validation and obtained an
average accuracy of 54% for the six gelators (using a one-vs-all method, where each class is
trained against all other classes). This lower accuracy of 54% should be compared with the
33% performance of a random classifier, so that the performance of the classifier is arguably
not fundamentally reduced with respect to the 2-class performances. The pseudocode for
the visualizing procedure is described in Fig. 4.
initialize 3D grid of 푛 × 푛 × 푛 points {푿푘,푙,푚} in HSP space
(
)
train SVM classifier on gelator dataset { 푿푖, 푦푖 , ∀푖}
get class probabilities 푝푆, 푝퐺, 푝퐼 from SVM classifier for each grid point 푿푘,푙,푚
for each class probability 푝 ∈ {푝푆, 푝퐺, 푝퐼}:
generate 3D density plot for points 푿푘,푙,푚 with 푝 > threshold (~0.33)
Fig. 4. Pseudocode for visual representation of gelating regions in HSP space, where the HSP
(훿푑, 훿푝, 훿퐻 ) of each solvent 푖 are taken as input 푋푖 for the classifier and the 3-class response
(S, G, or I) of the considered gelator as output 푦푖.
It should be noted that some features are beyond any classification method and show the
intrinsic limitations of the HSP method. The example of triethylamine is particularly telling.
Indeed, for several gelators 2, 3, 5 and 6, triethylamine is tightly surrounded by close
neighbors in HSP space that are all pointing to a different class than that of triethylamine.
There is strictly no way that a classification method, however advanced, would predict such

Journal Pre-proof
a feature that would almost appear to be an anomaly if the same behavior was not observed
with multiple gelators.
Nearly all gelators, except gelator 4 and perhaps gelator 3, exhibit two disconnected regions
of gelation, which are very remote from one another in HSP space and separated by large
regions associated with different classes. One gelation regions is typically associated with
low polarity and low hydrogen bonding (near benzene and toluene), whereas the other is
associated with high polarity and high hydrogen bonding (near water and glycerol).
Besides, compared to the classic approach of Hansen space found in the literature [26, 33],
due to non-spherical nature of each zone on our graphics, it appears not easy to interpret
the relationship between volume expressed by each respective gelation zone and some
different empiric results such as kinetic or undercooling without consistent and specific
values of Ra.
3.4.1. Gelator 1
Gelator 1 is known for its strong gelating capability in aromatic solvents such as benzene or
toluene. This behavior is confirmed in Fig. 5 (left). With high polar solvents, for example
dimethylsulfoxide or DMF, the solid remains insoluble even when introduced at a lower
concentration of 0.2 wt %. Interestingly, it is able to form a gel in hydrogen-bond acceptor
liquids, such as triethylamine or pyridine that can also be considered as aromatic solvents.
Surprisingly, even though ethylene glycol (EG) and glycerol are both water soluble polyols,
gelator 1 can form a gel in the former but not in the latter. This phenomenon might be due
to the smaller number of OH groups in EG, leading to more interactions between gelator 1
and its environment. In addition, a second gelation region that could correspond to
unknown solvents or liquid synthetic polyols appears in Fig. 5, but given that the classifier
seem to identify this region based on the known class of glycerol alone, it should rather be
interpreted as a potential region of interest that would call for further tests rather than as a
prediction.

Journal Pre-proof
Fig. 5. Gelation classification in HSP space for gelator 1 (left) and gelator 2 (right). Solvents
are marked with different colors depending on their gelation properties: soluble (blue), gel
(green) or insoluble (red). The regions predicted by the classifier (SVM) are displayed in the
same colors (blue, green and red) as probability clouds for all three classes. The opacity of
the clouds intensifies with the predicted probability above that of a fully random classifier.
Solvents that are misclassified by the classifier are marked with a center black disk.
3.4.2. Gelator 2
Gelator 2 is shown in Fig. 5 (right). It is characterized by the replacement of the two terminal
primary amine groups by OH moieties, which impacts only slightly the performances of the
derived gelator. This molecule is almost insoluble in common organic solvents, although it
can produce a gel in ethylene glycol, glycerol and water at a lower concentration of 1 wt %.
In addition, gelator 2 is capable of gelating non H-donor aromatic solvents such as benzene,
as well as more hydrophobic liquids such as short alkanes or CCl4.
The particular case of triethylamine is interesting. Given its HSPs values, one would have
expected that it behaves like other apolar liquids. The contribution of its non-binding pair of
nitrogen atom could however explain the differences observed with its neighbors in HSP
space, with the idea that different binding modes of a gelator could occur inside gel
networks, depending on the nature of the solvent employed for each test. By contrast with
gelator 1, with its strong hydrogen bonding between adjacent OH groups, gelator 2 could
self-assemble into some type of lipidic bilayer exposing either its hydrophobic tail or polar
head according the nature of the host liquid. Furthermore, these two possible modes of
assembly could arguably explain the presence of two distinct three dimensional gelation
spaces in the corresponding HSP solubility plot.
3.4.3. Gelator 3
The introduction of an additional amide group in gelator 3 can be interesting on several
accounts, as can be seen in Fig. 6 (left). Firstly, this gelator is expected to provide superior
hydrophilic-lipophilic balance value (HLB) compared to gelator 1, which could provide a
simple explanation to the better water solubility of the solid. Secondly, the gelation abilities
of gelators 1 and 3 are similar, except in nitrogen rich liquids where gelator 3 shows
difficulties in being solvated, which might be caused by the inhibition of the formation of
successful hydrogen bonds by these solvents. Thirdly, gelator 3 is unexpectedly almost
insoluble in all water miscible solvents, but it can nevertheless effectively act as a gelator in
glycerol at a lower concentration of 2 wt %. This phenomenon could be explained by the
normal viscosity of glycerol, which means that gelator 3 could possibly be acting as a
thickener, without playing the traditional role of a gelator.
3.4.4. Gelator 4
As shown in Fig. 6 (right), gelator 4 is obviously more hydrophobic than gelator 1 due to the
longer C22 alkyl chain in its structure. Unlike with previous gelators, we were able to obtain
opaque gel samples that were stable in unusual solvents such as acetonitrile, acetophenone,
2-methylfuran or DMF, even with only a small amount of this gelator. Curiously, this
molecule seems to be a good candidate to produce gels in the most polar water soluble

Journal Pre-proof
solvents (DMF, DMSO, and acetonitrile) but also in non-water miscible liquids, especially in
ketone, as well as in the smallest ethyl esters. In this case, the total HLB value was heavily
decreased. We will discuss later in the paragraph 3.5 of the real contribution of the polar
head in the packing, but the balance was definitively broken for this material. However, if we
superpose both models for gelators 1 and 4, it seems clear that the gelation ability of gelator
4 is complementary to that of gelator 1.
Fig. 6. Gelation classification in HSP space for gelator 3 (left) and gelator 4 (right). Solvents
are marked with different colors depending on their gelation properties: soluble (blue), gel
(green) or insoluble (red). Here, the classifier experiences some difficulties tackling with the
complexity of the gelation properties in HSP space, especially for gelator 4, as it misclassifies
the whole HSP space as a gel.
3.4.5. Gelator 5
As shown in Fig. 7 (left), gelator 5 exhibits three distinct and nested regions in HSP space.
This behavior is arguably a consequence of the introduction of urea moieties on both
extremities of the polar head of the gelator. Due to its high probability to form a large
number of additional intermolecular hydrogen bonding in complicated networks, gelator 5 is
surprisingly not a good candidate to form a gel in the majority of the liquids used in this
study, even in MeOH or pyridine. As was the case with gelator 3, the HSP solubility plots of
gelator 5 shows that increasing the amide bonds inside the structure of the gelator leads to
a drastic decrease of its solubility in various types of solvents. Obviously, not only the
number but their positions in relation to the hydrophilic head are responsible of the serious
solubility changes that could be potentially controlled.
3.4.6. Gelator 6
Finally, the solubility of gelator 6 shown in Fig. 7 (right) is affected by the presence of two
methyl groups on the periphery of the polar head, near the OH groups. The head of the
molecule is more hydrophobic, and overall bulkier. This behavior is however different than
that of gelator 2, as it might prevent the formation of head-to-head complexes between two

Journal Pre-proof
molecules promoted by intermolecular hydrogen bonding, which is the main cause of
hydrogen bond weakening. Fig. 5 (right) shows that, independently of the chirality of the
extremities of gelator 6, the held methyl groups have only a small influence on the gelation
ability of this gelator, and that although the solubility region is somewhat reduced, it
remains overall consistent with that of gelator 2, for which two clear gelation regions are
visible.
Fig. 7. Gelation classification in HSP space for gelator 5 (left) and gelator 6 (right). Solvents
are marked with different colors depending on their gelation properties: soluble (blue), gel
(green) or insoluble (red). The regions predicted by the classifier (SVM) are displayed in the
same colors (blue, green and red) as probability clouds for all three classes. The opacity of
the clouds intensifies with the predicted probability above that of a fully random classifier.
Solvents that are misclassified by the classifier are marked with a center black disk.
3.5. FE-SEM observations of the xerogels
Using gelators 1-6 at a concentration near their Gelation Minimum Concentration (GMC), we
obtained electron microscope images for air-dried self-assembled xerogels in toluene
organogels, except for gelator 4 where ethylacetate was used. In Fig. 8, in all images, the 3D-
networks are made of long thin fibers with average diameter of 1 m, except for gelator 4.

Journal Pre-proof
Fig. 8. FE-SEM images of air-dried organogels in toluene obtained for gelators 1-3, 5, 6, in
images a-c) to e) and f) respectively, and d) in ethylacetate for gelator 4.
For gelators 1 and 3, the fibers are very thin compared to those obtained from gelators 2
and 6. For these two diol-type gelators, it is reasonable to assume that two different binding
modes occur inside their gel matrix. Indeed, microscopic observations of the xerogels
obtained from gelators 2 and 6 in water show alternative surface morphology. That is,
gelator 2 gives a visibly weaker and altogether different structure network than the long
crystalline rods observed with gelator 6.
For gelator 4, the difficulties that we experienced in producing a gel from toluene convinced
us to use an ethylacetate (AcOEt) xerogel instead. Fig. 8 d) shows a gel matrix networks with
a visibly thick enough net to entrap the solvent, which could be directly connected to the
length of the fatty alkyl chain grafted on the remaining unchanged polar heads. Finally, for
gelator 5, a typical entangled three-dimensional -sheet network was formed in toluene,
with less apparent cavities.
In a recent work, Rogers and his collaborators have attempted to develop a correlation
between fiber lengths and the Hansen sphere radius or the H values of respective solvents
[33]. In our case, it appears difficult to get a specific exploitable value like radius Ra and we
can only interpret the volume of the gelation area surrounding the toluene location on the
graphics, especially for gelators 2, 3, 5 and 6. The values of the volumes are given from our
calculation models in arbitrary unity with values of 78.2, 42.6, 115.3 and 74.6 respectively
for 2, 3, 5, and 6. Interestingly, the values obtained from 5 were recorded as higher than the
others because of two nearest gelations lobes connected via a narrow area. In regards to
these results, it seems to be possible that the volume of a gelation area could become
enough informative. Thus, it is possible to imagine a direct connection with the thickness
and length of the networks fiber. For example, for gelators 2 and 6, we got the closest values
and similarities on the respective SEM pictures. By decreasing the value of the gelation space
around the toluene position, we observed thinner and longer fibers for 3.
3.6. IR spectroscopy of gel samples

Journal Pre-proof
As depicted in Fig. 9, FT-IR spectra of gel samples were measured in toluene-d8 to determine
the representative bands of physical gel, except for the more hydrophobic sample 4 where
DMSO-d6 was used because of its lack of solubility in apolar benzene-type solvents.
Fig. 9. FT-IR spectra: a) Toluene-d8 gels of gelators 2 and 6 (on left: NH bands variations; on
right: amide (I) band variations); b) Toluene-d8 gels of gelators 1, 3 and 5 (on left: NH bands
variations ; on right: amide (I) band variations).
For all gels samples, typical absorption bands that are characteristic of trans alkyl chains
were generally observed at a value superior to 2900 cm-1, which might be considered as
evidence that hydrophobic packing participates in the stabilization of the entire structure of
xerogels. Compared to gelator 6, gelator 2 shows lower OH vibration bands located at 3294
cm-1, with a broad band found between 3300 and 3400 cm-1. For gelator 6, this result can be
caused by the methyl group on intermolecular H-bonding networks, precisely between OH
extremities of two adjacent molecules. Moreover, lower value of amide (I) bond vibration
also confirms the creation of strong packing between compounds, which is a consequence of
stronger hydrogen bonding networks. This situation results in higher crystallinity of the
material associated with an increase of its GMC.

Journal Pre-proof
For gelator 3, the introduction of another inserted amide bond played an important role in
gel formation. Compared to 1, the lower signal at 1637 cm-1 can be seen as evidence of a
stronger packing and of a tendency to form crystals in a large variety of organic liquids.
Furthermore, the contribution of the additional amide group appears with two distinct
bands located at 3360 and 3282 cm-1, respectively. However, in case of gelator 5 the signal
was decomposed into distinct bands that also attributed to urea moieties. For the amine
signal, there are three consecutive NH bands located at 3453, 3325 and 3283 cm-1. For the
amide (I), they are located at 1659 and 1639 cm-1, which suggests that a part of the terminal
amide function, remains free and that the gelator only gets involved in a simple
interdigitated structure.
For gelator 4, the NH vibration band is interestingly as high as 3458 cm-1. This value provides
evidence of the lower contribution to the networks of terminal amine group components of
polar heads. Hydrophobic interactions dominate superficial hydrogen bonds.
3.7. Calorimetric characterization of formed supramolecular gels
On the Table 2, the phase transition temperature Tgel was determined in toluene for gelators
1, 2, 3, 5, and 6, and in ethylacetate for gelator 4. It was measured by differential scanning
calorimetry (DSC) in a range of temperatures between 15 and 80°C. The gel-to-sol melting
temperature (Tg-s) was measured for toluene organogels at 3 wt %. Gelators 1, 3, 4, and 5
gave values of 46, 43, 68 and 80°C, respectively. For gelators 4 and 5, due to the evidently
greater packing strength, the transition requires more energy. Against all expectation, the
Tg-s value of organogel 3 is not much higher than that of gelator 1, despite its additional
inserted amide group and its possibly higher number of intermolecular hydrogen bonds.
That said, with a value of 66°C, gelator 2 has a smaller melting temperature than that of
gelator 6, whose melting temperature of 77°C can be seen as a sign of higher crystallinity of
the material inside the gel matrix. Herein, despite our efforts, for this series of gelators, we
were not able to measure a coherent value of sol-to-gel crystallization (Ts-g) values. On the
other hand, the gelation space volume does not increase linearly with the measured melting
temperature values.
Table 2. Gel melting temperature from DSC measurements (from toluene).
Compound
1
2
3
4
5
6
Tg-s (°C)
46
66
43
68
80
77
3.8. Molecular packing and gel network organization
X-Ray diffraction (XRD) analysis of the xerogels was performed to study their network
structures. For all gelators, the reflection peaks observed at small angle regions are reported
in Table 3, together with the associated d-spacing values. The 3D hypothetical models of
organogels 1-6 were subjected to a simple empirical method for calculation of the molecular
electronic structure (AM1) found in the Argus lab a freeware available on the web, in order
to minimize these conformational energies.

Journal Pre-proof
Table 3. First XRD reflections peaks of air-dried toluene organogels 1-3, 5-6 and air-dried
AcOEt organogel 4 ; predicated length of gelators 1-6.
Gelators
d-spacing (nm)
first reflection peaks (nm)
Calculated length (nm)
1
2
3
4
5
6
3.76
3.47
5.06
4.08
3.0
2.38
2.66
3.47
3.94
3.54
2.80
2.97
3.08
3.17
3.55
3.21
3.07
3.4
For all gelators, in the wide angle region, the xerogels displayed also sharp diffraction peaks
representative of highly ordered lateral chain packing. As reported elsewhere [5, 8], for
lipopeptide-type gelators, the d-spacing ratios are practically 1 : 1/2 : 1/3. Indeed, gelators 1
and 3 have ratios of reciprocal of integers, which is characteristic of a lamellar structure. In
case of gelator 4, the ratio pattern evolved by the same manner. Interestingly, for gelator 5
the first d-spacing value of 3.0 nm was not compatible with the typical lamellar structure
generally observed with lipopeptide-type organogelators as depicted on Fig. 10. Even the
normal d-spacing ratio of a parallel interdigitated network was finally observed, the gelators
5 is suspected to show a real tendency to self-assemble into a parallel -sheet system
reinforced by a stronger intermolecular hydrogen networks between the sheets, due to the
presence of urea terminal groups. Thus, the d-spacing is really lower compared to those
expressed by other gelators. Hence, for all gelators except gelator 4 and 5, regarding the
computational simulation, the bilayer thickness appears to be smaller than twice the
molecular length at full extension. This situation is representative of partial interpenetration
of the fatty alkyl chain inside the gel networks.
Fig. 10. Molecular models of bilayers structures obtained from 1-3, 6.
Comparing gelator 2 and 6 is also interesting. Just like gelators 1 and 2 that are able to self-
assemble in toluene into parallel lamellar structures with approximately the same d-spacing
of 3.40 nm, gelator 6 displays almost the same d-spacing values but ultimately shows an
apparent distortion of its network that can be seen in its d-spacing ratios from 1: 1/√2 given
by 1, 2 and 3, which is always characteristic of a cubic system. This situation can be
explained by the presence of bulky lateral methyl groups on both sides of the molecule.
Indeed, they are the main source of enhanced steric hindrance around the hydroxyl groups

Journal Pre-proof
of the molecule. Also, compared to other compounds, gelator 3 shows in its xerogel
networks a degree of penetration of the alkyl chain that is lower than usual. Without a
doubt, the additional inserted amide group acts as a wall preventing complete entry of alkyl
chains inside the gel networks as depicted on Figure 10.
4. Conclusion
The machine learning methods that were tested performed similarly well as predictive tools,
but one stood out in particular (SVM) to provide useful visualizations to help interpreting the
results. Gelators 1, 2 and 6 thus showed two clear and distinct gelation regions.
Independently of the gelation of water, glycerol led to a gel in many cases. Polyols of higher
viscosity such as mixtures of di- and tri glycerol could be studied in the same way. Indeed, a
remarkable disequilibrium was observed for gelator 4, with hydrophobic bonds dominating
the intermolecular hydrogen bonds in alcohols and aromatic solvents. Finally, the case of
gelator 5 provided a good demonstration of the importance of a well-thought molecular
design of gelator candidates. Even though all required elements were present in its structure,
the sequence was indeed not the most effective to reduce its solubility. However, remained
as a challenge for us, how to find a method for interpreting through a machine learning
prediction the impact of carbon chirality, especially for the peptide-based supramolecular
gelators, but also the characteristic values of logP or the rheology parameters contribution.
In summary, this present work introduced a good alternative, but we are currently trying to
improve our models.
Acknowledgment
The authors did not receive any financial support for this present work. All programs
employed in this work remain available from the authors on demand.
APPENDIX A. Supplementary data
Supplementary datas (1H NMR charts of gelators 3-6 and some photographs of gel samples)
associated with this article can be found in the online version, at…
References
1. F. Delbecq, Supramolecular gels from lipopeptide gelators: Template improvement
and strategies for the in-situ preparation of inorganic nanomaterials and for the
dispersion of carbon nanomaterials. Adv. Colloid and Interface Sci. 209 (2014) 98-
108.
2. F. Delbecq, Y. Masuda, Y. Ogue, T. Kawai, Salt complexes of two-component N-
acylamino acid diastereoisomers: self-assembly studies and modulation of gelation
abilities. Tetrahedron Lett. 53 (2012) 6588-6593.
3. F. Delbecq, K. Tsujimoto, Y. Ogue, T. Kawai, Synthesis of green chiral organogelators
with a sulfide linkage via Michael addition: soft templates for the preparation of
size-controlled gold nanoparticles. Tetrahedron Lett. 54 (2013) 651-656.
4. C. K. Rouse, A. D. Martin, C. J. Easton, P. Thordarson, A Peptide Amphiphile
Organogelator of Polar Organic Solvents. Nature Scientific Reports 7 (2017) 43668.

Journal Pre-proof
5. C. Morita, C. Kawai, A. Kikuchi, Y. Imura, T. Kawai, Effect of amide moieties for
hydrogelators on gelation property and heating-free pH responsive gel-sol phase
transition. J. Oleo. Sci. 61 (2012) 707-713.
6. C. Morita, H. Sugimoto, Y. Imura, T. Kawai, Double-stimuli Responsive O/W Emulsion
Gel Based on a Novel Amidoamine Surfactant. J. Oleo Sci. 60 (2011) 557-562.
7. C. Morita, H. Sugimoto, K. Matsue, T. Kondo, Y. Imura, T. Kawai, Changes in viscosity
behavior from a normal organogelator to a heat-induced gelator for a long-chain
amidoamine derivative. Chem. Commun. 46 (2010) 7969-7971.
8. C. Morita, C. Kawai, K. Tsujimoto, K. Kasai, Y. Ogue, Y. Imura, T. Kawai, High
organogelation ability and soft-templating for ultrathin Au nanowires of long-chain
amidoamine derivatives. J. Oleo Sci. 62 (2013) 81-87.
9. C. Morita, H. Tanuma, C. Kawai, Y. Ito, Y. Imura, T. Kawai, Room-Temperature
Synthesis of Two-Dimensional Ultrathin Gold Nanowire Parallel Array with Tunable
Spacing. Langmuir 29 (2013) 1669-1675.
10. F. Delbecq, P. Delfosse, G. Laboureix, C. Paré, T. Kawai, Study of a gelated Deep
Eutectic solvent metal salt solution as template for the production of size-controlled
small noble metal nanoparticles. Colloid and Surface A 567 (2019) 55-62.
11. F. Delbecq, I. Yamano, T. Kawai, A pH controlled Reversible Phase Transfer and
Electrolytic Size-Fractionation of Stable Silver Nanoparticles Capped with a Long
Amino Amide Dicarboxylic Acid. Bull. Chem. Soc. Jpn 88 (2015) 1719-1725.
12. J. Bonnet, G. Suissa, M. Raynal, L. Bouteiller, Organogel formation rationalized by
Hansen solubility parameters: influence of gelator structure. Soft Matter (2015)
2308-2313.
13. Y. Lan, M. G. Corradini, X. Liu, T. E. May, F. Borondics, R. G. Weiss, M. A. Rogers,
Comparing and Correlating Solubility Parameters Governing the self-Assembly of
Molecular Gels Using 1,3:2,4-Dibenzylidene Sorbitol as the Gelator. Langmuir 30
(2014) 14128-14142.
14. E. R. Draper, D. J. Adams, Low-Molecular-Weight Gels: The State of Art. Chem 3
(2017) 390-410.
15. M. Hansen, Hansen solubility parameters: a user’s handbook, 2nd ed. Boca Raton:
CRC Press, 2007.
16. K. K. Diehn, H. Oh, R. Hashemipour, R. G. Weiss, S. R. Raghavan, Insights into
organogelation and its kinetics from Hansen solubility parameters. Toward a priori
predictions of molecular gelation. Soft Matter 10 (2014) 2632-2640.
17. M. G. Corradini, M. A. Rogers, Molecular gels: improving selection and design
through computational methods. Curr. Opin. Food Sci. 9 (2016) 84-92.
18. N. Yan, Z. Xu, K. K. Diehn, S. R. Raghavan, Y. Fang, R. G. Weiss, How Do Liquid
Mixtures Solubilize Insoluble Gelators? Self-Assembly of Pyrenyl-Linker-Glucono
Gelators in Tetrahydrofuran-Water Mixtures. J. Am. Chem. Soc. 135 (2013) 8989-
8999.
19. J. K. Gupta, D. J. Adams, N. G. Berry, Will it Gel? Successful computational prediction
of peptide gelators using physicochemical properties and molecular fingerprints.
Chem. Sci. 7 (2016) 4713-4719.
20. J. Bonnet, G. Suissa, M. Raynal, L. Bouteiller, Organogel formation rationalized by
Hansen solubility parameters: dos and don’ts. Soft Matter 10 (2014) 3154-3160.

Journal Pre-proof
21. Y. Lan, M. G. Corradini, R. G. Weiss, S. R. Raghavan, M. A. Rogers, To gel or not to gel:
correlating molecular gelation with solvent parameters. Chem Soc Rev 44 (2015)
6035-6058.
22. M. A. Rogers, Hansen Solubility Parameters as a Tool in the Quest for New Edible
Oleogels. J Am Oil Chem Soc 2018, https://doi.org/10.1002/aocs.12050.
23. B. S. Lengeling, L. M. Roch, J. D. Perea, S. Langner, C. J. Brabec, A. Aspuru-Guzik, A
Bayesian approach to predict solubility parameters. Theory and Simulations 2 (2019),
1800069.
24. S. Boobier, A. Osbourn, J. B. O. Mitchell, Can human experts predict solubility better
than computers? J Cheminform 9 (2017) 63-76.
25. Y. Lan, M. G. Corradini, M. A. Rogers, Do Molecular Gelator Cluster in Hansen Space ?
Cryst. Growth Des 14 (2014) 4811-4818.
26. C. Liu, M. Corradini, M. A. Rogers, Self-assembly of 12-hydroxystearic acid molecular
gels in mixed solvent systems rationalized using Hansen solubility parameters.
Colloid Polym Sci 293 (2015) 975-983.
27. B. Hossin, K. Rizi, S. Murdan, Application of Hansen Solubility Parameters to predict
drug-nail interactions, which can assist the design of nail medicines. Eur. J. Pharm.
Biopharm 102 (2016) 32-40.
28. F. Li, J. Han, T. Cao, W. Lam, B. Fan, W. Tang, S. Chen, K. Lamfok, L. Li, Design of self-
assembly dipeptide hydrogel and machine learning via their chemical features. PNAS
116 (2019) 11259-112264.
29. B. M. Jamcel, A. Huynh, A. Chadha, S. Pamdey, J. Duncam, M. Chandler, G. Baki, Int. J.
Pharm 569 (2019) 118549.
30. A. Yan, J. Gasteiger, Prediction of Aqueous Solubility of Organic Compounds by
Topological Descriptors. QSAR Comb. Sci. 22 (2013) 821-829.
31. J. P. Janet, H. J. Kulik, Resolving Transition Metal Chemical Space: Feature Selection
for Machine Learning and Structure-Property Relationships. J Phys Chem A 121
(2017) 8939-8954.
32. M. M. Batista, R. Guirardello, M. A. Krahenbuhl, Determination of the Hansen
solubility Parameters of Vegetable Oils, Biodiesel, Diesel, and Biodiesel-Diesel Blends.
J Am Oil Chem Soc 92 (2015) 95-109.
33. M. A. Rogers, A. G. Marangoni, Kinetics of 12-Hydroxyoctadecanoic Acid SAFiN
Crystallization Rationalized Using Hansen Solubility Parameters. Langmuir 32 (2016)
12833-12841.

Journal Pre-proof
Gelation Properties of Various Long Chain Amido Amines:
Prediction of Solvent Gelation via Machine Learning using Hansen
Solubility Parameters
Frederic Delbecqa*, Guillaume Adenierb, Yuki Oguec, Takeshi Kawaic

Journal Pre-proof
AUTHORSHIP STATEMENT
Manuscript title: Gelation Properties of Various Long Chain Amido Amines: Prediction
of Solvent Gelation via Machine Learning using Hansen Solubility Parameters
All persons who meet authorship criteria are listed as authors, and all authors certify that they have
participated sufficiently in the work to take public responsibility for the content, including
participation in the concept, design, analysis, writing, or revision of the manuscript. Furthermore,
each author certifies that this material or similar material has not been and will not be submitted to
or published in any other publication before its appearance in the Journal of Molecular liquid.
We would like to declare the CrediT author statement as follow:
Frederic Delbecq Project administration, Investigation, Conceptualization, Writing (original draft and
review/editing) ; Yuki Ogue Investigation ; Guillaume Adenier Methodology, Software, Validation,
Data Curation ; Takeshi Kawai Conceptualization.

Journal Pre-proof
Highlights
-
-
-
-
Six different amphiphilic amidoamines were tested in 28 liquids for their gelation abilities.
These compounds could assemble into pseudo-cubic or lamellar-like structures.
The phenomenons were dependent both their alkyl chain length and polar head nature.
A machine learning study was also efficient to predict gelation of various liquids.