Showcasing research from the Nanotechnology National

Laboratory and Biorenewables National Laboratory at the As featured in:

Brazilian Center for Research in Energy and Materials,
São Paulo, Brazil. Volume 14
Number 47
21 December 2022
Pages 17471-17806

How lignin sticks to cellulose—insights from atomic force Nanoscale

microscopy enhanced by machine-learning analysis and rsc.li/nanoscale

molecular dynamics simulations

This work reports a combination of experimental, data-

driven, and theoretical approaches to investigate the
cohesive interactions between cellulose and lignin. A large
amount of experimental data and unsupervised machine
learning algorithms help us to get deeper insights into ISSN 2040-3372

cellulose/lignin interactions at the nanoscale level. This MINIREVIEW

Christel Herold-Mende, Carlos Romero-Nieto et al.
Boron- and phosphorus-containing molecular/nano

combined approach established here expands the potential

platforms: exploiting pathological redox imbalance to fight
cancer

of big data science and machine learning algorithms in See Carlos Driemeier,
AFM and can be applied to other systems where surface Juliana S. Bernardes et al.,
nanoscale heterogeneities play an important role in adhesive Nanoscale, 2022, 14, 17561.
interactions.

rsc.li/nanoscale
Registered charity number: 207890
 Nanoscale
Published on 02 November 2022. Downloaded by UNIVERSIDADE ESTADUAL DE CAMPINAS on 5/22/2025 7:05:10 PM.

View Article Online

PAPER View Journal | View Issue

How lignin sticks to cellulose—insights from

Cite this: Nanoscale, 2022, 14, 17561
atomic force microscopy enhanced by
machine-learning analysis and molecular
dynamics simulations†
Diego M. Nascimento, a Felippe M. Colombari, b Bruno Focassio, a,c
Gabriel R. Schleder, a,c Carlos A. R. Costa,a Cleyton A. Biffe,a Liu Y. Ling,b
Rubia F. Gouveia, a,c Mathias Strauss, a George J. M. Rocha, b
Edson Leite, a,d Adalberto Fazzio, a,c Rodrigo B. Capaz, a,e
Carlos Driemeier *b and Juliana S. Bernardes *a,c

Elucidating cellulose–lignin interactions at the molecular and nanometric scales is an important research
topic with impacts on several pathways of biomass valorization. Here, the interaction forces between a
cellulosic substrate and lignin are investigated. Atomic force microscopy with lignin-coated tips is
employed to probe the site-specific adhesion to a cellulose film in liquid water. Over seven thousand
force-curves are analyzed by a machine-learning approach to cluster the experimental data into types of
cellulose-tip interactions. The molecular mechanisms for distinct types of cellulose–lignin interactions
Received 6th October 2022, are revealed by molecular dynamics simulations of lignin globules interacting with different cellulose Iβ
Accepted 1st November 2022
crystal facets. This unique combination of experimental force-curves, data-driven analysis, and molecular
DOI: 10.1039/d2nr05541d simulations opens a new approach of investigation and updates the understanding of cellulose–lignin
rsc.li/nanoscale interactions at the nanoscale.

Introduction molecular components and acts as a barrier for enzymes tar-

geting depolymerization of the polysaccharides into mono-
One of the leading scientific and technological challenges of meric sugars. Many efforts have been devoted to addressing
the present time is the development of more sustainable pro- biomass recalcitrance, including plant genetic engineering7
ducts, increasing the interest in using lignocellulosic biomass and chemical approaches,8,9 but with partial success. For
from plants as a platform for renewable fuels, chemicals, and instance, while thermochemical pre-treatments have been
materials.1,2 Biomass of plant cell walls is a hierarchical com- widely recognized as a way to enhance biomass digestibility,
posite consisting of cellulose fibrils embedded in a matrix of lignin redeposition on cellulose surfaces is a bottleneck for
hemicelluloses and lignin.3 However, this complex structure4–6 polysaccharides breakdown since it prevents enzymatic
hinders the fractionation of biomass for isolation of its macro- saccharification.10,11
Both cellulose and lignin have a combination of hydro-
phobic and hydrophilic characters, which modulate their inter-
a
Brazilian Nanotechnology National Laboratory (LNNano), Brazilian Center for actions with each other and with water.12 Cellulose is a
Research in Energy and Materials (CNPEM), CEP 13083-970 Campinas, São Paulo, polymer of β(1 → 4)-linked glucose units, having a predomi-
Brazil. E-mail: [email protected] nant hydrophobic character in the axial direction of the sugar
b
Brazilian Biorenewables National Laboratory (LNBR), Brazilian Center for Research rings, whereas the equatorial distribution of hydroxyls pro-
in Energy and Materials (CNPEM), CEP 13083-970 Campinas, São Paulo, Brazil.
vides hydrophilicity. This molecular feature is reflected in the
E-mail: [email protected]
c
Center for Natural and Human Sciences, Federal University of ABC (UFABC), CEP formation of the native cellulose crystallites, where molecular
09606-070 Santo André, São Paulo, Brazil chains hydrogen bond to form molecular sheets, which are
d
Department of Chemistry, Federal University of São Carlos (UFSCAR), CEP packed by hydrophobic stacking to form tridimensional
13565905 São Carlos, São Paulo, Brazil
e
crystallites.13–15 As a consequence of this structure and –OH
Instituto de Física, Universidade Federal do Rio de Janeiro (UFRJ), CEP 21941-972
groups orientation, native cellulose crystallites have a set of
Rio de Janeiro, Rio de Janeiro, Brazil
† Electronic supplementary information (ESI) available. See DOI: https://doi.org/ facets with varying degrees of hydrophilicity.16–19 On the other
10.1039/d2nr05541d hand, lignin is a macromolecule made of monomeric units

This journal is © The Royal Society of Chemistry 2022 Nanoscale, 2022, 14, 17561–17570 | 17561
 View Article Online

Paper Nanoscale

(monolignols) with hydrophobic aromatic rings and oxyge- be probed through nanomechanical mapping using atomic
nated functional groups (e.g., hydroxyls, methoxyls, and force microscopy (AFM) force curves.28,29
others) that provide some hydrophilic character.20 The confor- Here, we employ a combination of experimental, data-
Published on 02 November 2022. Downloaded by UNIVERSIDADE ESTADUAL DE CAMPINAS on 5/22/2025 7:05:10 PM.

mation of the lignin macromolecule in water is impacted by driven, and theoretical approaches to investigate the cohesive
this amphiphilic balance, tending to form globules with interactions between cellulose and lignin. Atomic force
internal hydrophobic contacts and projection of the hydro- microscopy (AFM) with lignin-coated tips is used to probe the
philic functional groups towards the solvent.12 nanoscale adhesion to cellulose films. Generation of thou-
Recent studies have been addressing the cohesive inter- sands AFM force–distance (F–d ) curves and analysis by unsu-
actions between cellulose and lignin. A theoretical work found pervised machine-learning (ML) methods provide a novel
that despite both hydrophobic and hydrophilic faces of cell- approach to elucidate the nanoscale variability of cellulose–
ulose interacting with lignin through hydrogen bonds, the lignin interactions. Furthermore, molecular dynamics (MD)
mean contact area and the consequent reduction of water simulations are employed to investigate the molecular mecha-
molecules required to solvate the biopolymers is the major nisms of lignin adhesion to different facets of crystalline cell-
factor for the binding energy (an entropy-driven mecha- ulose Iβ constructed as a model compound of plant-based cell-
nism).21 As a result, lignin molecules coalesce and precipitate ulose. The combination of experiments, data-driven analysis,
as globules onto cellulose fibers during water-based pretreat- and simulations provide a new approach to investigate the
ment processes.22–24 Recent experiments have shown that the mechanisms and the site-specific energetics of lignin adhesion
inclusion of a moderately polar cosolvent, such as tetrahydro- to cellulose.
furan (THF)25,26 and γ-valerolactone (GVL),27 can reduce
biomass recalcitrance due to local solvent demixing, which
leads to the preferential solvation of the hydrophobic surfaces Results and discussion
of biomass by the cosolvent, thus promoting more efficient
removal of lignin molecules from the cellulose fibers, and pre- Nanoscale adhesion of lignin to cellulose
venting redeposition.21 However, energetic aspects of cellulose To probe and quantify the adhesion between cellulose and
and lignin interaction still await experimental validation con- lignin, we used AFM, a versatile high-resolution technique to
sidering the inherent nanoscale heterogeneities of cellulosic measure interaction forces at the nanoscale. First, we prepared
substrates, including the various cellulose crystalline facets tips with colloidal lignin deposits, schematically represented
exposed to the solvent. This thermodynamic information at in Fig. 1a. Scanning electron microscopy (SEM) images pre-
the nanoscopic level is not accessible through traditional bulk sented in Fig. 2a exemplify morphological changes after
techniques, such as isothermal titration calorimetry, but can coating. The tip radius increases from 23 to 40 nm for this

Fig. 1 Experimental setup for AFM measurements. (a) Schematic of tip coating with lignin. (b) Schematic of the approach-retract procedure to
measure F–d curves in the aqueous medium. (c) Typical F–d curves obtained in approach and retraction modes, in the case of a net adhesion
between the tip and surface. Labels in each panel correspond to positions in the retract and approach trajectories indicated in Fig. 1b.

17562 | Nanoscale, 2022, 14, 17561–17570 This journal is © The Royal Society of Chemistry 2022
 View Article Online

Nanoscale Paper
Published on 02 November 2022. Downloaded by UNIVERSIDADE ESTADUAL DE CAMPINAS on 5/22/2025 7:05:10 PM.

Fig. 2 Morphology of tips and substrates. (a) SEM images of a gold-coated AFM tip before (top) and after (bottom) lignin coating. (b) AFM topogra-
phy images and height profiles across the corresponding dashed lines.

specific cantilever, where lignin precipitates can also be cessive adhesion measurements (Fig. S3†). This lack of auto-
observed along the cantilever body. By repeating the coating correlation demonstrates that each F–d curve across the 1024
for other cantilevers, evidence of lignin deposition is also sequential measurements with the same lignin-coated tip can
detected, nevertheless with significant variations. For instance, be considered statistically independent. That is, one F–d
increment in the tip radius ranges from 17 to 162 nm measurement is not influenced by any memory of previous F–d
(Fig. S1†), leading to quantitatively different tip coatings to be curves measured along the scan.
considered in the statistical analysis of the F–d curves.
The microfibrillated cellulose sample (MFC) in this study Unsupervised machine-learning analysis of the force–distance
was isolated from bleached sugarcane bagasse pulp through curves
high-pressure homogenization. Fig. 2b shows the morphology The collected F–d curves present an evident variability in
of the MFC film measured by AFM in the wet state. As shape and adhesion force. The experiments with the MFC
observed from this micrograph and the corresponding line films produced 7156 analyzable force curves, each curve made
scan, the surface has root-mean-square (RMS) roughness (Rq) of 1600 data points. Analysis of such a dataset is a formidable
of 4.3 nm, which is much smaller than the radius of the lignin task for traditional statistical approaches. Therefore, we chose
coated tips, thus ensuring the tips contact the outer surface of to analyze these experimental results using a combination of
the films and do not penetrate deeply into the MFC web. For machine learning techniques.
comparison, we used SiO2/Si with a roughness (Rq) of 0.2 nm Our statistical analysis procedure is composed of three
as a flat substrate reference. steps, as detailed in the Methods section. The first step is a fil-
When lignin-coated tips are brought into contact with the tering process to reduce noise in the data, and the second con-
surfaces (Fig. 1b), adhesion may be established, and adhesion sists of unsupervised clustering of the F–d curves into families
force, adhesion work, and rupture length are measured as the based on similarity. Specifically, we cluster the force curves
tip retracts from the contact (Fig. 1c). By convention, we define into three families, chosen as the optimal balance between
such experimental adhesion forces as positive. The force information and granularity, as given by measuring the quality
measurements are all carried out under liquid water to elimin- of the clustering with the Variance Ratio Criterion (Calinski-
ate the contribution of capillary forces and mimic the native Harabasz index),32,33 see Fig. S4.† Clustering is applied separ-
environment. Therefore, analysis of the adhesive forces pro- ately for each lignin-coated tip. In this way, variations among
duces information about intermolecular forces between tip tips do not influence the clustering process, which is exempli-
and substrate, with contributions from the local effects of the fied in Fig. 3a for a specific lignin tip interacting with MFC.
liquid medium.26,30,31 The clustered force curves and cluster centers for each family
Fig. S2† shows representative maps of topography and are shown in light gray and blue, respectively. It is important
adhesion measured with a lignin-coated tip. There are no to highlight that the reliability of the clusters obtained by
signs of correlation between topography and adhesion. unsupervised algorithms must be judged with care because
Besides, there is no evidence of autocorrelation between suc- distinct metrics may provide different results.34 Therefore, vali-

This journal is © The Royal Society of Chemistry 2022 Nanoscale, 2022, 14, 17561–17570 | 17563
 View Article Online

Paper Nanoscale
Published on 02 November 2022. Downloaded by UNIVERSIDADE ESTADUAL DE CAMPINAS on 5/22/2025 7:05:10 PM.

Fig. 3 AFM pull-off force measurements. (a) Set of F–d curves during retraction. (b) Frequency histogram of adhesion force for MFC probed with
lignin-coated tips.

dation by experts may be recommended, and in this work the specificity of the MFC films on the interaction with the lignin-
obtained clustering is coherent with molecular dynamic simu- coated tip.
lations discussed in the next section.
Finally, in the third step, we extract adhesion forces from
the cluster centers. For this, we treat the cluster centers in two Free energy of lignin adsorption depending on the cellulose
cases: (i) when the representative force presents a clear facet
minimum, i.e., there is a net attractive adhesion force between To clarify the mechanism of adhesion observed in AFM
tip and surface, and (ii) when no such minimum occurs, i.e., measurements (Fig. 3b), we simulated the approximation of
there are purely net repulsive forces between the tip and lignin to different cellulose surfaces using atomistic molecular
surface (mediated by the liquid). In the first case, the adhesion dynamics (MD) simulations along with the Umbrella Sampling
force (defined as positive) is given by the absolute value of the (US) formalism. Four different crystal facets of cellulose Iβ
minimum force of the representative cluster center curve, as (Fig. 4a) were exposed to the lignin macromolecular model
seen in Fig. S5a.† In the second case, we obtain the repulsion (Fig. 4b). The MD simulations were carried out with wide cell-
force at the contact point, defined by the intersection point ulose Iβ facets (Fig. 5), so the lignin does not interact with crys-
between the two linear portions of the force curve,35 as shown tallite edges present in fibrils like the ones observed experi-
in Fig. S5b.† In such cases, we follow the convention that the mentally in Fig. 2b and represented schematically in Fig. 4a.
repulsive forces have negative signal. This approach allows to probe the association free energy and
Clustering analysis indicates an inverse relation between the local conformation of lignin onto ideal cellulose Iβ
the fraction of curves belonging to the cluster and the repre- surfaces.
sentative adhesion forces (Fig. 3b). Cluster C1 (blue bars, For all surfaces, we can observe an attractive pattern from
Fig. 3b) has higher adhesion forces and lower frequency, free energy profiles (Fig. 5a), starting with separation distances
whereas cluster C3 (red bars, Fig. 3b) is the most prevalent, (dcellulose–lignin) farther than 3.0 nm. However, the surfaces
but adhesive forces are the lowest, ranging from −0.32 to present distinct free energy profiles at closer distances. For the
0.14 nN. Cluster C2 (orange bars, Fig. 3b) presents inter- (100) facet, the surface with less exposure of OH groups, we
mediate numbers in both frequency and adhesion force. It is note three distinct free energy minima at 1.79, 1.58, and
noteworthy that Fig. 3b reports frequencies and adhesion 1.28 nm, for which ΔG values are −43.2 ± 1.5, −50.1 ± 4.6, and
forces accounting across the clustering performed with the −67.8 ± 4.9 kJ mol−1, respectively. On the other hand, for the
different lignin tips. The physical origins of the distinct clus- (010) surface, which has intermediate hydroxyl exposure, a
ters are presumably related to the nanoscale heterogeneity of broader free energy minimum region is detected with two
the MFC films, most likely due to the distinct facets of Iβ minima at 1.77 nm and 1.61 nm. Free energy values at these
crystalline cellulose, as further discussed in the following points are −31.8 ± 0.9 and −30.6 ± 0.9 kJ mol−1, respectively.
sections. The other cellulose surfaces, (110) and (11 ˉ0), present lower
For comparison, we also analyse the force curves of lignin- lignin adsorption energies, with global free energy minima of
covered tips onto a flat SiO2/Si surface. For this substrate, we −21.1 ± 1.1 kJ mol−1 at 1.77 nm (11 ˉ0) and −23.0 ± 1.4 kJ mol−1
group the force curves on a single cluster (see Fig. S6†). The at 1.53 nm (110). F–d curves obtained numerically (Fig. 5a)
adhesion forces range from 0.09 to 0.54 nN, demonstrating a representing the interaction with the different facets (Fig. 5b)
narrower range of adhesion forces reflecting the surface are consistent with this trend, presenting the most significant
chemical and topological homogeneity (Fig. 2). Contrary to the values of adhesion forces for the (100) and (010) surfaces: 0.24
MFC films, SiO2/Si surface does not show any representative nN (at 1.35 nm) and 0.12 nN (at 1.90 nm). On the other hand,
force curve with repulsive forces, further highlighting the the (11 ˉ0) and (110) surfaces present adhesion forces lower

17564 | Nanoscale, 2022, 14, 17561–17570 This journal is © The Royal Society of Chemistry 2022
 View Article Online

Nanoscale Paper
Published on 02 November 2022. Downloaded by UNIVERSIDADE ESTADUAL DE CAMPINAS on 5/22/2025 7:05:10 PM.

Fig. 4 Schematic of cellulose surfaces and lignin model used in molecular dynamics simulations. (a) Cellulose Iβ facets (010), (100), (11̄0), and (110)
from Iβ crystallites in the hexagonal cross-section model of cellulose fibril. (b) Model of softwood lignin composed of 40 monomers.

Fig. 5 Molecular dynamics Umbrella Sampling simulations for the approximation of lignin onto cellulose surfaces. (a) Binding free energy and
Forces, and (b) visualization of structures for each free energy minimum.

This journal is © The Royal Society of Chemistry 2022 Nanoscale, 2022, 14, 17561–17570 | 17565
 View Article Online

Paper Nanoscale

than 0.1 nN. These results confirm that lignin macromolecules of plant-based cellulose fibrils (which determines facet
prefer to bind at the less hydrophilic cellulose facet (100).21,36 exposure) has been a matter of recent experimental and theore-
Average intermolecular energy components for each tical investigations.53–55 There is agreement on the main cell-
ˉ0),
Published on 02 November 2022. Downloaded by UNIVERSIDADE ESTADUAL DE CAMPINAS on 5/22/2025 7:05:10 PM.

Umbrella Sampling window simulation were obtained to ulose Iβ exposed surfaces, namely (100), (010), (110), and (11
understand further the driving force behind the cellulose– which are the surfaces modeled in this study (see Fig. 4).
lignin binding free energies and attractive forces (Fig. 5a and However, the cross-section of cellulose crystallites and, there-
b). The intermolecular energies between water molecules fore, the relative abundance of the surfaces is a matter of
(Fig. S7†) reveal that the cellulose–lignin interaction is affected debate.
by hydration at the lignin–cellulose interface. For the (100) The relative hydrophilicity of cellulose surfaces is also a
surface, the interaction between water molecules – mainly due contentious topic. The traditional view considers the (100) a
to coulombic potential –is more favored when the lignin is hydrophobic surface, whereas the remaining (010), (110), and
close to this surface, as shown in Fig. S7† (third row). The (11ˉ0) are though hydrophilic due to the exposed cellulose
attractive interactions between water molecules lead to further hydroxyls.16,17 However, in a recent evaluation of the nano-
partial dehydration of the lignin–cellulose (100) interface, scale wetting of crystalline cellulose by MD simulation,19 the
increasing the solvation entropy. These results corroborate highly hydrophilic characters of (010) and (110) surfaces
that the stronger hydration regions of cellulose have a lower were confirmed, and so was the relatively less hydrophilic
tendency to associate with lignin.37 Our data thus support that nature of the (100) surface. The unexpected finding was the
the adhesion observed at the lignin–cellulose (100) surface is relative hydrophobicity of the (11 ˉ0) surface. This result is
mainly driven by dehydration of the lignin–cellulose interface. part of a broader debate on the differences between the Iβ
(110) and (11 ˉ0) surfaces, which are differentiated by the Iβ
AFM enhanced by machine-learning to probe nanoscale unit cell monoclinic angle slight deviation from 90° [γ = 96.5
heterogeneity of lignocellulosics in the crystal structure56]. For instance, preferred cellulose
AFM has been a key technique to probe the nanoscale hetero- cohesion parallel to the Iβ (110) surface has been proposed
geneity of lignocellulosics,38,39 cellulose nanomechanics,40 to result in ribbon-like cellulose cross-sections in wood-
surface adhesion forces,35,41,42 and the role of hydration layers based fibrillated cellulose.52 Despite this structural difference
on surface interactions.30,31,43 The approach developed here to between (110) and (11 ˉ0) facets, the free energies of lignin
examine lignin–cellulose adhesion adds two elements to the adsorption on their surfaces calculated in this work are very
repertoire of AFM analysis. First, coating the sharp tip with similar (Fig. 5).
lignin allows to locally measure the interaction with distinct In addition, a word of caution is needed when trying to
nanometric regions of cellulose due to the AFM high resolu- understand lignin–cellulose interactions based solely on
tion. Second, our approach brings the power of high-dimen- hydrophobicity/hydrophilicity concepts. A key feature leading
sional data analysis using machine-learning44 to investigate to the prominent hydrophilicity character of the (010) surface
full F–d curves rather than performing conventional univariate highlighted by MD simulations19 was the ability of small water
statistics based on selected F–d curve parameters.45 molecules to penetrate and probe the hydroxyl groups located
Machine-learning has been applied with AFM experiments in the trenches of the corrugated (010) surface. This is cer-
primarily for the classifications tasks such as identifying nano- tainly not the case for the large lignin molecule, as seen in
materials and molecules,46,47 different phases of a material48 Fig. 5. For this reason, our results indicating that the lignin on
or detecting the number of layers of 2D materials.49 These the (010) cellulose surface has intermediate adhesion between
tasks usually rely on height maps or amplitude/phase the hydroxyl-depleted (100) and hydroxyl-rich (110) and (11 ˉ0)
images.50 We provide a novel approach using unsupervised surfaces are not incompatible with the simulations by Trentin
machine-learning to analyse AFM F–d curves and extract et al.19
surface-tip interactions. Our methodology avoids costly image It is tempting to consider three groups of cellulose surfaces
processing, escapes the endeavour of extracting a few features and relate them to the three clusters of the lignin–cellulose
from the force curves,50 and enables the use of well-known force curves (Fig. 3). That is, C1 ≪ (100); C2 ≪ (010); and C3 ≪
and straightforward clustering algorithms that generate repre- (110)&(11 ˉ0). If this association is correct, the cluster frequencies
sentative force curves that are later interpreted. Additionally, (Fig. 3b) would indicate the probability of the lignin-coated tip
such methodology can be coupled with theoretical contacting each family of cellulose surfaces, providing an indir-
modeling51,52 or atomistic simulations, providing further ect measurement of the preferential facet exposures in the MFC
insights about the molecular mechanisms underlying the film. It follows in this interpretation that the hydrophilic (110)
surface nanoscale heterogeneities. and (11 ˉ0) cellulose surfaces are preferentially exposed, in line
with the results from Rosén et al.55 for wood-based fibrillated
Exposed cellulose facets and their hydrophilicity cellulose as well as observations in larger cellulose crystals from
The results of MD simulations (Fig. 5) clearly show that lignin algae.57 Nevertheless, this interpretation linking cellulose sur-
adhesion to cellulose is a function of the exposed cellulose faces and F–d curve clusters should be carefully judged. It is
crystal facet. For the experimental investigation, however, pre- likely instrumental to consider this interpretation as a hypoth-
ferential facet exposure is unknown. Indeed, the cross-section esis that deserves further investigation.

17566 | Nanoscale, 2022, 14, 17561–17570 This journal is © The Royal Society of Chemistry 2022
 View Article Online

Nanoscale Paper

Conclusions Preparation of cellulose films

Spin-coated films were produced with a homemade spin
Understanding interactions between plant cell wall com- coating system. First, silicon wafers (0.5 × 0.5 cm) were cleaned
Published on 02 November 2022. Downloaded by UNIVERSIDADE ESTADUAL DE CAMPINAS on 5/22/2025 7:05:10 PM.

ponents at the molecular and nano scales is crucial for using with a piranha solution (3 : 1 volume ratio of H2SO4 and H2O2
biomass as a platform for fuels, chemicals, and materials. solutions, respectively), rinsed with Milli-Q water, and ultra-
Here, we used a high-resolution technique (AFM) and unsuper- sonically cleaned with baths of acetone, isopropanol, and
vised machine-learning methods to gather experimental evi- Milli-Q water at 50 °C for 5 min in each solvent. Then, SiO2/Si
dence that lignin adhesion to cellulose films is highly variable wafers were activated for 2 min in an oxygen plasma cleaner
and analysable through clustering algorithms. Moreover, the (Plasma Technology PLAB SE80), spin-coated with an APTMS
association of experiments, data-driven analysis, and MD aqueous solution (0.005% w/v) at 3000 rpm for 40 s and
simulations suggests that the three clusters of cellulose–lignin washed three times with Milli-Q water at the same spinning
interactions result from the distinct cellulose crystal facets condition.59 Finally. MFC dispersions (1 mg mL−1) were de-
exposed in microfibrillated cellulose films. The combined posited on the wafer previously coated by APTMS and then
approach established here expands the potential of big data spin-coated at 4400 rpm for 45 s.60
science and machine learning algorithms in AFM and can be
applied to other systems where surface nanoscale heterogene-
Surface topography
ities play an important role in adhesive interactions.
Structural details of SiO2/Si and spin-coated cellulose films
were investigated using a Nanowizard IV AFM (JPK/Bruker).
Quantitative image (QI) mode imaging was performed in Milli-
Methods Q water using a MSNL-10 probe (Bruker) with a nominal
Materials spring constant 0.1 N m−1 and resonant frequency at ∼38 kHz.
All the chemicals used were of analytical reagent grade quality The scanned image size was 2 × 2 μm2.
and were employed as received. 3-Aminopropyltrimethoxysilane
(APTMS), acetone, ethanol, hydrogen peroxide (H2O2), isopro- Force–distance measurements
panol, Kraft lignin (KL), and sodium hydroxide (NaOH) were
F–d curves between the lignin-coated AFM tips and the silicon/
purchased from Sigma-Aldrich. Sodium hypochlorite 12 wt%
spin-coated cellulose substrates were performed in contact
(NaClO) was acquired from Star Flash.
mode using the same AFM instrument mentioned above. The
experiments were carried out under Milli-Q water and con-
Preparation of lignin-coated AFM tips
trolled ambient conditions (22 °C and 55% RH) (Fig. 1b). The
Seven gold AFM probes (ContGB-G probes, BudgetSensorsTM, spring constant of the cantilever was obtained by the thermal
nominal spring constant of 0.2 N m−1, and tip radius of curvature tune method under liquid conditions. F–d curves were
of ca. 25 nm) were cleaned in O2 plasma (XEI Scientific acquired as a function of the distance by applying the follow-
EVACTRON) for 4 min to remove any organic contaminants. ing conditions: (i) approaching speed of 1000 nm s−1, (ii)
Afterward, each tip was immersed in a drop of 0.1 wt% KL solution contact force of 1.0 nN, (iii) approach curve ramp size of
(acetone/water 9 : 1 v/v) for 5 min using a FlexAFM (Nanosurf) to 200 nm, (iv) dwell time (interval between the approaching and
control the dipping and pulling velocity. Solvent evaporation retraction stages) of 3 s, (v) retraction speed of 500 nm s−1, and
induced the deposition of lignin on the tip surface. Finally, the (vi) retract curve ramp size of 200 nm. Fig. 1b shows a typical
lignin-coated tips were dipped into Milli-Q water (18.2 MΩ cm) for approach–retraction cycle: at the beginning, the tip is moved
1 min and then dried in air at room temperature (Fig. 1a). directly toward the sample until contact is established. At this
point, the cantilever undergoes a deflection due to the sub-
AFM tip imaging strate and tip contact. As soon as contact force is reached, the
All AFM tips were imaged under a scanning electron micro- cantilever is retracted from the substrate, and the interaction
scope (FEI Quanta 650 FEG). All images were captured using a between the tip and sample is measured after the cantilever
5 kV electron accelerating voltage and a working distance returns to its undisturbed state (Fig. 1b). An array of 32 × 32
between 9–10 mm. (1024) F–d curves were obtained across a 10 μm × 10 μm scan
area for each lignin-coated cantilever.
Preparation of MFC
First, bleached cellulose pulp was extracted from sugarcane Unsupervised machine-learning analysis
bagasse by organosolv pretreatment followed by a hydrogen Our machine-learning analysis of F–d curves is composed of
peroxide bleaching step, as described elsewhere.58 Then, the three steps. For the first step (filtering), we use principal com-
bleached pulp (1 wt%) was passed thirty times through a high- ponent analysis (PCA) for noise reduction.61,62 Using the first
pressure homogenizer (Microfluidizer M-101P, Microfluidics 50 principal components, we successfully represent more than
Crop., USA) with a serial chamber pair (400–100 μm) to 99% of the data variance, reducing noise in F–d curves. This
produce microfibrillated cellulose (MFC). filtering process is performed independently for each set of F–

This journal is © The Royal Society of Chemistry 2022 Nanoscale, 2022, 14, 17561–17570 | 17567
 View Article Online

Paper Nanoscale

d curves obtained for a tip–substrate pair. We use the globular lignin structure was placed ca. 4.0 nm above the cell-
implementation given by the scikit-learn package.63 ulose surface (considering the separation of both centers of
The second step is to cluster data. We use k-means cluster- masses) and a total of 30 000 TIP3P water molecules were used
Published on 02 November 2022. Downloaded by UNIVERSIDADE ESTADUAL DE CAMPINAS on 5/22/2025 7:05:10 PM.

ing since this is a simple and computationally efficient cluster- to fill the simulation box containing cellulose and lignin.
ing method. Clustering data imply measuring the similarity Heavy atoms from the cellulose slabs were restrained hereafter
between data instances. We apply full dynamic time warping to avoid structural deformations. The system was minimized
(DTW)64 as a similarity measure between force curves, elimi- and equilibrated during 25 ns with a harmonic potential (k =
nating shifts in the position axis. The cluster centers are com- 5000 kJ−1 mol nm−1) to keep the separation distance between
puted as the barycenters of each group with respect to DTW, the cellulose slab and the lignin model. This equilibration
yielding an average representative shape of the force curves. step allows both the temperature and the density of the system
Barycenters are calculated using the DTW Barycenter to relax and allow the solvent molecules to relax around both
Averaging (DBA) method.65 The k-means algorithm is initia- cellulose and lignin structures.
lized by the “k-means++” method.66 We group data into three
clusters, as previously discussed, and take the final clustering Umbrella sampling (US) simulations
result as the best out of 10 initializations. We use the
The initial configurations for the US calculations were gener-
implementation given by the tslearn package.67
ated by pulling the lignin center of mass into the cellulose
Finally, in the third step, we apply the Savitzky–Golay
surface along a 5 ns steered molecular dynamics simulation at
filter68 to the representative force curves as an efficient way to
the NpT ensemble. This procedure sampled the lignin–cell-
eliminate noise and sudden force shifts. After this procedure,
ulose separation coordinate from ca. 4.0 nm to ca. 1.2 nm. For
the adhesion forces are extracted. More details on the
each cellulose surface, 32 snapshots of the lignin structure at
adhesion forces extraction are given in the ESI (Fig. S5†).
different positions along the z coordinate were taken and used
Molecular dynamics (MD) simulations as inputs for the US simulations.
Within the US formalism, the convergence of the free-
Atomistic molecular dynamics simulations were performed using
energy profile is treated by setting up N independent simu-
the Charmm36 force field69 and the Gromacs 2018.4 package.70,71
lations with the intermolecular distance constrained at
Coulomb interactions were calculated in real space up to 1.2 nm,
different z0 values ranging from zmin to zmax, using a harmonic
and the PME algorithm72,73 was used to treat long-range inter-
bias potential (eqn (1)) for each i-th umbrella window.81 Such
actions. Lennard–Jones interactions were switched off smoothly
methodology is necessary because the Gibbs free energy
between 1.0–1.2 nm. Energy minimization steps were performed
profile G(z) is dependent on the probability distribution P(z)
using the steepest descent algorithm until forces reached values
along the whole coordinate (eqn (2)), which is limited by the
below 100 kJ mol−1 nm−1. All molecular dynamics simulations
sampling of high-energy/low-probability regions.
were performed using a 2 fs timestep to integrate Newton
equations of motion altogether with the LINCS algorithm74 to U i ðzÞ ¼ kumb ðz  zi0 Þ2 ; for i ¼ 1; …; N ð1Þ
i
keep all bonds containing hydrogen atoms rigid. Simulations
were performed at the isothermal–isobaric ensemble (T = 300 K, GðzÞ ¼ kB T ln PðzÞ þ C ð2Þ
p = 1 bar) with an anisotropic pressure coupling to let the box This work used N = 32 umbrella windows with the lignin–cell-
vectors independently relax while keeping the angles fixed. In all ulose separation coordinate varying from ca. 4.0 nm to ca.
equilibration steps, Berendsen weak coupling algorithms75 were 1.2 nm. In each simulation window, the initial configuration
used for both temperature and pressure control. In all data collec- (snapshot) was submitted to an additional 5 ns equilibration
tion steps, both Nose–Hoover76,77 and Parrinello–Rahman78 algor- step followed by 100 ns data collection run, both using a force
ithms were used for temperature and pressure coupling, respect- constant of 5000 kJ−1 mol nm−1 for the harmonic bias poten-
ively. Detailed procedures for the whole simulation protocol are tial (see eqn (1) and Fig. S8†). As a result, a total of 3.2 micro-
presented below. All structural renderings were performed using second sampling for each cellulose surface was obtained and
the VMD 1.9.3 software.79 the free energy profiles were estimated using the weighted his-
System setup togram analysis method (WHAM).82 The numerical integration
procedure to obtain the free-energy profiles was performed con-
The cellulose surface models were obtained using the cell- sidering 200 bins between 1.2 nm and 3.9 nm. Statistical
ulose-builder toolkit.80 Each slab corresponded to 30 cellulose errors of each free energy profile were estimated using a
chains of 14 glucose residues that were subjected to energy Bayesian bootstrap analysis method83 with 50 bootstraps. F–d
minimization and equilibrated during 5 ns and then placed at curves were obtained by taking the negative of the numeric
the bottom of the xy plane of an 18 nm-long rectangular box. derivative of free energy profiles (eqn (3)), considering 1 kJ−1
The lignin model containing 40 monomeric units with a mol nm−1u1.6606 pN and using profiles smoothed by taking
linkage proportion close to one found in softwood (50% β-O-4′, a 10-point running average.
30% 5–5′, 10% α-O-4′) was placed in a 10 nm cubic box and
was solvated with 25 000 TIP3P water molecules prior to @GðN; p; T; zÞ
F ð zÞ ¼  ð3Þ
energy minimization and a 50 ns equilibration. The resulting @T

17568 | Nanoscale, 2022, 14, 17561–17570 This journal is © The Royal Society of Chemistry 2022
 View Article Online

Nanoscale Paper

Author contributions 10 H. Li, Y. Pu, R. Kumar, A. J. Ragauskas and C. E. Wyman,

Biotechnol. Bioeng., 2014, 111, 485–492.
J. S. B., C. D., E. L., A. F. and R. B. C. conceived and designed 11 J. K. Saini, A. K. Patel, M. Adsul and R. R. Singhania,
Published on 02 November 2022. Downloaded by UNIVERSIDADE ESTADUAL DE CAMPINAS on 5/22/2025 7:05:10 PM.

the study. D. M. N., C. A. B. and C. A. R. C. performed AFM Renewable Energy, 2016, 98, 29–42.
characterization and L. Y. L., B. F. and G. S. performed data 12 L. Petridis and J. C. Smith, Nat. Rev. Chem., 2018, 2, 382–
analysis. F. M. C. executed MD simulations. R. F. G., M. S., 389.
G. J. M. R. contributed to the data analysis and result 13 M. Jarvis, Nature, 2003, 426, 611–612.
interpretation. D. M. N., F. M. C., R. B. C., C. D. and J. S. B. 14 A. S. Gross and J.-W. Chu, J. Phys. Chem. B, 2010, 114,
wrote the manuscript. All authors approved the final version of 13333–13341.
the manuscript. 15 G. H. Silvestre, L. O. Pinto, J. S. Bernardes, R. H. Miwa and
A. Fazzio, J. Phys. Chem. B, 2021, 125, 3717–3724.
16 O. Biermann, E. Hädicke, S. Koltzenburg and F. Müller-
Conflicts of interest Plathe, Angew. Chem., Int. Ed., 2001, 40, 3822–3825.
17 K. Mazeau and A. Rivet, Biomacromolecules, 2008, 9, 1352–
There are no conflicts to declare. 1354.
18 A. N. Fernandes, L. H. Thomas, C. M. Altaner, P. Callow,
V. T. Forsyth, D. C. Apperley, C. J. Kennedy and M. C. Jarvis,
Acknowledgements Proc. Natl. Acad. Sci. U. S. A., 2011, 108, E1195–E1203.
19 L. N. Trentin, C. S. Pereira, R. L. Silveira, S. Hill, M. Sorieul
This study was supported by São Paulo Research Foundation and M. S. Skaf, Biomacromolecules, 2021, 22, 4251–4261.
(FAPESP, grants #14/50884-5, #16/04514-7, #17/02317-2, #19/ 20 F. M. Colombari, V. M. Nascimento, Y. L. Liu,
04527-0 and #20/07794-6), Rio de Janeiro Research Foundation G. J. M. Rocha and C. Driemeier, ACS Sustainable Chem.
(FAPERJ, grant #E-26/201.192/2021), Brazilian National Eng., 2022, 10, 10870–10878.
Council for Scientific and Technological (CNPq, grants 21 J. V. Vermaas, M. F. Crowley and G. T. Beckham, ACS
#300564/2022-8 and #380739/2019-4), INCT – Nanomateriais Sustainable Chem. Eng., 2019, 7, 19570–19583.
de Carbono and INCT –Bioetanol. The authors are thankful to 22 B. S. Donohoe, S. R. Decker, M. P. Tucker, M. E. Himmel
the National Laboratory for Scientific Computing (LNCC/ and T. B. Vinzant, Biotechnol. Bioeng., 2008, 101, 913–925.
MCTI, Brazil) for providing HPC resources of the SDumont 23 C. Driemeier, M. M. Oliveira and A. A. Curvelo, Ind. Crops
supercomputer, which have contributed to the research results Prod., 2016, 82, 114–117.
reported within this paper. https://sdumont.lncc.br. 24 S. V. Pingali, V. S. Urban, W. T. Heller, J. McGaughey,
H. O’Neill, M. Foston, D. A. Myles, A. Ragauskas and
B. R. Evans, Biomacromolecules, 2010, 11, 2329–2335.
References 25 A. S. Patri, B. Mostofian, Y. Pu, N. Ciaffone, M. Soliman,
M. D. Smith, R. Kumar, X. Cheng, C. E. Wyman and
1 L. R. Lynd, G. T. Beckham, A. M. Guss, L. N. Jayakody, L. Tetard, J. Am. Chem. Soc., 2019, 141, 12545–12557.
E. M. Karp, C. Maranas, R. L. McCormick, D. Amador- 26 S. V. Pingali, M. D. Smith, S.-H. Liu, T. B. Rawal, Y. Pu,
Noguez, Y. J. Bomble and B. H. Davison, Energy Environ. R. Shah, B. R. Evans, V. S. Urban, B. H. Davison and
Sci., 2022, 15, 938–990. C. M. Cai, Proc. Natl. Acad. Sci. U. S. A., 2020, 117, 16776–
2 T. Li, C. Chen, A. H. Brozena, J. Y. Zhu, L. Xu, C. Driemeier, 16781.
J. Dai, O. J. Rojas, A. Isogai, L. Wågberg and L. Hu, Nature, 27 H. Q. Lê, Y. Ma, M. Borrega and H. Sixta, Green Chem.,
2021, 590, 47–56. 2016, 18, 5466–5476.
3 M. C. McCann and N. C. Carpita, J. Exp. Bot., 2015, 66, 28 Y. F. Dufrêne, T. Ando, R. Garcia, D. Alsteens, D. Martinez-
4109–4118. Martin, A. Engel, C. Gerber and D. J. Müller, Nat.
4 T. J. Simmons, J. C. Mortimer, O. D. Bernardinelli, Nanotechnol., 2017, 12, 295–307.
A.-C. Pöppler, S. P. Brown, E. R. Deazevedo, R. Dupree and 29 R. Garcia, Chem. Soc. Rev., 2020, 49, 5850–5884.
P. Dupree, Nat. Commun., 2016, 7, 1–9. 30 Z. Zhang, S. Ryu, Y. Ahn and J. Jang, Phys. Chem. Chem.
5 X. Kang, A. Kirui, M. C. Dickwella Widanage, F. Mentink- Phys., 2018, 20, 30492–30501.
Vigier, D. J. Cosgrove and T. Wang, Nat. Commun., 2019, 10, 31 S. Sinniah, A. Steel, C. Miller and J. Reutt-Robey, J. Am.
1–9. Chem. Soc., 1996, 118, 8925–8931.
6 A. Kirui, W. Zhao, F. Deligey, H. Yang, X. Kang, F. Mentink- 32 T. Caliński and J. Harabasz, Commun. Stat. Theory Methods,
Vigier and T. Wang, Nat. Commun., 2022, 13, 1–12. 1974, 3, 1–27.
7 M. Sticklen, Curr. Opin. Biotechnol., 2006, 17, 315–319. 33 P. J. Rousseeuw, J. Comput. Appl. Math., 1987, 20, 53–65.
8 V. B. Agbor, N. Cicek, R. Sparling, A. Berlin and D. B. Levin, 34 M. Wegmann, D. Zipperling, J. Hillenbrand and
Biotechnol. Adv., 2011, 29, 675–685. J. Fleischer, arXiv, preprint, arXiv:2106.12792.
9 K. H. Kim and C. G. Yoo, Front. Chem. Eng., 2021, 3, 35 H.-J. Butt, B. Cappella and M. Kappl, Surf. Sci. Rep., 2005,
785709. 59, 1–152.

This journal is © The Royal Society of Chemistry 2022 Nanoscale, 2022, 14, 17561–17570 | 17569
 View Article Online

Paper Nanoscale

36 S. Besombes and K. Mazeau, Plant Physiol. Biochem., 2005, 60 A. Griffo, B. J. Rooijakkers, H. Hähl, K. Jacobs, M. B. Linder
43, 277–286. and P. Laaksonen, Biomacromolecules, 2019, 20, 769–777.
37 B. Lindner, L. Petridis, R. Schulz and J. C. Smith, 61 K. Pearson, London, Edinburgh Dublin Philos. Mag. J. Sci.,
Published on 02 November 2022. Downloaded by UNIVERSIDADE ESTADUAL DE CAMPINAS on 5/22/2025 7:05:10 PM.

Biomacromolecules, 2013, 14, 3390–3398. 1901, 2, 559–572.

38 J. Fahlén and L. Salmén, Biomacromolecules, 2005, 6, 433– 62 I. T. Jolliffe, Principal component analysis for special types of
438. data, Springer, 2002.
39 E. Lambert, V. Aguié-Béghin, D. Dessaint, L. Foulon, 63 F. Pedregosa, G. Varoquaux, A. Gramfort, V. Michel,
B. Chabbert, G. Paës and M. Molinari, Biomacromolecules, B. Thirion, O. Grisel, M. Blondel, P. Prettenhofer, R. Weiss
2019, 20, 515–527. and V. Dubourg, the J. Mach. Learn. Res., 2011, 12, 2825–
40 P. N. Ciesielski, R. Wagner, V. S. Bharadwaj, J. Killgore, 2830.
A. Mittal, G. T. Beckham, S. R. Decker, M. E. Himmel and 64 H. Sakoe and S. Chiba, IEEE Trans. Acoust., Speech, Signal
M. F. Crowley, Proc. Natl. Acad. Sci. U. S. A., 2019, 116, Process., 1978, 26, 43–49.
9825–9830. 65 F. Petitjean, A. Ketterlin and P. Gançarski, Pattern
41 C. Qin, K. Clarke and K. Li, Biotechnol. Biofuels, 2014, 7, 1– Recognition, 2011, 44, 678–693.
10. 66 D. Arthur and S. Vassilvitskii, k-means++: The advantages of
42 B. Arslan, K. Egerton, X. Zhang and N. I. Abu-Lail, careful seeding, Stanford, 2006.
Langmuir, 2017, 33, 6857–6868. 67 R. Tavenard, J. Faouzi, G. Vandewiele, F. Divo, G. Androz,
43 A. Noy, D. V. Vezenov and C. M. Lieber, Annu. Rev. Mater. C. Holtz, M. Payne, R. Yurchak, M. Rußwurm and K. Kolar,
Sci., 1997, 27, 381–421. J. Mach. Learn. Res., 2020, 21, 1–6.
44 G. R. Schleder, A. C. Padilha, C. M. Acosta, M. Costa and 68 A. Savitzky and M. J. Golay, Anal. Chem., 1964, 36, 1627–
A. Fazzio, J. Phys. Mater., 2019, 2, 032001. 1639.
45 Y. Chen, H. J. Busscher, H. C. van der Mei and W. Norde, 69 J. Huang and A. D. MacKerell Jr., J. Comput. Chem., 2013,
Appl. Environ. Microbiol., 2011, 77, 5065–5070. 34, 2135–2145.
46 J. Carracedo-Cosme, C. Romero-Muñiz and R. Pérez, 70 M. J. Abraham, T. Murtola, R. Schulz, S. Páll,
Nanomaterials, 2021, 11, 1658. J. C. Smith, B. Hess and E. Lindahl, SoftwareX, 2015, 1,
47 H. Bai and S. Wu, Nanotechnol. Precis. Eng., 2021, 4, 19–25.
013002. 71 D. Van Der Spoel, E. Lindahl, B. Hess, G. Groenhof,
48 F. Cellini, F. Lavini, C. Berger, W. De Heer and E. Riedo, 2D A. E. Mark and H. J. Berendsen, J. Comput. Chem., 2005, 26,
Mater., 2019, 6, 035043. 1701–1718.
49 M. Xu, B. Tang, Y. Lu, C. Zhu, Q. Lu, C. Zhu, L. Zheng, 72 T. Darden, D. York and L. Pedersen, J. Chem. Phys., 1993,
J. Zhang, N. Han and W. Fang, J. Am. Chem. Soc., 2021, 143, 98, 10089–10092.
18103–18113. 73 U. Essmann, L. Perera, M. L. Berkowitz, T. Darden, H. Lee
50 B. Huang, Z. Li and J. Li, Nanoscale, 2018, 10, 21320–21326. and L. G. Pedersen, J. Chem. Phys., 1995, 103, 8577–8593.
51 W. Fu and W. Zhang, Phys. Chem. Chem. Phys., 2018, 20, 74 B. Hess, H. Bekker, H. J. Berendsen and J. G. Fraaije,
24434–24443. J. Comput. Chem., 1997, 18, 1463–1472.
52 K. L. Johnson, K. Kendall and A. D. Roberts, Proc. R. Soc. 75 H. J. Berendsen, J. van Postma, W. F. Van Gunsteren,
London, Ser. A, 1971, 324, 301–313. A. DiNola and J. R. Haak, J. Chem. Phys., 1984, 81, 3684–
53 J. D. Kubicki, H. Yang, D. Sawada, H. O’Neill, D. Oehme 3690.
and D. Cosgrove, Sci. Rep., 2018, 8, 13983. 76 S. Nosé, J. Chem. Phys., 1984, 81, 511–519.
54 K. Daicho, T. Saito, S. Fujisawa and A. Isogai, ACS Appl. 77 W. G. Hoover, Phys. Rev. A, 1985, 31, 1695.
Nano Mater., 2018, 1, 5774–5785. 78 M. Parrinello and A. Rahman, J. Appl. Phys., 1981, 52,
55 T. Rosén, H. He, R. Wang, C. Zhan, S. Chodankar, A. Fall, 7182–7190.
C. Aulin, P. T. Larsson, T. Lindstrom and B. S. Hsiao, ACS 79 W. Humphrey, A. Dalke and K. Schulten, J. Mol. Graphics,
Nano, 2020, 14, 16743–16754. 1996, 14, 33–38.
56 Y. Nishiyama, J. Sugiyama, H. Chanzy and P. Langan, J. Am. 80 T. C. Gomes and M. S. Skaf, J. Comput. Chem., 2012, 33,
Chem. Soc., 2003, 125, 14300–14306. 1338–1346.
57 J. Lehtiö, J. Sugiyama, M. Gustavsson, L. Fransson, 81 B. Reischl, M. Watkins and A. S. Foster, J. Chem. Theory
M. Linder and T. T. Teeri, Proc. Natl. Acad. Sci. U. S. A., Comput., 2013, 9, 600–608.
2003, 100, 484–489. 82 S. Kumar, J. M. Rosenberg, D. Bouzida, R. H. Swendsen
58 L. O. Pinto, J. S. Bernardes and C. A. Rezende, Carbohydr. and P. A. Kollman, J. Comput. Chem., 1992, 13, 1011–
Polym., 2019, 218, 145–153. 1021.
59 M. A. Herrera, A. P. Mathew and K. Oksman, Carbohydr. 83 J. S. Hub, B. L. De Groot and D. Van Der Spoel, J. Chem.
Polym., 2014, 112, 494–501. Theory Comput., 2010, 6, 3713–3720.

17570 | Nanoscale, 2022, 14, 17561–17570 This journal is © The Royal Society of Chemistry 2022