Classification and Characterization of Internal Structure

of Incipient Soot from Acetylene Pyrolysis obtained via

Molecular Dynamics Simulations
Khaled Mosharraf Mukuta , Anindya Gangulyb , Eirini Goudelib , Georgios
Kelesidisc,d , Somesh Roya,∗
a
Department of Mechanical Engineering, Marquette
University, Milwaukee, Wisconsin, 53233, USA
b
Department of Chemical Engineering, University of Melbourne, Victoria, Australia
c
Particle Technology Laboratory, Institute of Process Engineering, Department of
Mechanical and Process Engineering, ETH Zürich, Sonneggstrasse 3, CH-8092
Zürich, Switzerland
d
Department of Environmental and Occupational Health and Justice, Rutgers School of
Public Health, New Brunswick, New Jersey, USA

Abstract
A series of reactive molecular dynamics simulations is used to study the
internal structure of incipient soot particles obtained from acetylene pyroly-
sis. The simulations were performed using ReaxFF potential at four different
temperatures. The resulting soot particles are cataloged and analyzed to ob-
tain statistics of their mass, volume, density, C/H ratio, number of cyclic
structures, and other features. A total of 3324 incipient soot particles were
analyzed in this study. Based on the features obtained, the incipient soot
particles were found to be well-suited to be classified into two classes, referred
to as type 1 and type 2 incipient soot particles in this work. The radial distri-
bution of density, 5-/6-/7-member ring structures and C/H ratio inside the
particles revealed a clear difference in the internal structure between type 1
and type 2 particles. These classes were further found to be well represented
by the size of the particles with smaller particles in type 1 and larger particles
in type 2. The radial distributions of ring structures, density, and C/H ratio
indicated the presence of a dense core region in type 2 particles, whereas no
clear evidence of the presence of a core was found in type 1 particles. The

∗
Corresponding author: [email protected]

Preprint submitted to Carbon December 1, 2023

 boundary between the core and shell regions in type 2 incipient soot particles
was also identified.
Keywords: Soot, Molecular Dynamics, Core-Shell structure, ReaxFF,
Machine learning

1 1. Introduction
2 Soot is a harmful carbonaceous nanoparticle generated during combus-
3 tion of hydrocarbon fuels. Soot, also known as black carbon, can cause
4 serious health issues [1, 2] and acts as a major forcing factor in climate
5 change[3, 4]. The exact mechanism of the formation of soot particulates
6 from gaseous precursors is still unknown due to the complex chemical na-
7 ture of the hydrocarbon system and time- and length-scale of the soot for-
8 mation processes. According to the present understanding, soot formation
9 occurs by a series of complex physicochemical events such as the formation
10 of gas-phase soot precursors including, but not limited to, polycyclic aro-
11 matic hydrocarbons (PAHs), nucleation of incipient soot particles, growth
12 and maturation of incipient soot particles due surface reactions, aggregate
13 formation via coagulation and coalescence, and decay of the particles by
14 fragmentation and oxidation [5, 6, 7, 8, 9]. The inception of soot particles is
15 arguably the least understood phenomenon among these processes and the
16 exact chemical pathways of soot inception are not completely known yet.
17 Researchers agree that soot formation starts with forming small gas-phase
18 precursor molecules such as acetylene which leads to PAHs like benzene,
19 pyrene, and coronene [10, 11, 12]. The freshly formed PAHs then combine to
20 form the solid or liquid-like incipient soot particles [13, 14, 15]. Such formed
21 particles then start to grow due to surface interactions and coalescence to
22 form larger soot particles [13, 16, 17, 18, 19].
23 Due to the complexity and scale of the incipient particles, their exact
24 internal structures are not very well characterized yet. Recent studies have
25 shown incipient soot particles tend to have a condensed core of ring-like struc-
26 tures surrounded by a shell of chain-like structures [20]. As these incipient
27 particles mature, their internal structures evolve, which in turn affects their
28 physical and chemical properties.
29 There have been some recent breakthroughs in the experimental explo-
30 ration of the internal structure of incipient soot. For example, Chang et
31 al. [21] employed high-resolution transmission electron microscopy (HRTEM)

2
32 and scanning electron microscopy (SEM) to investigate the structural evolu-
33 tion and fragmentation of coal-derived soot and carbon black particles under
34 high-temperature air oxidation conditions. They also explored the onset
35 of micropores and the internal graphitic microcrystals using X-ray diffrac-
36 tion (XRD) and Raman spectra. Morajkar et al. [22] utilized HRTEM,
37 XRD, Raman spectroscopy, and inductively coupled plasma mass spectrom-
38 etry (ICPMS) to examine the transmission of trace metals from biodiesel
39 fuels to soot particles and the nanostructural irregularities of the soot. In
40 their study, Gleason et al. [23] indicated that the formation of soot nuclei in
41 an ethylene/nitrogen flame can be attributed exclusively to aromatic com-
42 pounds comprising one or two rings. Carbone et al. [24] conducted a compre-
43 hensive investigation of soot inception in a laminar premixed ethylene flame
44 and found that soot particles undergo an aging transformation from being
45 nearly transparent in the visible spectrum to a more graphitic-like compo-
46 sition. Using low-fluence laser desorption ionization (LDI) in conjunction
47 with HRTEM, Jacobson et al. [25], investigated the molecular composition
48 of soot particles to determine the PAH concentration in soot particles. The
49 capabilities of atomic force microscopy (AFM) were exploited by Barone et.
50 al. [26] to calculate particle size distribution functions under different sam-
51 pling conditions. In their study, Schulz et al. [27] conducted an investigation
52 into the initial phases of soot formation using AFM and observed the pres-
53 ence of multiple aromatic compounds, some of which displayed noticeable
54 aliphatic side chains. Commodo et al. [28] investigated the initial phases of
55 soot formation using X-ray, ultraviolet photoemission spectroscopy (UPS),
56 UV-visible, and Raman spectroscopy to show the coexistence of sp3 carbon
57 and a more advanced graphitic structure, which exhibits a slightly larger
58 aromatic island, a reduced band gap, and an increased density of states.
59 In another study, they [29] identified a noteworthy occurrence of aliphatic
60 pentagonal rings in the early stages of soot formation, particularly in close
61 proximity to the outer region of aromatic soot molecules, and it has been
62 suggested that the elimination of hydrogen from these molecules can result
63 in the creation of resonantly stabilized π-radicals [30]. This phenomenon has
64 also been theorized by Johansson et al. [31], Gentile et al. [32], and Rundel
65 et al. [33].
66 Even with such recent advancement in experimental findings, there is
67 still a lot of unknowns about the internal structure of soot. The limita-
68 tions of experimental methods can be compensated and complemented by
69 first-principle modeling such as molecular dynamics. With the development

3
 70 of high-performance computational resources, reactive molecular dynamics
71 (RMD) simulation has become more affordable for studying complex reactive
72 networks. For soot-relevant RMD studies, the reactive force field (ReaxFF)
73 potential developed by van Duin et al. [34] for carbon, hydrogen, and oxygen
74 chemistry (CHO-parameters [35, 36]) is a popular choice. The ReaxFF poten-
75 tial can capture the physicochemical evolution of hydrocarbon systems in an
76 extensive range of temperatures and pressures. It is based on the bond order
77 between different atoms, which carry valuable information related to bond
78 breakage and formation. In recent years, RMD simulations have been suc-
79 cessfully used to investigate soot nucleation using pyrene dimerization [37],
80 to shed light into the nucleation and growth of incipient soot using different
81 PAHs, e.g. naphthalene, pyrene, coronene, ovalene and circumcoronene [16],
82 to explore the initial mechanism of soot nanoparticle formation [38] and to
83 examine the effect of oxygenated additives on diesel soot reduction [39].
84 Since RMD simulation provides detailed structural information at the
85 atomic scale, it can be an excellent tool for analyzing the internal structure
86 of incipient soot particles. For example, recently Pascazio et al. [40] looked
87 into the internal structure and the mechanical properties of incipient soot
88 particles using RMD simulation and quantified the amount of cross-linking
89 in the core and shell region of developing and mature soot particles. Mature
90 soot primary particles exhibit a distinct core-shell structure with a disorderly
91 condensed core of ring-like structures surrounded by a shell of chain-like
92 structures [20, 41].
93 In this work, we utilize RMD to characterize the internal structure of in-
94 cipient soot particles in further detail at different process temperature. Pro-
95 cess temperature plays an important role in the development and maturation
96 of soot particles. For example, in a recent study, Pathak et. al. [42] studied
97 graphitization induced structural transformation of candle soot at different
98 temperatures and found that increasing the temperature increases the rate
99 of graphitization that leads to more spherical and mature soot and weaken-
100 ing of the correlation between graphitic nanostructure and surface functional
101 groups (SFGs). SFGs have been found to be connected to the characteristics
102 of soot aggregates, including the fractal dimension [43]. Since soot morphol-
103 ogy, maturity and reactivity are expected to be influenced by temperature
104 during acetylene pyrolysis [44], it is important to study the internal structure
105 of incipient soot particles at different temperatures.
106 In the present study, a series of isothermal RMD simulations using the
107 ReaxFF potential is conducted mimicking acetylene pyrolysis at different

4
108 temperatures (1350, 1500, 1650, and 1800K). A variety of physicochemical
109 features of these RMD-generated soot particles are then analyzed to char-
110 acterize the internal evolution of the incipient soot particle. The focus of
111 this article is to shed light on different types of incipient soot particles and
112 to characterize the internal structure of these particles obtained from RMD
113 simulations.

114 2. Numerical Methodology

115 2.1. Simulation configurations
116 Following the methodology described in [45], 1000 acetylene molecules
117 are randomly placed in a cubic domain (75Å × 75Å × 75Å) at four different
118 temperatures, i.e., 1350 K, 1500 K, 1650 K, and 1800 K. The temperatures
119 are chosen to capture soot particles from various thermally activated sys-
120 tems. For statistical significance, multiple instances of identical simulations
121 are performed for up to five times with different initial conditions. In total,
122 24 RMD simulations were performed for four different temperatures. The
123 RMD simulations are done using the Large-scale Atomic/Molecular Mas-
124 sively Parallel Simulator (LAMMPS) [46] software. ReaxFF potential for
125 hydrocarbons [34, 47] is used to capture the chemical changes (bond break-
126 age and formation) during the molecular collisions between the acetylene
127 molecules. The bond length between individual atoms is calculated at each
128 timestep (0.25 fs) based on the changes in the chemical environment to de-
129 scribe bond cleavage and formation accurately [36]. This helps the model
130 capture the chemical reactions leading to radical formation during soot nu-
131 cleation. Periodic boundary conditions are assumed in all three dimensions.
132 The coordinates of each atom are calculated and updated using the velocity-
133 Verlet algorithm [48] in conjunction with the Nose-Hoover thermostat [49].
134 A constant number, volume, and temperature (NVT) ensemble strategy are
135 used to run each simulation up to 10 ns. The simulation results are probed
136 every 0.10 ns, and the meaningful clusters of hydrocarbons that resembles
137 primary soot particles are isolated, cataloged, and analyzed. Features such
138 as surface area and volume of primary particles are calculated using MSMS
139 software developed by Sanner [50] and other physicochemical characteristics
140 are analyzed mostly using MAFIA-MD [51]. The open visualization tool
141 (OVITO) [52] is used for visualization of the molecular clusters.

5
142 2.2. Workflow
143 The workflow in this study can be summarized as
144 1. Conduct RMD simulations at different temperature with different ini-
145 tializations and extract incipient soot particles from the trajectory re-
146 sults.
147 2. Calculate chemical and morphological characteristics such as the num-
148 ber of atoms, C/H ratio, the radius of gyration, atomic fractal dimen-
149 sion, density, surface area, and volume.
150 3. Classify the soot particles based on all the calculated features using
151 machine learning techniques such as k-means clustering [53] and t-
152 distributed stochastic neighbor embedding (t-SNE) [54].
153 4. Investigate the internal distribution of several relevant features such as
154 distribution of cyclic/non-cyclic molecules, C/H ratio, density etc. and
155 find identifiable patterns in the distribution.

156 2.3. Extraction of physicochemical properties

157 From the RMD simulations, we extract the coordinates of each atom
158 present in the simulation box at regular time intervals via the trajectory
159 file. Each individual timestep is investigated separately by analyzing the
160 atom coordinates within the entire simulation domain, which contains both
161 large molecular clusters and small molecules. The large soot-like molecular
162 clusters are identified as the ones that have more than 20 carbon atoms
163 and have at least one 5-, 6-, or 7-member ring [55]. It is noted that in our
164 case, the smallest such cluster was found to have 65 carbon atoms. These
165 clusters are isolated using the cluster analysis tool from the OVITO Python
166 module [52] implemented in a unified Python script developed inhouse. Then
167 the isolated clusters are analyzed individually to calculate their physical,
168 morphological and chemical attributes. Some attributes are obtained trivially
169 from the trajectory files, e.g., number of atoms (N ), carbon to hydrogen
170 ratio (Θ ), mass (Mp ), and molar mass (M ). Some other attributes like
C/H

171 the radius of gyration (Rg ), atomic fractal dimension (Df ), and density (ρ)
172 are extracted by simple algebraic and geometric analysis or by using some
173 empirical correlations proposed in the literature and listed in Appendix A.
174 The volume and the surface area of incipient particles are calculated using
175 MSMS software [50] using a probe radius of 1.5 Å.
176 The identification and analysis of 5- /6- /7-member ring structures are
177 done using MAFIA-MD [51]. MAFIA-MD can analyze RMD trajectory files

6
178 to identify cyclic/ring structures in an atomic cluster. Not all cyclic struc-
179 tures identified are necessarily aromatic. As discussed in [51], it is difficult
180 to exactly confirm which cyclic structures are aromatic as the information
181 about aromaticity requires some approximations regarding the bond order
182 of aromatic bonds and establishment of planarity. To remove any confusion,
183 therefore, we used the terms “ring” or “cyclic” in this work instead of aro-
184 matic when discussing these internal structures in the soot clusters. The
185 numbers of 5-, 6-, and 7-member rings are denoted as N5 , N6 , N7 , respec-
186 tively, and the total number of rings is denoted as N⃝ . Similarly, the number
187 of carbons in rings is denoted as N© and the number of non-cyclic carbons
188 in a particle is denoted as N¢ .
189 A sample of two particles and a list of their respective properties are
190 provided in Appendix B. This entire set of features is used in the classification
191 of particles as discussed in Sec. 3.2. It must be noted here that while all the
192 above-mentioned properties were evaluated for each particle, this article only
193 focuses on the internal structure of the particles, which are characterized as
194 discussed in Sec. 2.4. Therefore, beyond their use in the classification of
195 particles, the detailed analysis of physical and morphological features such
196 as volume, surface area, radius of gyration, and atomic fractal dimension are
197 not the focus of this work.

198 2.4. Characterization of internal structure

199 We analyzed the internal structure of soot particles via the radial distri-
200 bution of carbon atoms, C/H ratio, and density inside the particle. In order
201 to compare different-sized particles on the same scale, we first normalized
202 the radius of particles by scaling each particle by its radius of gyration (Rg ).
203 Then each particle is divided in equal number of radial bins. Each such radial
204 bin creates a spherical shell or strip as shown by the shaded yellow region of
205 interest in Fig. 1. We calculate various internal features in these spherical
206 strips and present them as a function of the normalized radial distance from
207 the center of mass of each spherical strip (r∗ = r/Rg ).
208 For example, let’s consider a strip of width ∆r, whose midplane is r
209 distance away from the center of mass. The number of cyclic and non-
210 cyclic carbon atoms in this strip (i.e., within a radial distance of r ± ∆r/2) are
211 counted and represented as functions of the normalized radius of the midplane
212 (r∗ = r/Rg ), as n© (r∗ ) and n¢ (r∗ ), respectively. The radial distribution of
213 cyclic and non-cyclic carbon per unit area (indicated by ′′ ) at a normalized

7
 Figure 1: Schematic representation of the calculation of radial distribution of internal
features in the soot particle.

214 distance r∗ are then, respectively:

′′ n© (r∗ ) n¢ (r∗ )
N© (r∗ ) = ; N¢′′ (r∗ ) = (1)
4π(r∗ × Rg )2 4π(r∗ × Rg )2
215 In a similar manner, the radial distribution of C/H ratio is also ana-
216 lyzed for each particle. The C/H ratio of the entire particle is calculated as
217 Θ = NC/NH , where NC and NH are, respectively, number of carbon and
C/H

218 hydrogen atoms in the entire particle. This C/H ratio is termed as the parti-
219 cle C/H ratio (Θ ) to differentiate from the local C/H ratio (θ ), which is
C/H C/H

220 calculated using the number of carbon and hydrogen atoms in the spherical
221 strips as shown in Fig. 1. The local C/H ratio (θ ) is calculated by cal-
C/H

222 culating the number of carbon (nC (r∗ )) and hydrogen (nH (r∗ )) atoms in a
223 spherical strip with the midplane at a normalized distance r∗ from the center
224 of mass (θ (r∗ ) = nC (r∗ )/nH (r∗ )). Finally, the local C/H ratio is normalized
C/H

225 by the corresponding particle C/H ratio

∗ ∗ θ (r∗ )
C/H
θ (r ) = (2)
C/H
Θ C/H

226 Similarly, the radial distribution of local density is also evaluated by di-
227 viding the actual density of the thin spherical strip using Eqn. A.3 for the
228 strip (referred as local density, ϱa (r∗ )) by the actual density of the particle
229 (ρa ) as
ϱa (r∗ )
ρ∗a (r∗ ) = (3)
ρa

8
230 3. Results and Discussion
231 3.1. Formation of incipient soot particles in RMD
232 During the RMD simulations, the system of atoms goes through differ-
233 ent chemical and physical interactions resulting in the formation of larger
234 atomic clusters due to the pyrolysis of acetylene. The evolution of one of
235 these atomic clusters is depicted in Fig. 2. Carbon and hydrogen atoms
236 are represented using black and red dots, respectively. First, the acetylene
237 molecules combine to form small linear chains (Fig. 2B: linearization) and
238 then transform into cyclic structures (Fig. 2C: cyclization). After cyclization,
239 the small clusters start growing due to both bond formation at the surface
240 and internal reorganization. These larger atomic clusters resemble incipient
241 soot particles (Fig. 2D–F). It is important to note that, the collisions are
242 stochastic in nature and the time required for an event, i.e. linearization,
243 cyclization, surface growth. etc., varies based on the initial configurations,
244 and therefore are omitted from the figure for generality. A similar forma-
245 tion mechanism is also reported by Zhang et al. [56] for their carbon-black
simulations and Sharma et al [45] in their acetylene pyrolysis simulations.

Figure 2: A general representation of steps during the formation and evolution of incipient
soot cluster during acetylene pyrolysis (from a simulation performed at 1650 K).
246

9
247 The incipient soot clusters are extracted from the RMD simulations at
248 different timesteps to capture the growth. Each simulation is run multiple
249 times (five or more) with a randomly initialized velocity field in each case to
250 generate soot particles with different evolutionary histories. In total, 3324
251 individual soot clusters are isolated from the RMD simulations. The particle
252 C/H ratios (Θ ) of these particles are compared with the theoretical limits
C/H

253 for PAHs in Fig. 3. Based on the compactness, the PAHs can be classified
254 into two categories: (a) peri-condensed PAHs, where the carbon atoms in
255 the aromatic structures can be shared by more than two aromatic rings and
256 (b) cata-condensed PAHs, where the carbon atoms in the aromatic structure
257 can be shared by at most two aromatic rings. Siegmann and Sattler [57]
258 proposed a relationship between the number of carbon and hydrogen for both
259 peri-condensed and cata-condensed PAHs. Fig. 3 presents the C/H ratio and
260 molar mass of the soot clusters from different temperatures and compares it
261 with the peri-condensed and cata-condensed PAH zones derived from [57]. As
262 observed from Fig. 3, the soot clusters fall between the peri-condensed and
263 cata-condensed boundaries, indicating an intricate network of different types
264 of aromatic and aliphatic structures in incipient particles. Commodo et. al.
265 [29] reported, using atomic force microscopy (AFM), that smaller aromatic
266 clusters (number of carbon atoms ranging from 6 to 55, lower than what
267 studied in this work) in the early stage of soot formation in a slightly sooting
268 premixed ethylene flame tend to be close the peri-condensed line. Beyond
269 that, to the best of the authors’ knowledge, there is no data available for the
270 PAH characteristics inside larger incipient soot particles as studied here.

271 3.2. Classification of incipient soot particles

272 We tagged, extracted, or calculated the following features for each parti-
273 cle:
274 1. Temperature (T )
275 2. Number of carbon atoms (NC )
276 3. Number of hydrogen atoms (NH )
277 4. Number of atoms (N )
278 5. Molar mass (M )
279 6. C/H ratio (Θ )C/H

280 7. Radius of gyration (Rg )

281 8. Atomic fractal dimension (Df )
282 9. Actual density (ρa )

10
 Figure 3: C/H ratio vs. molar mass of soot clusters at different temperatures.

Figure 4: A t-SNE diagram generated using all the 3324 incipient soot clusters with two
different k-means clusters.

11
283 10. Bulk density (ρb )
284 11. Total number of cyclic structures (N⃝ )
285 12. Fraction of cyclic carbon atoms (N©/NC )
286 13. Fraction of 5-member rings (N5/N⃝ )
287 14. Fraction of 6-member rings (N6/N⃝ )
288 15. Fraction of 7-member rings (N7/N⃝ )
289 16. Surface area (A)
290 17. Volume (V )
291 18. Area to volume ratio (A/V )
292 The entire feature sets for two sample particles are shown in Appendix B.
293 Our initial observation of trends of various internal and physicochemical
294 features revealed a wide variation, indicating that these particles can be
295 classified into multiple groups based on their features. We attempted two
296 unsupervised machine-learning techniques to extract unique classifications
297 that may exist in the incipient particle sample space. The first method is the
298 k-means clustering algorithm [53], which is used to label particles of different
299 classes based on all extracted features of the particles. Then we used the
300 t-dispersion stochastic neighbor embedding (t-SNE)[54] plot to display the
301 particle properties on a 2D map. In the t-SNE diagram, similar clusters (i.e.,
302 potentially belonging to the same class) are expected to be close to each
303 other. Although the number of classes was not known a priori, trial and
304 error with k-means clustering revealed good results with two classes. For
305 identification purposes, these two classes are referred to as “type 1” and
306 “type 2” particles, respectively. The resulting t-SNE diagram is shown in
307 Figure 4.
308 Looking closely at the two classes, we see that the particles exhibiting
309 similarity fall into a nearly continuous size range. For example, in the first
310 class, (type 1) the incipient particles have a lower total number of carbon
311 atoms (65 − 818) whereas in the second class (type 2) the particle have a
312 higher number of total atoms (759 − 1503). This essentially points to the
313 fact that the characteristics of the incipient particles change after a certain
314 level of growth: smaller particles (type 1) show different features and trends
315 than larger particles (type 2). It should be noted here that the number of
316 total carbon atoms is not a unique marker of the threshold between type 1
317 and type 2 – as indicated by a small overlap in the number of carbon atoms
318 range between the types – but acts as a very good surrogate for threshold
319 identifier. In total, we have obtained 670 type 1 and 2654 type 2 incipient

12
320 particles from a total of 3324 particles. Fig. 5 depicts some example particles
321 from the analyzed sample space. Here, the non-cyclic carbon atom structures
322 are shown in blue dots and the cyclic structures are shown in black, and
323 hydrogen atoms are omitted from the visualization for clarity.

Figure 5: Some example particles from each class obtained from RMD simulations. The
non-cyclic carbon atom structures are shown in blue dots and the cyclic structures are
shown in black. Hydrogen atoms are omitted from the visualization for clarity.

324 3.3. Comparison with experimental data

325 Before analyzing the internal structure, we compared some mean statistics
326 of soot particles obtained in our simulation with reported data available in
327 the literature. Johansson et al. [58] reported a bulk density of 1.51 g/cm3 for
328 incipient soot particles. In another study, Minutolo et. al. [59] calculated
329 the average density of soot to be 1.5 gm/cm3 using relationships proposed
330 in [60] for PAHs with C/H ratio ranging from 1.0-2.4. Similar values are also
331 reported and assumed in other studies [15, 61, 62]. These reported values of
332 density match well with the mean density of type 1 and type 2 particles in
333 our study, which was found to be 1.50 and 1.53 g/cm3 . Furthermore, Schulz
334 et. al [27] reported an average H/C ratio of 0.43 ± 0.03 (i.e., C/H ratio of
335 2.33 ± 0.16) for soot particles collected from a laminar premixed ethylene

13
336 flame which matches with the mean C/H ratio obtained for the type 2 (2.3)
337 particles in this study.

338 3.4. Internal structure of incipient soot particles

339 The radial distributions of various quantities relevant to the analysis of
340 the internal structure of the incipient particles are presented as box and
341 whiskers plots in subsequent figures. In this visualization, every box includes
342 the data points within the second and third quartiles, the horizontal line
343 inside the box indicates the median value and whiskers represent the range
344 of the data. The statistics were found to be insensitive to the process tem-
345 perature. Hence only aggregate data for all temperatures is presented here.

(a) (b)

(c) (d)

Figure 6: Radial distribution of non-cyclic and cyclic carbon atoms (N¢′′ (r∗ ) and N© ′′
(r∗ ))
∗ r
in the soot particle as a function of the normalized radial distance (r = Rg ) from the
center of mass. The blue vertical line is at a radial distance equivalent to Rg . Left column:
non-cyclic carbon, right column: cyclic carbon, top row: Type-1 particles, and bottom
row: Type 2 particles.

346 The radial distribution of cyclic and non-cyclic carbon atoms per unit area
347 for type 1 (Figs. 6(a)-6(b)) and type 2 (Figs. 6(c)-6(d)) incipient particles

14
348 as a function of the normalized radial distance from the center of mass is
349 shown in Fig. 6. The blue vertical line depicts the location where the radial
350 distance becomes equal to the radius of gyration (Rg ) of individual particles.
351 For type 1 particles, an abundance of non-cyclic carbon atoms is observed in
352 the central region. The number of non-cyclic carbon atoms is almost twice
353 the number of carbon atoms in cyclic structures in type 1 particles. Almost
354 all the carbon atoms reside near the central region of type 1 particles and
355 the number of carbon atoms quickly drops to zero as we go away from the
356 center of mass. This indicates that in type 1 particles, i.e., at the very early
357 stages of soot formation, the number of non-cyclic structures is significantly
358 higher than the number of cyclic structures.
359 The type 2 particles, on the other hand, show a different trend where a
360 very concentrated region of cyclic carbon atoms is observed in the central
361 region of the particles. The concentration of cyclic carbon atoms slowly
362 decreases as the distance from the center of mass increases. The number
363 of non-cyclic carbon atoms increases from a very low value in the central
364 region, then reaches a steady value just before the radius of gyration and
365 then gradually drops to zero as the distance increases beyond the radius of
366 gyration. This indicates that the non-cyclic carbon atoms are more likely to
367 be present in the outer region (what can be presumably considered near the
368 particle surface) of the type 2 incipient particles while the central region is
369 dominated by the cyclic carbon atoms.
370 Figure 6 suggests that the internal chemical structure of incipient soot
371 particles changes as the particles transition from type 1 to type 2. This indi-
372 cates the development of core and shell structures as the incipient soot grows
373 and matures. The presence of such core-shell structure has been theorized in
374 the literature. For example, Michelsen et. al. [63] used a fractal core-shell
375 model to explain the changes in the structure of soot aggregates and primary
376 particles at different heights of a laminar co-flow ethylene-air flame. Kholghy
377 et. al. [64] proposed a surface shell formation model to predict maturity of
378 soot primary particles. More directly, recently Botero et. al. [20] studied the
379 internal structure of soot particles using high-resolution transmission elec-
380 tron microscopy (HRTEM) to identify the PAH structures in the core and
381 shell regions, and suggested the presence of a stabilized core region indicat-
382 ing nano-structural mobility. Pascazio et. al. [40] utilized RMD simulation
383 and identified different levels of crosslinking in core and shell in hypothetical
384 soot particles. The clear difference in the radial distribution of cyclic and
385 non-cyclic carbon atoms between type 1 and type 2 particles supports these

15
386 findings.

(a)

(b)

θ (r ∗ )
Figure 7: Radial distribution of normalized C/H ratio (θ∗ (r∗ ) = C/H
C/H ΘC/H ) in incipient
∗ r
particles as a function of normalized radial distance (r = Rg ) from the center of mass.
C
The blue vertical line is at a radial distance equivalent to Rg . The H ratio of the strip
and the particle are the same along the red horizontal line.

387 The changes in chemical properties inside the incipient soot particles can
388 also be observed in the radial distribution of carbon to hydrogen ratio. The
389 normalized local C/H ratio for type 1 (Fig. 7(a)) and type 2 (Fig. 7(b))
390 particles are plotted as a function of the normalized radial distance from the
391 center of mass in Fig. 7 using box and whisker plots. Once again, the blue
392 vertical line depicts the location where the radial distance becomes equal to

16
393 the radius of gyration (Rg ) of individual particles. The red horizontal line
394 indicates where the local C/H ratio is equal to the particle C/H ratio.
395 The normalized local C/H ratio increases up to a certain distance from
396 the center of mass and then starts to drop. For type 1 particles, the increase
397 in the local C/H ratio is slow, and the median value gradually reaches a
398 peak value slightly higher than the particle C/H ratio (about 1.3 times).
399 After that, the local C/H ratio starts to drop and the median value reaches a
400 value close to the particle C/H ratio at around 65% of the radius of gyration.
401 The value stays close to the particle C/H ratio up to the radius of gyration,
402 and then slowly drops to zero. This indicates a very small or an absence
403 of a dense core region in type 1 particles. The demarcation between the
404 core and shell regions is not clear in type 1 particles because of the absence
405 of a pronounced core. For type 2 particles, however, the local C/H ratio
406 increases rapidly in the central region and the median value reaches a peak
407 value of about 1.7 times the particle C/H ratio. After that, the local C/H
408 ratio starts to drop and reaches a value equal to the global C/H ratio at
409 the radius of gyration. Unlike type 1 particles, the region where the median
410 of the normalized local C/H ratio is close to unity is very narrow in type 2
411 particles. This indicates that the dense core region of type 2 particles is well-
412 developed and the demarcation between the core and shell regions is more
413 pronounced than in type 1 particles.
414 The radial distribution of the normalized density inside type 1 (Fig. 8(a))
415 and type 2 (Fig. 8(b)) incipient soot particles is presented in Fig. 8 as a
416 function of the normalized radial distance from the center of mass of soot
417 particles (r∗ ) as box and whisker plots. For type 1 particles, Fig. 8 shows a
418 very small dense central region that extends up to about 40% of the radius of
419 gyration. The local density of type 1 particles is maximum near the center of
420 mass, and it drops gradually as the distance from the center of mass increases.
421 For type 2 particles, on the other hand, the dense core region is larger and
422 extends up to about 50-60% of the radius of gyration. The local density of
423 type 2 particles stays close to the particle density up to around 60% of the
424 radius of gyration, and then it quickly drops.

425 3.5. The boundary between the core and shell

426 The results discussed so far indicate the presence of a core and shell
427 structure in incipient soot particles. This is further examined by looking
428 at the median values of normalized local density (ρ∗a ), normalized C/H ratio
429 C/H
′′
(θ∗ ), and radial distribution of carbon atoms (N© and N¢′′ ) in Fig. 9 for both

17
 (a)

(b)

∗
Figure 8: Radial distribution of normalized local density (ρ∗ (r∗ ) = ϱ(r )
ρa ) inside incipient
particles as a function of normalized radial distance (r∗ = Rrg ) from the center of mass.
The blue vertical line is at a radial distance equivalent to Rg . The density of the strip and
the particle are the same along the red horizontal line.

18
 ϱa (r ∗ ) θC/H (r ∗ )
(a) Normalized density, ρ∗a = ρa (b) Normalized C/H ratio, θ∗ =C/H ΘC/H

(c) Radial distribution of cyclic carbon (d) Radial distribution of non-cyclic carbon
n (r ∗ ) n (r ∗ )
atoms per unit area, N©′′
= 4π(r©∗ ×Rg )2 atoms per unit area, N¢′′ = 4π(r¢∗ ×Rg )2

Figure 9: Identification of core and shell based on the radial distribution of physicochemical
properties (medians values) in incipient particles as a function of normalized radial distance
(r∗ = Rrg ) from the center of mass.

19
430 type 1 (in red) and type 2 (in blue) incipient soot particles. No significant
431 trend is noticed from the radial distribution of non-cyclic carbon per unit
432 area (Fig. 9(d)). However, the normalized density (Fig. 9(a)), C/H ratio
433 (Fig. 9(b)) and the radial distribution of cyclic carbon atoms per unit area
434 (Fig. 9(c)) demonstrate unique common trends in type 2 particles, which is
435 not observed in type 1 particles. All of the quantities
436 1. reach a maximum value near the center of mass and decrease gradually
437 to a local minimum at a distance of about 50% of the radius of gyration,
438 2. show the presence of a plateau region between 50% and 60% of the
439 radius of gyration,
440 3. and drop monotonously after approximately 60% of the radius of gy-
441 ration.
442 The first region or the central region can be identified as the core of the
443 soot as it is near the center of mass, denser and contains more rings than
444 other regions. The local minima mark the beginning of the boundary between
445 the core and the shell. The narrow plateau region can be thought of as an
446 overlap between the core and shell. And finally, the gradual descent of these
447 quantities indicates the shell region. In the type 1 particles, we can only see
448 the gradual descent stage, indicating that the core-shell demarcation is not
449 yet developed, i.e., there is no developed core. The reason for such a trend
450 in density, C/H ratio and ring structures can be explained by the nature of
451 the stacking of cyclic molecules (disordered and ordered), as shown in Fig.
452 10. As observed in the schematic in Fig. 10, the core region comprises an
453 interconnected cross-linked network of cyclic molecules while the shell region
454 contains sheet-like organization of cyclic molecules. Such structural differ-
455 ences in core and shell are also supported by results presented by Pascazio
456 et al. [40].

457 4. Conclusion
458 A series of reactive molecular dynamics (RMD) simulations were per-
459 formed to study the evolution of incipient soot particles during acetylene
460 pyrolysis at four different temperatures. A total of 3324 incipient soot par-
461 ticles were obtained at different stages of evolution from these simulations.
462 The mass, volume, surface area, radius of gyration, density, C/H ratio, and
463 the number of 5-/6-/7-member rings were calculated for each particle. The

20
 Figure 10: Schematic of core and shell structure of incipient particles.

464 internal structures of RMD-derived soot and their classification were inves-
465 tigated. Using unsupervised machine-learning techniques, the incipient soot
466 particles are classified into two types – type 1 and type 2 – based on their
467 morphological and chemical features. This classification was found to be
468 very well predicted by the size of the particles with smaller particles in the
469 type 1 class and larger particles in type 2. The internal structures of type 1
470 and type 2 particles show clearly distinct trends and features, indicating the
471 two types correspond to early and late stages of incipient soot. The internal
472 structure of the incipient particles from this investigation shows no direct or
473 obvious sensitivity to temperature.
474 Other conclusions drawn from the study are
475 1. Incipient soot particles comprise both cyclic and non-cyclic structures.
476 2. The core of type 1 particles is dominated by non-cyclic structures while
477 the core of type 2 particles is dominated by cyclic or ring structures.
478 3. The internal distribution of ring and non-ring structures indicates the
479 presence of a dense core region and a less dense shell region in type 2
480 incipient particles.
481 4. The core comprises an interconnected cross-linked network of cyclic
482 molecules while the shell region contains a sheet-like organization of
483 cyclic molecules.
484 5. The core of type 2 particle extends up to approximately 0.5Rg , followed

21
485 by an overlap region between core and shell, and shell in the periphery
486 beyond 0.6Rg .
487 6. The core and core-shell demarcation is not developed in type 1 particles.

488 5. Nomenclature
489 a: A parameter in Eqn A.4
490 A: Surface area of a particle (Å2 )
491 c: A parameter in Eqn A.4
492 Df : Atomic fractal dimension of a particle
493 mp,i : Mass of ith particle (kg)
494 M : Molar mass of a particle (kg/kmol)
495 Mp : Mass of a particle (kg)
496 n© : Number of cyclic carbon atoms at a specific location
497 n¢ : Number of non-cyclic carbon atoms at a specific location
498 N : Total number of atoms in the entire particle
499 N⃝ : Number of rings in the entire particle
500 N5 : Number of 5-membered rings in the entire particle
501 N6 : Number of 6-membered rings in the entire particle
502 N7 : Number of 7-membered rings in the entire particle
503 NC : Number of carbon atoms in the entire particle
504 NH : Number of hydrogen atoms in the entire particle
505 N© : Number of cyclic carbon atoms in the entire particle
506 N¢ Number of non-cyclic carbon atoms in the entire particle
507 r: Local radius (Å)
508 Req : Volume equivalent radius (Å)
509 Rg : Radius of gyration (Å)
510 ρa : Actual density of a particle (kg/m3 )
511 ρb : Bulk density of a particle (kg/m3 )
512 ϱ: Local (actual) density (kg/m3 )
513 T : Temperature (K)
514 θ : Local C/H ratio
C/H

515 Θ : C/H ratio of the entire particle

C/H

516 V : Volume of a particle (Å3 )

517 wC : Mass of a carbon atom (kg)
518 wH : Mass of a hydrogen atom (kg)
519

520 Superscripts:

22
521 ∗: Denotes normalized value
522 ′′: Denotes per unit area value

523 6. Acknowledgments
524 The research benefited from computational resources provided through
525 the NCMAS, supported by the Australian Government, The University of
526 Melbourne’s Research Computing Services and the Petascale Campus Ini-
527 tiative. K.M.M. and S.R. acknowledge funding support from the National
528 Science Foundation as some of this material is based upon work supported
529 by the National Science Foundation under Grant No. 2144290.

530 References
531 [1] N. R. Jacobsen, G. Pojana, P. White, P. Møller, C. A. Cohn, K. S.
532 Korsholm, U. Vogel, A. Marcomini, S. Loft, H. Wallin, Genotoxicity,
533 cytotoxicity, and reactive oxygen species induced by single-walled carbon
534 nanotubes and C60 fullerenes in the FE1-Muta™Mouse lung epithelial
535 cells, Environ. Mol. Mutagen. 49 (6) (2008) 476–487. doi:10.1002/em.
536 20406.

537 [2] M. Jerrett, R. T. Burnett, B. S. Beckerman, M. C. Turner, D. Krewski,

538 G. Thurston, R. V. Martin, A. van Donkelaar, E. Hughes, Y. Shi, S. M.
539 Gapstur, M. J. Thun, C. A. P. Iii, Spatial Analysis of Air Pollution and
540 Mortality in California, Am. J. Respir. Crit. Care Med. (Aug. 2013).
541 URL https://www.atsjournals.org/doi/10.1164/rccm.
542 201303-0609OC

543 [3] J. Hansen, L. Nazarenko, Soot climate forcing via snow and ice albedos,
544 Proc. Natl. Acad. Sci. U.S.A. 101 (2) (2004) 423–428. doi:10.1073/
545 pnas.2237157100.

546 [4] T. C. Bond, S. J. Doherty, D. W. Fahey, P. M. Forster, T. Berntsen,

547 B. J. DeAngelo, M. G. Flanner, S. Ghan, B. Kärcher, D. Koch, S. Kinne,
548 Y. Kondo, P. K. Quinn, M. C. Sarofim, M. G. Schultz, M. Schulz,
549 C. Venkataraman, H. Zhang, S. Zhang, N. Bellouin, S. K. Guttikunda,
550 P. K. Hopke, M. Z. Jacobson, J. W. Kaiser, Z. Klimont, U. Lohmann,
551 J. P. Schwarz, D. Shindell, T. Storelvmo, S. G. Warren, C. S. Zender,
552 Bounding the role of black carbon in the climate system: A scientific

23
553 assessment, Journal of Geophysical Research: Atmospheres 118 (11)
554 (2013) 5380–5552. doi:10.1002/jgrd.50171.
555 [5] C. Irimiea, A. Faccinetto, X. Mercier, I.-K. Ortega, N. Nuns,
556 E. Therssen, P. Desgroux, C. Focsa, Unveiling trends in soot nucleation
557 and growth: When secondary ion mass spectrometry meets statistical
558 analysis, Carbon 144 (2019) 815–830. doi:10.1016/j.carbon.2018.
559 12.015.
560 [6] J. Appel, H. Bockhorn, M. Frenklach, Kinetic modeling of soot for-
561 mation with detailed chemistry and physics: laminar premixed flames
562 of C2 hydrocarbons, Combust. Flame 121 (1) (2000) 122–136. doi:
563 10.1016/S0010-2180(99)00135-2.
564 [7] H. Wang, Formation of nascent soot and other condensed-phase ma-
565 terials in flames, Proc. Combust. Inst. 33 (1) (2011) 41–67. doi:
566 10.1016/j.proci.2010.09.009.
567 [8] S. Rigopoulos, Modelling of Soot Aerosol Dynamics in Turbulent Flow,
568 Flow, Turbulence and Combustion 103 (3) (2019) 565–604. doi:10.
569 1007/s10494-019-00054-8.
570 [9] D. D. Li, C. Wang, Q. N. Chan, G. H. Yeoh, Soot: A review of
571 computational models at different length scales, Experimental and
572 Computational Multiphase Flow 5 (1) (2023) 1–14. doi:10.1007/
573 s42757-021-0124-4.
574 URL https://link.springer.com/10.1007/s42757-021-0124-4
575 [10] R. A. Dobbins, H. Subramaniasivam, Soot Precursor Particles in Flames,
576 in: Soot Formation in Combustion, Springer, Berlin, Germany, 1994, pp.
577 290–301. doi:10.1007/978-3-642-85167-4_16.
578 [11] M. Balthasar, M. Kraft, A stochastic approach to calculate the par-
579 ticle size distribution function of soot particles in laminar premixed
580 flames, Combust. Flame 133 (3) (2003) 289–298. doi:10.1016/
581 S0010-2180(03)00003-8.
582 [12] B. Wang, S. Mosbach, S. Schmutzhard, S. Shuai, Y. Huang, M. Kraft,
583 Modelling soot formation from wall films in a gasoline direct injection
584 engine using a detailed population balance model, Appl. Energy 163
585 (2016) 154–166. doi:10.1016/j.apenergy.2015.11.011.

24
586 [13] H. A. Michelsen, M. B. Colket, P.-E. Bengtsson, A. D’Anna, P. Des-
587 groux, B. S. Haynes, J. H. Miller, G. J. Nathan, H. Pitsch, H. Wang, A
588 Review of Terminology Used to Describe Soot Formation and Evolution
589 under Combustion and Pyrolytic Conditions, ACS Nano 14 (10) (2020)
590 12470–12490. doi:10.1021/acsnano.0c06226.

591 [14] R. L. Vander Wal, Soot precursor carbonization: Visualization using

592 LIF and LII and comparison using bright and dark field TEM, Combust.
593 Flame 112 (4) (1998) 607–616. doi:10.1016/S0010-2180(97)00171-5.

594 [15] B. Zhao, K. Uchikawa, H. Wang, A comparative study of nanoparticles

595 in premixed flames by scanning mobility particle sizer, small angle neu-
596 tron scattering, and transmission electron microscopy, Proc. Combust.
597 Inst. 31 (1) (2007) 851–860. doi:10.1016/j.proci.2006.08.064.

598 [16] Q. Mao, A. C. T. van Duin, K. H. Luo, Formation of incipient soot

599 particles from polycyclic aromatic hydrocarbons: A ReaxFF molecular
600 dynamics study, Carbon 121 (2017) 380–388. doi:10.1016/j.carbon.
601 2017.06.009.

602 [17] M. Chen, W. Li, H. Zhang, M. Liu, J. Zhang, X. Li, Y. Han, Recent
603 ReaxFF MD studies on pyrolysis and combustion mechanisms of avi-
604 ation/aerospace fuels and energetic additives, Energy Advances 2 (1)
605 (2023) 54–72. doi:10.1039/D2YA00285J.

606 [18] H. Sabbah, L. Biennier, S. J. Klippenstein, I. R. Sims, B. R. Rowe, Ex-

607 ploring the Role of PAHs in the Formation of Soot: Pyrene Dimerization,
608 J. Phys. Chem. Lett. 1 (19) (2010) 2962–2967. doi:10.1021/jz101033t.

609 [19] N. E. Sánchez, A. Callejas, Á. Millera, R. Bilbao, M. U. Alzueta, Poly-

610 cyclic Aromatic Hydrocarbon (PAH) and Soot Formation in the Pyrol-
611 ysis of Acetylene and Ethylene: Effect of the Reaction Temperature,
612 Energy Fuels 26 (8) (2012) 4823–4829. doi:10.1021/ef300749q.

613 [20] M. L. Botero, Y. Sheng, J. Akroyd, J. Martin, J. A. H. Dreyer, W. Yang,

614 M. Kraft, Internal structure of soot particles in a diffusion flame, Carbon
615 141 (2019) 635–642. doi:10.1016/j.carbon.2018.09.063.

616 [21] Q. Chang, R. Gao, M. Gao, G. Yu, F. Wang, The structural evolution
617 and fragmentation of coal-derived soot and carbon black during high-

25
618 temperature air oxidation, Combust. Flame 216 (2020) 111–125. doi:
619 10.1016/j.combustflame.2019.11.045.

620 [22] P. P. Morajkar, M. K. Abdrabou, A. Raj, M. Elkadi, S. Stephen,

621 M. Ibrahim Ali, Transmission of trace metals from fuels to soot parti-
622 cles: An ICP-MS and soot nanostructural disorder study using diesel
623 and diesel/Karanja biodiesel blend, Fuel 280 (2020) 118631. doi:
624 10.1016/j.fuel.2020.118631.

625 [23] K. Gleason, F. Carbone, A. J. Sumner, B. D. Drollette, D. L. Plata,

626 A. Gomez, Small aromatic hydrocarbons control the onset of soot
627 nucleation, Combust. Flame 223 (2021) 398–406. doi:10.1016/j.
628 combustflame.2020.08.029.

629 [24] F. Carbone, K. Gleason, A. Gomez, Probing gas-to-particle transition

630 in a moderately sooting atmospheric pressure ethylene/air laminar pre-
631 mixed flame. Part I: gas phase and soot ensemble characterization, Com-
632 bust. Flame 181 (2017) 315–328. doi:10.1016/j.combustflame.2017.
633 01.029.

634 [25] R. S. Jacobson, A. R. Korte, A. Vertes, J. H. Miller, The Molecular

635 Composition of Soot, Angew. Chem. Int. Ed. 59 (11) (2020) 4484–4490.
636 doi:10.1002/anie.201914115.

637 [26] A. C. Barone, A. D’Alessio, A. D’Anna, Morphological characterization

638 of the early process of soot formation by atomic force microscopy, Com-
639 bust. Flame 132 (1) (2003) 181–187. doi:10.1016/S0010-2180(02)
640 00434-0.

641 [27] F. Schulz, M. Commodo, K. Kaiser, G. De Falco, P. Minutolo, G. Meyer,

642 A. D`anna, L. Gross, Insights into incipient soot formation by atomic
643 force microscopy, Proc. Combust. Inst. 37 (1) (2019) 885–892. doi:
644 10.1016/j.proci.2018.06.100.

645 [28] M. Commodo, A. D’Anna, G. De Falco, R. Larciprete, P. Minutolo,

646 Illuminating the earliest stages of the soot formation by photoemission
647 and Raman spectroscopy, Combust. Flame 181 (2017) 188–197. doi:
648 10.1016/j.combustflame.2017.03.020.

26
649 [29] M. Commodo, K. Kaiser, G. De Falco, P. Minutolo, F. Schulz,
650 A. D’Anna, L. Gross, On the early stages of soot formation: Molec-
651 ular structure elucidation by high-resolution atomic force microscopy,
652 Combust. Flame 205 (2019) 154–164. doi:10.1016/j.combustflame.
653 2019.03.042.
654 [30] D. S. N. Parker, F. Zhang, Y. S. Kim, Ralf. I. Kaiser, A. M. Mebel,
655 On the Formation of Resonantly Stabilized C5H3 Radicals—A Crossed
656 Beam and Ab Initio Study of the Reaction of Ground State Carbon
657 Atoms with Vinylacetylene, J. Phys. Chem. A 115 (5) (2011) 593–601.
658 doi:10.1021/jp109800h.
659 [31] K. O. Johansson, M. P. Head-Gordon, P. E. Schrader, K. R. Wilson,
660 H. A. Michelsen, Resonance-stabilized hydrocarbon-radical chain reac-
661 tions may explain soot inception and growth, Science 361 (6406) (2018)
662 997–1000. doi:10.1126/science.aat3417.
663 [32] F. S. Gentile, F. Picca, G. De Falco, M. Commodo, P. Minutolo,
664 M. Causà, A. D’Anna, Soot inception: A DFT study of σ and π
665 dimerization of resonantly stabilized aromatic radicals, Fuel 279 (2020)
666 118491. doi:10.1016/j.fuel.2020.118491.
667 [33] J. A. Rundel, C. M. Thomas, P. E. Schrader, K. R. Wilson, K. O.
668 Johansson, R. P. Bambha, H. A. Michelsen, Promotion of particle for-
669 mation by resonance-stabilized radicals during hydrocarbon pyrolysis,
670 Combust. Flame 243 (2022) 111942. doi:10.1016/j.combustflame.
671 2021.111942.
672 [34] A. C. T. van Duin, S. Dasgupta, F. Lorant, W. A. Goddard, ReaxFF:
673 A Reactive Force Field for Hydrocarbons, J. Phys. Chem. A 105 (41)
674 (2001) 9396–9409. doi:10.1021/jp004368u.
675 [35] C. Ashraf, A. C. T. van Duin, Extension of the ReaxFF Combustion
676 Force Field toward Syngas Combustion and Initial Oxidation Kinetics,
677 J. Phys. Chem. A 121 (5) (2017) 1051–1068. arXiv:28072539, doi:
678 10.1021/acs.jpca.6b12429.
679 [36] K. Chenoweth, A. C. T. van Duin, W. A. Goddard, ReaxFF Re-
680 active Force Field for Molecular Dynamics Simulations of Hydrocar-
681 bon Oxidation, J. Phys. Chem. A 112 (5) (2008) 1040–1053. doi:
682 10.1021/jp709896w.

27
683 [37] C. A. Schuetz, M. Frenklach, Nucleation of soot: Molecular dynamics
684 simulations of pyrene dimerization, Proc. Combust. Inst. 29 (2) (2002)
685 2307–2314. doi:10.1016/S1540-7489(02)80281-4.

686 [38] S. Han, X. Li, F. Nie, M. Zheng, X. Liu, L. Guo, Revealing the
687 Initial Chemistry of Soot Nanoparticle Formation by ReaxFF Molec-
688 ular Dynamics Simulations, Energy Fuels 31 (8) (2017) 8434–8444.
689 doi:10.1021/acs.energyfuels.7b01194.

690 [39] C. Chen, X. Jiang, Molecular dynamics simulation of soot formation

691 during diesel combustion with oxygenated fuel addition, Phys. Chem.
692 Chem. Phys. 22 (36) (2020) 20829–20836. doi:10.1039/D0CP01917H.

693 [40] L. Pascazio, J. W. Martin, K. Bowal, J. Akroyd, M. Kraft, Exploring

694 the internal structure of soot particles using nanoindentation: A reactive
695 molecular dynamics study, Combust. Flame 219 (2020) 45–56. doi:
696 10.1016/j.combustflame.2020.04.029.

697 [41] T. Ishiguro, Y. Takatori, K. Akihama, Microstructure of diesel soot

698 particles probed by electron microscopy: First observation of inner core
699 and outer shell, Combust. Flame 108 (1) (1997) 231–234. doi:10.1016/
700 S0010-2180(96)00206-4.

701 [42] A. D. Pathak, D. Potphode, C. S. Sharma, Graphitization induced struc-

702 tural transformation of candle soot carbon into carbon nano-onion as a
703 functional anode for metal-ion batteries, Mater. Adv. 3 (8) (2022) 3610–
704 3619. doi:10.1039/D2MA00042C.

705 [43] Y. Liu, G. Lv, C. Fan, N. Li, X. Wang, Surface functional groups and
706 graphitization degree of soot in the sooting history of methane premixed
707 flame, in: WCX™ 17: SAE World Congress Experience, SAE Interna-
708 tional, 2017. doi:https://doi.org/10.4271/2017-01-1003.
709 URL https://doi.org/10.4271/2017-01-1003

710 [44] M. Ruiz, R. Villoria, A. Millera, M. Alzueta, R. Bilbao, Influence of the

711 temperature on the properties of the soot formed from c2h2 pyrolysis,
712 Chemical Engineering Journal 127 (2007) 1–9. doi:10.1016/J.CEJ.
713 2006.09.006.

28
714 [45] A. Sharma, K. M. Mukut, S. P. Roy, E. Goudeli, The coalescence of
715 incipient soot clusters, Carbon 180 (2021) 215–225. doi:10.1016/j.
716 carbon.2021.04.065.

717 [46] A. P. Thompson, H. M. Aktulga, R. Berger, D. S. Bolintineanu, W. M.

718 Brown, P. S. Crozier, P. J. In ’t Veld, A. Kohlmeyer, S. G. Moore, T. D.
719 Nguyen, R. Shan, M. J. Stevens, J. Tranchida, C. Trott, S. J. Plimp-
720 ton, LAMMPS - a flexible simulation tool for particle-based materials
721 modeling at the atomic, meso, and continuum scales, Comput. Phys.
722 Commun. 271 (2022) 108171. doi:10.1016/j.cpc.2021.108171.

723 [47] F. Castro-Marcano, A. M. Kamat, M. F. Russo, A. C. T. van Duin, J. P.

724 Mathews, Combustion of an Illinois No. 6 coal char simulated using an
725 atomistic char representation and the ReaxFF reactive force field, Com-
726 bust. Flame 159 (3) (2012) 1272–1285. doi:10.1016/j.combustflame.
727 2011.10.022.

728 [48] W. C. Swope, H. C. Andersen, P. H. Berens, K. R. Wilson, A computer

729 simulation method for the calculation of equilibrium constants for the
730 formation of physical clusters of molecules: Application to small water
731 clusters, J. Chem. Phys. 76 (1) (1982) 637–649. doi:10.1063/1.442716.

732 [49] D. J. Evans, B. L. Holian, The Nose–Hoover thermostat, J. Chem. Phys.

733 83 (8) (1985) 4069–4074. doi:10.1063/1.449071.

734 [50] M. F. Sanner, A. J. Olson, J.-C. Spehner, Reduced surface: An efficient

735 way to compute molecular surfaces, Biopolymers 38 (3) (1996) 305–
736 320. doi:10.1002/(SICI)1097-0282(199603)38:3<305::AID-BIP4>
737 3.0.CO;2-Y.

738 [51] K. M. Mukut, S. Roy, E. Goudeli, Molecular arrangement and fringe

739 identification and analysis from molecular dynamics (MAFIA-MD): A
740 tool for analyzing the molecular structures formed during reactive molec-
741 ular dynamics simulation of hydrocarbons, Comput. Phys. Commun. 276
742 (2022) 108325. doi:10.1016/j.cpc.2022.108325.

743 [52] A. Stukowski, Visualization and analysis of atomistic simulation data

744 with OVITO–the Open Visualization Tool, Model. Simul. Mater. Sci.
745 Eng. 18 (1) (2009) 015012. doi:10.1088/0965-0393/18/1/015012.

29
746 [53] S. Lloyd, Least squares quantization in PCM, IEEE Trans. Inf. Theory
747 28 (2) (1982) 129–137. doi:10.1109/TIT.1982.1056489.

748 [54] L. van der Maaten, G. Hinton, Visualizing Data using t-SNE, Journal
749 of Machine Learning Research 9 (86) (2008) 2579–2605.
750 URL https://www.jmlr.org/papers/v9/vandermaaten08a.html

751 [55] K. M. Mukut, A. Sharma, E. Goudeli, S. P. Roy, A closer look into the
752 formation of soot particles: A molecular dynamics study, in: 12th US
753 National Combustion Meeting, College Station, TX, USA, 2021.

754 [56] C. Zhang, C. Zhang, Y. Ma, X. Xue, Imaging the C black formation
755 by acetylene pyrolysis with molecular reactive force field simulations,
756 Phys. Chem. Chem. Phys. 17 (17) (2015) 11469–11480. doi:10.1039/
757 C5CP00926J.

758 [57] K. Siegmann, K. Sattler, Formation mechanism for polycyclic aromatic

759 hydrocarbons in methane flames, J. Chem. Phys. 112 (2) (2000) 698–
760 709. doi:10.1063/1.480648.

761 [58] K. O. Johansson, F. El Gabaly, P. E. Schrader, M. F. Campbell, H. A.

762 Michelsen, Evolution of maturity levels of the particle surface and bulk
763 during soot growth and oxidation in a flame, Aerosol Sci. Technol.
764 51 (12) (2017) 1333–1344. doi:10.1080/02786826.2017.1355047.

765 [59] P. Minutolo, M. Commodo, A. D’Anna, Optical properties of incipient

766 soot, Proc. Combust. Inst. (Oct. 2022). doi:10.1016/j.proci.2022.
767 09.019.

768 [60] H. A. Michelsen, Effects of maturity and temperature on soot density

769 and specific heat, Proc. Combust. Inst. 38 (1) (2021) 1197–1205. doi:
770 10.1016/j.proci.2020.06.383.

771 [61] A. Veshkini, Understanding Soot Particle Growth Chemistry and Par-
772 ticle Sizing Using a Novel Soot Growth and Formation Model, Ph.D.
773 thesis (Nov. 2015).
774 URL https://tspace.library.utoronto.ca/handle/1807/71398

775 [62] P. Desgroux, Recent advances in soot nucleation understanding, in: Eu-
776 ropean Combustion Meeting, Lisboa, Portugal, 2019.
777 URL https://hal.archives-ouvertes.fr/hal-03326326

30
778 [63] H. A. Michelsen, M. F. Campbell, K. O. Johansson, I. C. Tran,
779 P. E. Schrader, R. P. Bambha, E. Cenker, J. A. Hammons, C. Zhu,
780 E. Schaible, A. van Buuren, Soot-particle core-shell and fractal struc-
781 tures from small-angle X-ray scattering measurements in a flame, Car-
782 bon 196 (2022) 440–456. doi:10.1016/j.carbon.2022.05.009.
783 [64] M. R. Kholghy, A. Veshkini, M. J. Thomson, The core–shell internal
784 nanostructure of soot – A criterion to model soot maturity, Carbon 100
785 (2016) 508–536. doi:10.1016/j.carbon.2016.01.022.
786 [65] P. F. DeCarlo, J. G. Slowik, D. R. Worsnop, P. Davidovits, J. L.
787 Jimenez, Particle Morphology and Density Characterization by Com-
788 bined Mobility and Aerodynamic Diameter Measurements. Part 1: The-
789 ory, Aerosol Sci. Technol. 38 (12) (2004) 1185–1205. doi:10.1080/
790 027868290903907.
791 [66] J. Theiler, Estimating fractal dimension, J. Opt. Soc. Am. A, JOSAA
792 7 (6) (1990) 1055–1073. doi:10.1364/JOSAA.7.001055.
793 [67] S. R. Forrest, T. A. Witten, Jr., Long-range correlations in smoke-
794 particle aggregates, J. Phys. A: Math. Gen. 12 (5) (1979) L109. doi:
795 10.1088/0305-4470/12/5/008.

796 Appendix A. Expressions for physical properties of soot particles

797 The trajectory files obtained from RMD simulations contain coordinates
798 of each atom with reference to a global reference frame. This coordinate
799 information along with the mass of each atom is used to calculate the coor-
800 dinate of the center of mass of each particle. The mass of a particle Mp is
801 calculated by summing up the mass of the atoms in the cluster. The volume
802 (V ) is calculated using MSMS [50] with a pore size of 1.5 Å. The volume
803 equivalent radius of a particle with volume V is calculated via Eqn. A.1.
 1/3
3V
Req = (A.1)
4π
804 The radius of gyration (Rg ) is calculated following the standard definition
805 using Eqn. A.2. sP
N
mp,i ri2
Rg = Pi=1
N
, (A.2)
i=1 mp,i

31
806 where ri is the distance of the ith atom from the center of mass, mp,i is the
807 mass of individual atoms, and N is the total number of atoms in the cluster.
808 The actual density (ρa ) is calculated using the particle mass (Mp ) and
809 volume (V ) of the incipient particle using Eqn. A.3.

Mp
ρa = (A.3)
V
810 Bulk density [58, 65] of an incipient particle is calculated using Eqn. A.4.
 
2 −1 wC Θ +wH
ρb = (0.260884a c) Θ +1
C/H

C/H
, (A.4)

811 where wC and wH are the molar masses of a carbon and hydrogen atoms, a
812 is the length of the graphite unite cell in the basal plane, c is the interlayer
813 spacing in Angstroms, and Θ represents the carbon to hydrogen ratio of
C/H

814 the cluster. More details can be found in [58, 65].

815 The atomic fractal dimension (Df ), following the approach used in [45],
816 is calculated using the sandbox method [66, 67] using Eqn. A.5.

log Mp (r)
Df = , (A.5)
log r

817 where Mp (r) is the mass of atoms in the cluster as a function of radial distance
818 from the center of mass. Please note that this “atomic” fractal dimension
819 is for a single incipient particle and is different from the traditional fractal
820 dimension used in aggregate characterization [45].

821 Appendix B. A sample of physicochemical data used and analyzed

822 in this study
823 Figure B.11 shows two sample soot clusters and their properties as exam-
824 ples. These clusters, labeled as A and B, were extracted from a simulation at
825 1500 K at two different times. The left side shows the molecular structure
826 of the particle and the three-dimensional volumetric representation by con-
827 structing a surface mesh using OVITO [52] (this is what the incipient particles
828 would actually look like). The physicochemical properties of these particles
829 as analyzed in this work for classification via machine-learning are tabulated
830 on the right side of the figure.

32
Figure B.11: Two sample soot particles and their attributes investigated in this study

33