Evolution of Glassy Carbon Microstructure: In Situ Transmission Electron Microscopy of the Pyrolysis Process

Glassy carbon is a graphene-rich form of elemental carbon obtained from pyrolysis of polymers, which is composed of three-dimensionally arranged, curved graphene fragments alongside fractions of disordered carbon and voids. Pyrolysis encompasses gradual heating of polymers at ≥ 900 °C under inert atmosphere, followed by cooling to room temperature. Here we report on an experimental method to perform in situ high-resolution transmission electron microscopy (HR-TEM) for the direct visualization of microstructural evolution in a pyrolyzing polymer in the 500–1200 °C temperature range. The results are compared with the existing microstructural models of glassy carbon. Reported experiments are performed at 80 kV acceleration voltage using MEMS-based heating chips as sample substrates to minimize any undesired beam-damage or sample preparation induced transformations. The outcome suggests that the geometry, expansion and atomic arrangement within the resulting graphene fragments constantly change, and that the intermediate structures provide important cues on the evolution of glassy carbon. A complete understanding of the pyrolysis process will allow for a general process tuning specific to the precursor polymer for obtaining glassy carbon with pre-defined properties.


Results and Discussions
Sample morphology and imaging locations. Different sample types that were probed by in situ TEM imaging are presented in Fig. 1. Figure 1A is an optical micrograph of a SU-8 fiber patterned across three imaging windows of a heating chip. TEM micrographs of a freestanding cantilever-like structure, the edge of a fiber and a thin-film region are shown in Fig. 1B-D, respectively. Microstructural evolution of these structures during pyrolysis was analyzed to provide specific information on fullerene-like and other closed-cage geometries, migration and reorganization of the graphene fragments, graphitic stacking, fragment bifurcations and edge dynamics during pyrolysis. Structures with and without substrate were also compared for their post-pyrolysis shrinkage, which may be useful for microsystem engineers in determining the initial geometry of a pattern for nano-device fabrication. The substrate-supported fiber exhibited a unidirectional shrinkage (along z-axis), while the substrate-less structures (such as the one in Fig. 1B) shrank isometrically.
Post-pyrolysis glassy carbon. TEM images shown in Fig. 2 were recorded at the edge of a SU-8 nanofiber pyrolyzed at 1200 °C inside the TEM chamber. Three regions that resemble strongly folded or closed-cage structures (numbered 1-3) are shown in higher magnification to the left. Region 1 represents a curved graphene sheet that seems to lead towards a closed-cage formation. However, this feature is most likely a projection of a graphene fragment spread in 3D, which may actually have a significantly different curvature than what appears in the 2D projection. The possibilities of its closing or unfolding remain open at higher annealing temperatures. The angular  formation shown in Region 2 that contains near-120° bends, has a higher probability of representing a floating sheet than a completely closed, 3D structure. This feature is most likely a flat, disc-like graphene fragment. Region 3 gives the impression of a circular structure, but on a closer observation one can detect minor contrast variations on its periphery. This structure is probably a tilted graphene fragment, where the tilting is causing different defocus conditions at the edges. Similar structures are often mistaken for fullerenes or nanoparticles in the pre-manufactured glassy carbons, owing to the lack of their structural evolution data. We elucidate this with simplified drawings in Fig. 3, describing the various geometric possibilities for a projected pattern. Some of these schematics also facilitate discussions for the in situ TEM data detailed in the subsequent text. As can be observed in Fig. 3A,B, the projected views representing a near-circular structure, and a set of concentric circles, can be rendered in various 3D shapes. Similarly, the illustration in Fig. 3C could depict the bifurcation of a graphene fragment at a trigonal junction, or simply present a set of hexagons in different (xy) planes. Figure 3D is the demonstration of a collective projection effect of various graphene fragments existing at different depths, which give the impression of intersections or crossovers, since only their edges are visible in a TEM image.
In situ observation of fullerene-like structures. Figure 4 is a compilation of detailed in situ TEM micrographs of a freestanding film during pyrolysis. The focus here is to evaluate the formation of fullerene-like structures with the help of the intermediates formed during pyrolysis. Prior to this, it is important to mention that the term 'fullerene-like' in this report is used specifically for carbon nanostructures that are compared with completely closed buckminsterfullerenes. In the literature, this term is occasionally used in a more general sense, i.e., to describe a variety of graphene-based structures that contain pentagons/ heptagons or feature a high curvature. This has been cautiously avoided.
As can be observed in Fig. 4, near-circular projections of carbon nanostructures are already traceable during the preliminary pyrolysis stages (see image at 520 °C). However, at these stages (typically up to 900 °C), the material undergoes a rapid reconfiguration. After pyrolysis (see RT images), 3 well-defined circular carbon nanostructures (labeled 1-3) are identified, which are further examined for being fullerenes using in situ structural evolution data. Structures 1 and 2, which first appear around 800 °C (indicated by an arrow), display negligible change in shape or size up to 1200 °C. They slightly move apart, but do not make any attempt to attach to a neighboring fragment, despite a rapid rearrangement in the surrounding material. This structural stability supports the idea of a closed-cage formation, which is otherwise challenging to distinguish from a flat fragment detected in a single HRTEM image. The approximate equivalent diameter, measured for two such structures varied between 0.7 to 0.9 nm, which is similar to the size of a Buckminsterfullerene. Even with this information, only structure-1 can be claimed as a fullerene. Minor shape and edge-contrast variations between 1080-1200 °C add uncertainty to defining structure-2. One cannot overrule the possibility that it is simply a protruding edge of another fragment. Structure 3 is slightly larger, which can either be a closed-cage (spherical) structure or a flat graphene flake resembling a disc. Without the knowledge of its evolution history, there is a high probability that it is misconstrued for a fullerene or a nanoparticle. However, on a careful backward trace one can witness the frequent distortion in its shape, as well as its migration, which corresponds to that of a floating disc-like fragment, not a spherical particle. A similar structural development was observed at the tip of the cantilever-like structure shown in Fig. 1B. Sequential in situ TEM images (100 °C intervals) recorded at this location are presented in Fig. 5(A-H). Similar to the experiment shown in Fig. 4, a significant structural change could be seen between 800 and 900 °C (Fig. 5C,D), where presumably a small closed caged structure is formed at the tip, which is stable during further heating, despite the constantly reorganizing surrounding material ( Fig. 5E-H). From the micrographs in Figs 4 and 5, one obtains visual evidence of a prevailing microstructural change around 900 °C. This fits to the well-established sudden enhancement of glassy carbon's electrical conductivity and Young's modulus around 900 °C 4 . These micrographs efficiently demonstrate how in situ imaging capabilities provide previously unknown essential information on the evolution of glassy carbon.
A comparison of in situ TEM data with the known physicochemical properties of low-temperature pyrolytic carbons (e.g. EPR characteristics 37 ) suggests that the mobile graphene fragments with short-range order carry a large fraction of edge-radicals, and are constantly attempting to attain thermodynamically stable arrangements. A majority of fragments contain non-six membered rings, which cause them to curl and fold. As a result, the pyrolyzing material during the initial pyrolysis stages contains constantly migrating, curved and bent, floating graphene fragments with highly reactive edges. Occasionally, these mobile fragments form completely closed structures such as fullerenes. Some such defect-containing fragments also stack up, which is essentially turbostratic, since the defects cause misalignments in the basal planes and thus restrict a graphitic stacking (ABABA type).
Many structural defects are carried forward from the backbone structure 40 that are in turn influenced by polymer's chemical composition (distribution and nature of the C-heteroatom bonds) and the arrangement/ entanglement of its chains. We speculate that some defects are also generated during pyrolysis. Generally, symmetric, six-membered ring based configurations are energetically favored in a graphene sheet. But in the case of a dynamic 3D material such as glassy carbon, it is possible that the surrounding matrix (composed of various reactive edges and already curved sheets) facilitates the formation of non-six-membered rings at a lower overall energy. Plausible reasons include steric hindrance, topological frustrations 61,62 or just a lack of neighboring atoms, e.g., near the sample boundaries and surfaces.
Separation and merger of graphene fragments. Figure 6 is a subcollection of the TEM micrographs shown in Fig. 4 with an emphasis on the migration of small, floating graphene flakes during early pyrolysis stages. In Fig. 6A-C, the separation of a small fragment (diameter: < 2 nm) from a larger mass of graphene can be witnessed. Initially, a neck-like structure is formed at an unstable region around 640 °C, which subsequently develops into an independent flake at higher pyrolysis temperatures. The micrographs in Fig. 6D-F represent the merger of another flake, which entered the field of view of these micrographs at around 700 °C, into a larger fragment. Notably, the two transitions take place in the same temperature range and are perhaps interdependent. From these micrographs it is clear that the mobile graphene fragments are capable of completely merging or separating out, which would cause a sudden change in their size along the basal plane (L a ). Such migration patterns are more prominent during the early pyrolysis stages, likely due to the instabilities caused by their higher mobility.
Multi-layer graphene and trigonal junctions. Figure 7A,B are the micrographs recorded for a relatively thick fiber sample (multiple graphene fragments in all directions) after pyrolysis at 1200 °C. Similar micrographs have been reported by various researchers, since common glassy carbon samples are much thicker than the few-layer graphene. Here various graphene fragments present at different depths are projected on the image plane such that they seem to intersect or cross. One such location is magnified in Fig. 7C. Such apparent  intersections, previously denoted as a glassy carbon motif 26 , are merely a visual effect. Various such geometries can be traced in the parent images (Fig. 7A,B). Since only the edges of a fragment are visible in TEM, the planes also appear to be more densely packed compared to their actual configuration (see Fig. 3D). Figure 7D is yet another projection-derived geometry. Similar features have been confused for voids or nanoparticles in the literature due to their seemingly closed shape and ~5 nm size. In order to confirm whether the dense edges represent L c or are merely visual effects caused by multiple planes in projection, an analysis of the regional Fourier transforms was performed that allows for the determination of local inter-layer distance. Region 1 in Fig. 7A most likely corresponds to an edge of few-layer graphene. The maximum in the FFT corresponds to a distance of 0.42 nm. In the corresponding real space line profile of the edge (shown to the right), distances between 0.4 and 0.5 nm are measured. While the measured outermost distance is affected by the Fresnel fringes around the graphene layer, the other lattice distances are significantly higher than the characteristic inter-layer separation in graphite (0.335 nm) or in turbostratic graphene (0.336-0.344 nm 63 ). In contrast, the lattice distance measured at another location (marked Region 2) exhibits inter-layer spacing of ~0.34 nm, indicative of a graphitic stack imaged edge-on. Moiré effects causing these layer spacings as a projection effect can be excluded from the Fourier transforms.
We also observed that the graphene fragments often stack up on top of each other with a certain offset, rather than being perfectly edge-matched. Due to the edge-only projections, the top view created by such an offset can be confused for L c . Intensity profiles can provide some clarity, but only in the case of few-layer graphene. The identification of the orientation of the stacks becomes increasingly difficult with an increase in the sample thickness. Nonetheless, from these intensity profiles we can deduce that there is definitely some graphitic stacking and inter-fragment bonds in glassy carbon, which is often manifested by an increase in L c in the XRD patterns. Discrete fragments with no or very little stacking, as proposed by Harris, would also not justify glassy carbon's good electrical conductivity.
In Fig. 7E a trigonal junction can be observed, which could be an example of the bifurcation of graphene sheets with 120° angles, featuring a confluence-like geometry. It is also possible that it is an illusion created by various hexagonal graphene fragments lying in different planes. Both possibilities are schematically represented in Fig. 3C. It has been proposed that such junctions may contain tetrahedral geometries or diamond-like carbon (DLC) 26,33 , Since the nano-scale tetrahedral carbons are only stable up to 700 °C, such claims have also been repeatedly rejected 16,26 . Extreme sample preparation conditions, or beam-damage can potentially introduce some sp 3 -hybridized atoms, especially on the surface of glassy carbon. If the analysis is based solely on TEM, it is also possible that the trigonal geometries (such as in Fig. 7E) rotated from the image plane are projected at an angle other than 120° and are therefore confused for DLC. In any case, the presence of DLC cannot be evaluated by the reported TEM experiment due to the complex projection effects in the micrographs.
The micrographs in Fig. 7 are visually different from those recorded for the thin-films (Figs 4-6). This is a clear evidence of the influence of sample size and geometry on the resulting microstructure. A relatively thick sample contains a large fraction of angular bends (several 120° angles identified in Fig. 7A,B), which is not the case for the thin films. Multilayered structures allow for 3D interactions between various developing fragments. As a result, the fragments can spread and bond across different planes, and the stacked crystallites can orient themselves in any direction. On the other hand, graphene fragments in thin-films tend to expand primarily in the xy plane, since any out-of-plane protrusions would require a very high surface energy.
It is evident from the data shown in Figs 4-7 that the material (above 900 °C) is composed of randomly shaped graphene fragments of highly variable sizes, interconnected in a complex manner, rather than long and narrow, frequently bifurcating ribbons. These fragments do contain non-six-membered carbon species that introduce curvature in the fragments resulting in occasional formation of nanoparticles or fullerenes that are typically 0.7-2.0 nm in diameter. Slightly larger nanostructures are most likely floating flakes or projection effects. The term ribbon gives the impression that graphene sheets in glassy carbon feature a much greater length compared to their stack thickness, and also that the width of these so-called ribbons is more or less uniform along their entire length. Such a proposition is based on the hypothesis that the polymeric backbone serves as the nucleation point or the central axis for the growth of a graphene fragment 26 . Since the backbone is derived from long polymer chains, a graphene sheet originating from it would look like a ribbon, fibril or chain. The expansion of such ribbons was suggested to be restricted by the edge atoms featuring valance angle variations (non-sp 2 hybridizations), or by steric hindrance due to the neighboring ribbons 22 . Based on the in situ TEM data, we conclude that such ribbon-like graphene growth is a highly unlikely scenario, since the developing fragments contain reactive edges in all directions with no restriction to spread. Additionally, smaller flakes are quite mobile and can attach to other fragments. The separation and merger of graphene fragments indicates that instead of always expanding, they are constantly rearranging to achieve a more stable overall structure. The assumption that several stacked ribbons exhibit almost identical geometries is also improbable. In the reported work it was difficult to find even two identical fragments. The ribbon model is often supported by the XRD data that indicates an increase in L c and L a , before and after the pyrolysis temperature of ~1000 °C. Here one needs to be careful, since L a is defined as the expansion of the graphene basal plane that can occur in any, or simultaneously in all directions along the edges (e.g. radial growth). Often the ribbon or fibril models assume the increase in L a to be unidirectional.
In the model proposed by Harris, the change in L c has been completely overlooked and the material is believed to contain a very high fraction of strongly folded or closed-cage structures. Most of these structures, including the relatively large ones (5-10 nm), are considered fullerene-like. Such structures may indeed be complex 2D projections from multiple planes, 2D (disc-like or hexagonal) flakes, or the top-view of graphene layers arranged with an offset. A clear explanation of what is considered fullerene, and what is 'fullerene-like' , is also necessary for a better understanding of this model. Graphene edge dynamics during glassy carbon evolution. Finally, in Fig. 8 the patterns that evolve at the graphene edges during pyrolysis are revealed. One can witness a continuous transformation of these edges, present at an unhinged boundary of a thin-film sample, from a highly disordered structure to well-defined facets similar to those observed in synthetic graphene nanoribbons 64 . The 3-layer (or 3-stack) arrangement at this dynamic protruding edge is clearly detectable around 1000 °C. Evidently, the edges are highly dynamic as they undergo continuous reorganization. Their curvature changes constantly and the structure ultimately acquires a symmetry pattern (which could be zig-zag or armchair at a molecular level) with bends close to 120°. Interestingly, the edges become angular, slightly rounded, and then angular again between 1000 °C to RT, confirming that during pyrolysis ramp-up, the material continues to reorder even at higher temperatures. A cool-down at any given temperature leads to an energetically favorable arrangement. Also noticeable is the frequent change in the interlayer separation, which clearly indicates that these micrographs feature the top view of graphene layers rather than the side-view or L c (see schematic in Fig. 8J). Glassy carbon contains a considerable fraction of reactive, defect-containing edges that are responsible for the high degree of fragment mobility. In Fig. 8K-M, the slightly out of focus edges can be observed, which appear coarse and discontinuous. These defective edges may contain open rings, dangling bonds, voids, or other intermediate structures, in addition to non-six membered rings. Here, the reactive edge atoms have the option of forming σ-bonds with a neighboring flake, or of attaching with the upper (or lower) layer by π-bonds. An in-plane bonding of the two defective edges could also lead to grain boundaries 65 that can be partially annealed out at higher temperatures 66 .

Conclusions
By comparing the data from all reported experiments we conclude that fullerenes, strongly curved graphene sheets, and small 2D graphene flakes, co-exist with relatively much larger, stacked (< 10 layers) and interconnected graphene fragments of highly variable sizes and shapes in glassy carbon. The fraction and distribution of fullerenes and fullerene-like structures, as well as the extent and nature of 3D inter-fragment bonding are strongly dependent on the surface area of the sample, in addition to the chemical structure of the precursor polymer. Microporosity in glassy carbon should not be completely attributed to fullerene-like structures. Arbitrarily shaped voids caused by inter-fragment bonding across multiple planes also contribute to the material's low density. These random voids are more prevalent in 3D samples as opposed to the few-layer graphene version of glassy carbon. The fragment interconnects contain both σ and π bonds. Some of them may be highly strained resulting in a range of C-C bond-lengths in glassy carbon. Inherent non-six-membered rings, which are also responsible for a predominantly turbstratic arrangement, further contribute to bond-length variations. Graphene fragments in pyrolytic carbon are not always expanding. Occasionally smaller flakes separate from, or merge into the larger ones, due to local instabilities, especially at low pyrolysis temperatures.
The reported in situ TEM imaging protocol can be extended to practically any polymer in order to determine the exact microstructure of a specially designed glassy carbon. A complete understanding of the pyrolysis also allows for process tuning, for example, the change in dwell time or temperature ramp-up specific to the precursor polymer for obtaining the material with pre-defined properties. Various nanostructures in glassy carbon, for example, disc-like flakes as small as 2 nm, can only be deconvoluted with their complete evolution history, which is only possible with low-voltage TEM at the nano-scale. Other commonly used characterization techniques only yield the average values of material's properties and thus limit the microstructural information. The application of low-voltage scanning-TEM can be potentially employed for obtaining more detailed information regarding glassy carbon microstructure in the future.
Glassy carbon is not only an excellent engineering material, it is also a platform suitable for understanding the origin of graphene from a hydrocarbon, and the behavior of graphene fragments in a dynamic 3D environment. There are numerous possibilities for theoretical studies that would provide further information on the minimum energy configurations, despite strained bonds. Mathematical simulations can also shed light on the exact mechanism of fullerene formation in glassy carbon. Our current experiments were limited to 1200 °C due to the chip design, which can be further improved for obtaining microstructural data at higher temperatures.

Methods
TEM experiments were carried out on an aberration corrected FEI Titan 80-300 microscope operated at 80 kV, equipped with an Aduro 200 (Protochips Inc.) sample holder. SU-8 (a photopatternable thermosetting phenol-formaldehyde resin commonly used for carbon-MEMS fabrication 67 ; procured from MicroChem) fibers were manually placed onto Aduro E chips (Protochips Inc.) such that at least one fiber could be observed through the imaging window. Fibers were UV cross-linked and where necessary, further thinned by oxygen plasma etching prior to pyrolysis. Longer etch times resulted in fiber breakage at the center, thus yielding a cantilever-like structure (freestanding beam anchored at one end). Thin films were obtained when the patterned fiber barely touched the imaging window, allowing for extremely small quantities of SU-8 to be imaged. Chips were then placed in the Aduro 200 TEM holder and were heated inside the TEM chamber (pressure: ~10 −7 torr) at 5 °C/ minute ramp rate up to 1200 °C, followed by cooling at a rate of 10 °C/minute. Images were recorded at 20 °C intervals in the 500-1200 °C range. The beam was only turned on during imaging. Schematics of the top and cross-sectional views of a heating chip and a digital photograph of the accompanying TEM holder are provided in Fig. 9. The chip consists of a freestanding ceramic membrane coated with a thin amorphous silicon nitride (SiN) film with patterned holes as imaging windows. The ceramic membrane serves as the heating element and the SiN membrane (thickness: 50 nm) as the sample support 68 . Chips with and without the SiN membrane were used for substrate-attached and freestanding samples respectively. Images were post-processed using DigitalMicrograph (Gatan, Inc.) and ImageJ 69 software.

Data Availability
Detailed TEM images (as recorded) are available from SS on request.