Introduction

Graphene, a single atomic layer of carbon connected by sp2 hybridized bonds, has attracted intense scientific interest since its recent discovery1. Much of the research on graphene has been directed towards the exploration of its novel electronic properties which open up new avenues to both exciting experiments in basic science2,3,4,5 and electronic applications6. Further experiments and novel devices could be envisaged but remain to be demonstrated due to technological challenges in fabrication such as the lack of precision for locating or growing graphene of a specific size on a substrate of choice.

Whilst significant strides have been made in understanding graphene synthesis7, the mechanisms behind growth remain highly debated. Graphene growth cannot be captured by a universal mechanism with specific routes and conditions but a variety of synthesis strategies and growth modes exist. The best-known mechanism is the use of metal catalysts whereby free carbon radicals are formed, carbon is dissolved in the catalyst and finally precipitates at the surface. The free carbon radicals usually are supplied from a hydrocarbon feedstock, but there also are a few reports where the carbon feedstock is provided by a-C5,6,7,8,9,10.

Another surface that can provide suitable sites for growth is a bulk oxide support without any metal catalyst present where the carbon precursor is supplied by a hydrocarbon feedstock11,12,13,14. In the case of graphene growth from stable oxides as the support material, carbon dissolution is unlikely and therefore the growth probably occurs through surface diffusion processes. Oxides without a metal catalyst can also be used for the growth of carbon nanotubes (CNTs)15,16.

The growth of sp2 structures without a catalyst relies on a mechanism that largely remains to be understood17. Another example of such a process is the formation of CNTs on the cathode in the arc-discharge route which can occur without catalyst addition above 4000°C18,19,20,21. More recently, other growth routes without catalyst have emerged such as the formation of CNTs on graphitic surfaces22,23, the substrate-free gas-phase synthesis of graphene sheets24, or the growth of graphene sheets by microwave chemical vapour deposition (CVD)25.

Recently, the non-catalytic graphitization of a–C into small (~10 nm) polycrystalline graphene26 and into additional shells on multi-walled (MW) CNTs21,27,28 by current-induced annealing of graphene or of MWCNTs, respectively, has been reported. Moreover, catalyst-free crystallization of a–C nanowires led to the formation of tubular graphitic shells with nano-onions in their interior29. Unfortunately, the quality of all these graphitized nanostructures was rather poor as compared to arc-discharge grown CNTs or mechanically exfoliated graphene, presumably because temperatures were insufficiently high (below 3000°C) to induce perfect graphitization21,30. Additionally, a recent theoretical report points towards template assisted graphene growth31.

In this Article we report on in-situ transmission electron microscopy (TEM) studies of the structural changes that lead from a–C to crystalline graphene patches of over 10 x 10 nm in size and to even larger patches up to 100 x 300 nm. Furthermore, we use molecular dynamics (MD) simulations in order to get more insight in the process that transforms a–C to graphene when on top of a graphene substrate. Both in experiment and in theory, we observe that small a–C clusters on top of a graphene substrate rearrange and crystallize into graphene patches. The process is aided by the high temperatures involved and by the van der Waals32 interactions with the substrate. Finally, in the presence of a–C, graphene can grow from the borders of holes and form a seamless graphene sheet, a novel finding that has not been reported before and that is reproduced by the simulations as well.

Results

We perform in-situ current-induced annealing of suspended graphene devices by taking the samples to the high bias regime, specifically up to 2 – 3 V by stepwise increasing the voltage bias in 10 mV steps33,34, please see the Methods Section for further details. In this regime, the samples are at such a high bias and, therefore such a high current is flowing through them, that they are close to a complete and irreversible electrical breakdown and we start to sublime different regions and layers of the graphene device35. Via this procedure contaminants from the fabrication process are also effectively removed as they were on top of the removed layers35 and we observe that we obtain atomically clean graphene devices, as can be resolved from TEM imaging. After the current annealing process, the bias is taken back to 0 V and the samples cool down. Exposure of the cold samples to the electron beam allows us to intentionally deposit a–C on the previously clean graphene surfaces36, see figure 1 and section S1 in the Supporting Information. The carbon source originates from the beam-aided decomposition of hydrocarbons in the TEM column and/or from organic impurities adsorbed on the chip, the chip carrier and the sample holder. The regions where the a–C preferentially deposits are the edges of the individual layers in few layer graphene flakes, edges and other defects, fig. 137. Amorphisation of the graphene sheet because of disorder introduced by the electron beam is unlikely at an acceleration voltage of 80 keV, which is below the “knock-on” damage threshold of carbon nanostructures, see section 3 in the Supporting Information38,39. Thus graphene sheets remain stable and defect free in clean regions40. However, holes can form in contaminated areas by beam-driven chemical modifications with contaminants and adsorbates at energies below the knock-on threshold41. These holes seem to concentrate around edges and other defects, fig. 1 b, c37. Interestingly, during the current-annealing process itself, we never observe deposition of a–C. Presumably, hydrocarbon precursors for a–C formation desorb before being able to reach the graphene flake due to the high temperatures and deposit on colder areas around the hot graphene.

Figure 1
figure 1

TEM images of the stepwise deposition of a–C on an initially clean graphene sheet due to imaging.

The scale bar is 20 nm. (a) Preferential deposition of a–C at edges of other graphene layers. The a–C is the darker and rough surface. (b) Formation of holes (bright spots in the images, marked with arrows) and further deposition of a–C. (c) Growth of holes (marked with arrows) and almost complete coverage of a–C on the graphene template.

Full size image

After deposition of a–C on the previously atomically clean graphene surfaces, the samples are brought back once more to the high bias regime, specifically up to 2 – 3 V and current-annealed again without reaching the high-current limit where also the graphene substrate starts to sublime35. We proceed by stepwise increasing the voltage, wait for changes to occur and then slowly increase the bias voltage further. Temperatures as high as 2000°C26,42, or even 3000°C29, have been estimated to be reached due to Joule heating. Surprisingly, during this process we observe that it is not possible to sublime the a–C but instead it gradually transforms into graphene patches. We have observed the transformation of a–C to graphene by current-induced annealing on 15 of 15 samples where a–C had been intentionally deposited by the TEM beam. Figures 2 and 3 illustrate the evolution of the process from amorphous matter to crystalline few-layer graphene terraces.

Figure 2
figure 2

TEM images of the stepwise transformation of a–C into few-layer graphene terraces by means of current annealing.

(a) a–C on graphene. The scale bar is 20 nm. (b) Gradual crystallization of the a–C through a glass-like phase (2.2 V, 0.4 mA/μm). (c) Transformation into graphene patches (2.68 V, 0.8 mA/μm). The time elapsed between frame (a) to (c) is 5 min and 30 seconds.

Full size image
Figure 3
figure 3

Aberration corrected HR-TEM images of the stepwise transformation of a–C into graphene patches by means of current annealing at 3.32 V, 0.55 mA.

(a) a–C on graphene. The scale bar is 2 nm. (b,c) Gradual crystallization of the a–C through a glass-like phase, see figure S6 in the Supplementary Information for further illustration of this process39. (d) Transformation into highly ordered graphene patches. The time elapsed between the four frames is 22 minutes.

Full size image

Due to the nature of our in-situ TEM experiments, we can unequivocally testify to the circumstances during growth by performing atomic resolution imaging. Small a–C clusters rearrange and crystallize due to the high temperatures reached during current annealing without the involvement of any catalyst. Before reaching temperatures high enough to sublime the a–C, it gradually rearranges into high-quality graphene, see figures 3 and 4. The supplementary information (SI) contains low magnification TEM images (section S-4) and a video (movie S-1) of a different device where an overview can be obtained regarding the gradual transformation of a–C to graphene by current-induced annealing.

Figure 4
figure 4

HR-TEM images of the transformation of a–C to graphene and the graphene supporting layer.

(a) Aberration corrected HR-TEM image of the graphene support (top) and a graphene patch grown on top of it (bottom) from a–C by means of current annealing. The scale bar is 2 nm. The contrast of the micrograph was enhanced through Wiener filtering to suppress noise. Note that obtaining high resolution TEM images at such high bias voltages is challenging due to thermal vibrations. (b) FT of the initial graphene layer in the upper part of the TEM image marked by a square. (c) FT of the whole micrograph containing the graphene support layer and the graphene patch grown on top (circles), which are rotated with respect to each other by 22°.

Full size image

Based on high resolution (HR) TEM (see Figure 4) we were able to confirm that indeed the newly grown patches are graphene. From the corresponding Fourier transform (FT) in Fig. 4 (b) we can obtain the typical lattice parameter of graphene and the orientation of the newly grown layer which is rotated by 22 degrees with respect to the substrate. These patches can reach more than 100 nm x 100 nm in size (see figure S-4). The fact that we obtain a clear FT signal from an overlayer of an area of approx. 30 nm2 suggests that the graphene is not disordered and has a “long range order”, i.e. consists of a single grain.

Another interesting finding is that in the presence of a–C at high bias it is possible to repair holes in the graphene lattice. In Figure 5 we observe that holes formed by the reaction of contaminants with graphene due to the electron beam41, are self-repaired by growing new graphene healing the holes. Recently, it was found that multi-vacancies in a graphene lattice can be quickly reoccupied by C ad-atoms and graphene can recover its crystallinity. This repairing mechanism works best at temperatures above 600°C and was attributed to lattice reconstructions36 but can also occur spontaneously43. Healing of multivacancies in carbon nanotubes with up to 20 missing atoms can also be achieved with lattice reconstructions due to the TEM beam44. The holes in our graphene lattice are much bigger and can have diameters up to 5 nm37. Indeed, figure 5 suggests that the holes are closing step by step, presumably by the formation of new bonds with carbon radicals originating from the a–C. In this sense, the healing mechanism of the holes can be understood as substrate-free growth starting from the borders of holes in the graphene lattice by carbon atom addition to the reactive dangling bonds at the edges.

Figure 5
figure 5

Aberration corrected HR-TEM images of the gradual healing of a hole in the graphene lattice by means of current annealing of graphene in the presence of a–C at 2.75 V, 2.2 mA.

The arrows point to the initial hole (a), that gradually gets smaller (b) until it completely heals out (c). The scale bar is 2 nm. The time elapsed between the three frames is 30 s.

Full size image

Low magnification TEM images (figure S-4 in the electronic supplementary material) and a video (movie S-2) of a different device showing an overview regarding the gradual healing of holes in graphene in the presence of a–C by current-induced annealing can be found in the SI39.

Discussion

In order to shed light on the catalyst-free transformation mechanism of a–C into graphene and to explore the role of the graphene substrate, we performed molecular dynamics simulations of a perfect graphene substrate and four a–C clusters of 1 nm diameter on top at a distance of ~3.5 Å (Figure 6); for more details see the supporting information. During the MD simulations, the graphene substrate and the four a–C clusters on top are subject to stepwise increasing temperatures, namely, 300, 600, 1200 and 1800 K. This scenario resembles our experimental procedure well. In the experiments, a temperature distribution with a maximum close to the center of the graphene flake is present because the heat is only evacuated through the electrodes.

Figure 6
figure 6

Molecular dynamics simulations of the stepwise transformation of a–C to graphene.

(a) Initial 4 a–C clusters on top of a graphene unit cell marked by a yellow square. (b) Intermediate stage after annealing at 1800 K. (c) Structural shape after further a–C addition at 1800 K. The time elapsed between each frame is ~50 ps (see text for details). Further annealing at 1800 K with more carbon feedstock helped to overcome pentagon/heptagon defect structures, as high temperatures help to overcome Stone-Wales barriers to obtain hexagonal graphene lattices31. (d) Zoom in into the red square in panel (c) displaying a perfect graphene region without defects.

Full size image

In our theoretical model we assume a constant temperature, which is a reasonable approximation for the small windows used in the simulations, since within the hot spot the thermal gradient is small. Upon increasing temperature, the a–C starts transforming, goes through a glass-like phase in the range 600 – 1200 K and finally forms a graphene structure at 1800 K, see figure 6b37. Indeed, the need for higher temperatures to overcome Stone-Wales barriers to obtained perfect hexagonal graphene lattices has recently been reported31. The structure formed is flat and is located 3.5 Å above the initial graphene template. Nevertheless, holes are still present because insufficient carbon feed material was available to grow graphene over the whole area of the underlying graphene template. Upon further addition of a–C at 1800 K the graphene structure grows and, defects are progressively healed out (Fig. 6c). Indeed, some areas display defect-free graphene such as in figure 6d. The remaining holes and defects could be healed by further addition of a–C and a longer annealing time but this would require excessively long calculation times. In movie S-3, which shows the transformation of a–C to graphene, one can observe that only in the final stages of the growth process when the graphene precursor flakes merge into a bigger unit, they become stationary on the surface. This is due to the energy gain provided by the π-π coupling, suggesting that any atomically smooth substrate could serve as a template.

Recently, similar MD simulations modeled the synthesis of fullerenes45,46,47. An important difference between those simulations and ours is that in our case there is a graphene substrate while the fullerene synthesis was obtained in a substrate free model. Regardless of the initial geometry and velocity of the a–C, at the end of the MD runs, graphene is reproducibly formed on top of the graphene template. Although the substrate is only weakly coupled to the a–C, it apparently strongly influences the transformation of the nanostructure on top of it and prevents the formation of fullerene-like structures, demonstrating its influence on the formation of graphene. Experimental evidence that confirms the graphene template is only weakly coupled to a–C is given by the fact that the newly grown graphene patch is rotated with respect to the initial support layer, see figure 4. Again, these results suggest that in a more general picture our graphene growth method is universal for atomically smooth template-supported processes such as graphene, hBN or other two-dimensional layered materials such as MoS2. Indeed, a similar “substrate effect” has recently been recently reported for the growth of graphene on Ni31. In this theoretical study, the authors have predicted that Haeckelite is preferentially nucleated from ensembles of C2 molecules on a clean Ni(111) face, with graphene as a metastable intermediate phase. To the contrary, in the presence of a coronene-like C24 template, hexagonal ring formation is clearly promoted and finally anneal to graphene. Experimentally, in another study it was possible to grow graphene on hBN by CVD48, further supporting the universality of this growth method on atomically flat 2D systems.

To gain further insight into the experimental observation of healing holes, we performed MD simulations. We create a hole with a 1 nm radius in an ideal graphene flake and place 3 a–C clusters (of 1 nm3 size each) on top of it, see figure 7. We then heat our system to 1800 K. First, long fibers and big polyedres (C8-10) are formed across the hole. With further annealing, the hole is healed completely. For several independent runs with different initial structural and velocity conditions it takes 25 – 30 ps to completely heal the hole. The newly grown graphene contains at least one Stone-Wales defect (two pentagons (C5) and two heptagons (C7) forming a double pair)37, which we anticipate would fully heal out if significantly longer simulation time were available. This process of graphene forming in and healing a hole can be seen in movie S-439.

Figure 7
figure 7

Molecular dynamics simulations at 1800 K describing the hole healing process.

a–C is represented by the semi-transparent spheres. (a) Initial configuration displaying a hole in the graphene lattice. (b) Intermediate state of the structural transformation. (c) Repaired hole. The colors represent the actual number of vertices of the rings and their bending order. The full color code is described in Table 1, page 134 in Ref. 52.

Full size image

Interestingly, running the same MD simulations at temperatures below 600 K instead of at 1800 K does not yield healed out holes, suggesting that higher temperatures are required for repairing holes. Effective changes in a reasonable timeframe for the simulations (around 10 ps) only take place above 600 K. Due to the high temperatures reached during current annealing, untangled and unsaturated bonds from the a–C diffuse on the graphene and act as a source of radicals. They react with the dangling bonds at the edge of holes, gradually healing them out forming a new graphene lattice.

The speed of the transformation from a–C to graphene or the growth of graphene in holes in our MD simulations can be markedly fast, down to about 50 ps. However, the time elapsed for the transformation of a–C to graphene observed experimentally takes up to 1 – 15 minutes. The large difference in velocity between the experiments and simulations is attributed to the difference in the system dimensions, note that the graphene patches in the simulations are approx. 5 x 5 nm2 in size and still contain many holes and defects, which would heal out if the simulation times could be significantly increased. In contrast, the patches grown experimentally can be as big as 100 x 300 nm2, see figure S5 in the Supporting Information39. Moreover, the experiments were conducted as slowly as possible to prevent bringing the sample to an excessively high bias where the transformation would occur quicker but the risk of a complete electrical breakdown of the sample is larger.

In conclusion, our in situ real-time TEM observations correlated with MD simulations shed light on the catalyst-free transformation of a–C to flat graphene sheets. Small a–C clusters rearrange and crystallize into graphene at high temperatures on a graphene substrate or from the edges of holes. This finding opens up new avenues for engineering novel graphene-based devices in which additional graphene layers are needed on top of a graphene substrate49,50. To this end, clusters of a–C could be deposited on specified locations on top of graphene via e-beam deposition51 and then be transformed to additional graphene patches in-situ by a further (current-)annealing step. In a more general picture, our graphene growth method seems to be universal for atomically smooth template-supported processes such as graphene, hBN or other strongly layered 2D crystals such as MoS2 or NbSe2. Transforming a–C to graphene could open up new avenues for novel devices consisting of graphene on top of 2D materials of choice.

Methods

Chips with single-layer and few-layer graphene flakes supported by metal contacts were mounted on a custom-built sample holder for TEM with electric terminals, enabling simultaneous TEM imaging and electrical measurements. For imaging, a FEI Titan3 80–300 transmission electron microscope with a CEOS third-order spherical aberration corrector for the objective lens was used. It was operated at an acceleration voltage of 80 kV to reduce knock-on damage. All studies were conducted at room temperature with a pressure of approx. 10−7 mbar. The graphene device fabrication and measurement procedures are described in detail in references 32 and 53. In brief, a graphene flake is transferred onto Cr/Au electrodes that are freely suspended over an opening in a Si/SiO2 wafer. The device is voltage biased and the current is measured. In total, we measured 15 devices, with spacings between the electrodes between 1 and 20 μm.