A vertical silicon-graphene-germanium transistor

Graphene-base transistors have been proposed for high-frequency applications because of the negligible base transit time induced by the atomic thickness of graphene. However, generally used tunnel emitters suffer from high emitter potential-barrier-height which limits the transistor performance towards terahertz operation. To overcome this issue, a graphene-base heterojunction transistor has been proposed theoretically where the graphene base is sandwiched by silicon layers. Here we demonstrate a vertical silicon-graphene-germanium transistor where a Schottky emitter constructed by single-crystal silicon and single-layer graphene is achieved. Such Schottky emitter shows a current of 692 A cm−2 and a capacitance of 41 nF cm−2, and thus the alpha cut-off frequency of the transistor is expected to increase from about 1 MHz by using the previous tunnel emitters to above 1 GHz by using the current Schottky emitter. With further engineering, the semiconductor-graphene-semiconductor transistor is expected to be one of the most promising devices for ultra-high frequency operation.


8.
The authors also mentioned that it would require 1μm thick emitter to achieve THz operation. Would it increase the emitter transit time?
Reviewer #2 (Remarks to the Author): The paper reports on a successful fabrication of vertical silicon-graphene-germanium transistors with performance that notably surpasses other results published so far. Based on the measured data, the authors predict for such devices high-frequency operation with cut-off frequency of 1 THz at 1 kA/cm<sup>2</sup>, that is, at current levels comparable to those achieved by the devices they made. This makes the concept of a graphene base transistor (GBT), i.e., of a transistor in which the input-output current flows in the direction normal to the graphene sheet, a realistic alternative to the usually studied concept of a graphene field-effect transistor, i.e., of a transistor in which the input-output current flows in the plane of the graphene sheet. This achievement is therefore of sufficient novelty and of sufficient general interest to justify the publication of the report in Nature Communications.
Nevertheless, my opinion is that the distribution of accents in the main message as well as in the write-up should be optimized. The paper would also benefit from some improvements in the presentation of the acquired data and in the discussion. There are several points to be addressed, so I recommend publication of the paper after a major revision.
1. The importance of the hot-electron character of the device is stressed in the title, and from the lecture of the text the reader gets the impression that the main message is performance improvement by moving grom tunnel emitters to Schottky emitters. Personally, I would not put too much weight to these two aspects, for the following two reasons: -First, the base transition time is already negligible in a GBT due to extremely narrow base formed by a single graphene sheet (and the authors note this in the manuscript, summarily ignoring this time in their analysis). In fact, the potential mismatch between the emitter and graphene making the electrons in graphene hot may be her more a hindrance than a help, because it increases the quantum scattering of electrons back to the emitter. So I would treat the role of hot electrons with care, not over-stressing their beneficial contribution to the high-frequency performance of the device.
-Second, the difficulties posed by the tunnelling process have been recognized already in the first GBT publication [25], where the prediction of the cut-off frequency of about 1.5 THz was done assuming the emitter built from 3 nm of undoped Ge on Er2Ge3 electrode (though lines 43-45 and 197-198 of the manuscript properly comment on the related problem with the capacitance). Schottky barriers in the emitter and collector electrodes have been afterwards analysed in 2013 by theorists from Bologna: Di Lecce, IEEE Trans. Electr. Dev. Lett. 60, 3584 (2013) and Di-Lecce, IEEE Trans. Electr. Dev. 60, 4263 (2013); their work should most certainly be cited in the final version of the paper. The important progress reported by the submitted manuscript is in my view that such devices have finally been fabricated and the superior performance confirmed, as the authors properly note in the last sentence of the Abstract.
2. For the reasons explained above, I would recommend to put much more weight on the analysis of the achieved and prospective device parameters, in particular on the estimate of the cut-off frequency and on the perspective of this estimate being valid for useful devices working in the common base configuration. Towards this end, I propose to consider the following: -Lines 61-78 on device fabrication. Since the fabrication involves transfer of graphene and then of a SOI membrane, it is to be expected that the surfaces of Si and possibly also of Ge are covered by a chemical oxide. Such an oxide is reduced by the metal used to make a metal-semiconductor Schottky diode, but this is not the case here, because graphene does not reduce anything. Is there any information available on the interfacial chemistry? To what extent does this oxide have an effect on high-frequency performance of the device?  Fig. 2b is as high as 1.5 eV, and this is already at 3 V bias, that is, already including Schottky barrier lowering. This is much higher than a barrier expected on a semiconductor with 1 eV energy gap, and indeed exceeds the values reported for n-Si/graphene Schottky barriers, which are in the range of 0.3 to 0.9 eV; see e.g. the review by Di Bartolomeo, Physics Reports 606, 1 (2016). So I would recommend to plot the supplementary Fig. 2 in the traditional Schottky coordinates, ln(J/T</sup>2<sup>) vs. sqrt(V)/T, so that the accuracy of the barrier extraction can be better appreciated. If the barrier is indeed high, could it be attributed to the presence of the interfacial oxide? The barrier estimated for the Ge collector is lower (albeit also somewhat high, given the smaller band gap of Ge); can this be attributed to less efficient oxidation of Ge?
-Lines 101-107, on the cut-off frequency. The cut-off frequency is obtained from the emitter characteristics assuming the demonstrated ballistic transport (current gain close to 1) at sufficiently high emitter-base voltages. While this is correct, please note that the fabricated device has high input and low output impedance, that is, its power gain is much lower than 1 when Si is considered as the input and Ge as the output terminal. On the other hand, Fig. 3b indicates that reversing the role of these terminals would also lead to a working device: due to quantum capacitance effects in graphene, one can still control the Si-graphene potential difference by varying the POSITIVE bias on Ge, so that reasonable power gain can be obtained. But this rises the question of the role of the leakage current in the magnitude of this effect; the authors may want to address this more directly, which might be facilitated by providing plots showing base and collector currents and also plotting the input characteristic showing in this case IC versus VCB at various constant VEB (there one can readily see, e.g., if and how much does the separation between the IV lines drawn for various VEB depend on the choice of VCB, that is, by how much does the collector indeed influence the emitter). How would the output characteristics (families of IE vs VEB at constant IC) in such a configuration look like?
-Lines 191-198, on predicted THz operation, and the corresponding supplementary page. Please provide more information on the estimate, including the formula and the assumed and measured quantities that enter into it, so that one can more easily appreciate the role of the measured data and of the assumptions in this estimate. In particular, how was the ideal Si-Gr emitter capacitance at 1000 A/cm<sup>2</sup> calculated?
-Line 186 (Fig. 4) and the related discussion. The Figure is valid for small voltages, much lower than the several volts that must be applied between graphene and the terminals to achieve high frequency operation.
-Lines 258-262, characterization. How many working devices have been successfully fabricated and characterized? Are the reported results typical for many working devices? How reproducible is the process?
4. The paper is in general clearly and nicely written, but it should be linguistically improved in a few places. For example: -Lines 37-38: "atomic thickness /.../ of graphene will benefit /../ the base resistance". Actually, it degrades the base resistance, because the base is so thin. So this sentence should be split into two: one about the thickness and the current gain, and the other about the mobility and the base resistance.
-Line 128: "Previously those scattered electrons which could not cross the collector barrier can now tunnel under the barrier". The word "previously " is superfluous or misplaced (BTW, "which" should be preceded by a comma).
Reviewer #3 (Remarks to the Author): The paper describes a novel silicon-grapheme-germanium hot electron transistor with record performance. Before the paper can be accepted, the following issues need to be addressed: 1/On p. 4 it is stated that with a positive voltage applied to the emitter junction, electrons are emitted... In my opinion, this statement is wrong or at least not carefully formulated.
Electrons will be emitted from the emitter when a negative bias is applied to it.(see also Fig. 2a).
2/ The barrier of 1.47 eV seems to be rather high in my opinion. How do we have to interpret this in terms of electron affinity and work function of silicon and grapheme? In addition, barrier seems to be dropping at higher temperatures. Can you explain this degradation of the leakage current for higher temperatures?
3/ The Authors state that the depletion in the silicon emitter can easily reach several micrometers. However, in the present structure, with a thickness of 880 nm, the silicon emitter is in my opinion, for the quoted doping density, fully depleted. This should also modify the band structure of the transistor in Fig. 1e. It has perhaps also consequences for the barrier height. 4/ I do not understand the statement on p. 12 that ohmic contacts were obtained due to the generation current at the etched surface. It is also not clear from the supplementary material ( Fig.  1). In Fig. 1c, one can indeed observed that after prolonged etching (7 min) ohmic behavior is found. But I do not agree with the explanation or at least I do not understand it. It is known that it is easier to make a good ohmic contact to a damaged surface, by defect-assisted leakage, so this could be the reason. At the same time, it is known that dry etching damage can lead to the introduction of shallow donors, so that the doping density can be changed. By the way, I presume that the silicon membrane is p-type (the SOI film?).
I would suggest to verify the resistivity of the membrane if possible to check whether there are some changes in the net doping density. This will also have a drastic impact on the fully or partially depleted nature of the 880 nm film. minor detail: p. 2 supplementary line 33: can now tunnel through the collector barrier to be collected. This is a strange sentence -perhaps to be collected can be removed or replaced.

Responses to the comments of reviewers
Reviewer #1: The authors fabricated a graphene-base transistor with Si emitter and Ge collector.
The main claims of the paper are record current density and potential for THz operation.
However, I find that the claims are not well supported by the experimental data and are incorrect in some cases. The manuscript is not suitable for publication in the current format. Here are my comments,

Response:
Thank you very much for your review. In accordance with the reviewer's valuable and insightful comments, we have carefully revised our manuscript and addressed all the concerns of the reviewer in the revised manuscript. Details are discussed as follows.

Response:
As the reviewer pointed out, a pioneering theoretical study on graphene-base heterojunction transistors (GBHT) has been done with a device structure of silicongraphene-silicon by Di Lecce et al 1 . In our study, we have demonstrated the first vertical silicon-graphene-germanium (Si-Gr-Ge) transistor experimentally. In the theoretical study of Di Lecce et al. 1 , using neglected contact resistance and ideal silicon-graphene interface, as well as heavily-doped silicon emitter (3×10 19 cm −3 ) and collector (10 18 or 3×10 18 cm −3 ), the THz operation was expected when the collector current was larger than 10 6 A/cm 2 . Indeed, there exists a performance gap between the theoretical results and our experimental results, which is mainly induced by the key factors of contact and interface quality, as well as emitter doping and collector selection. Based on our experimental results, an ideal case, ignoring the series resistance, with an ideal interface was considered for a transistor with a heavily-doped n + -Si-Gr emitter and a thin collector, and the THz operation is expected with an emitter current of about 4.7×10 6 A/cm 2 , which is consistent with the theoretical predictions 1 reported by Di Lecce et al.
The details are presented as below.
As shown in Fig. R1a, the experimental data (black solid line) of the I-V characteristics of the Si-Gr emitter was fitted using a Schottky junction model without considering series resistance J=J0[exp(V/Vt)1], where the fitted leakage current density is J0=3.3×10 5 A/cm 2 , the ideality factor =1.85 and Vt is the thermal voltage (red dashed line). When the voltage is larger than 0.2 V, the model deviates from the experimental data because of the series resistance. An ideal case was considered. The series resistance was ignored and an ideal interface with unit ideality factor =1 was assumed (blue dash-dot line), which further increases the current and conductance. where  is the relative dielectric constant, 0 is the permittivity in vacuum, W is the width of the depletion width for Si, q is elementary charge, Nd is the doping concentration of Si of 8×10 19 cm 3 , qψbi is the potential barrier height at the Si side when no bias is applied, Vt is the thermal voltage, qBn is the Schottky barrier height of about 0.64 eV, Vn is the Fermi potential of the Si, and Nc is the effective density of states in conduction band of Si. The emitter charging time e can thus be calculated ase=Ce/gm. A small collector delay timec can be achieved by heavily-doped semiconductors 1 or thin collector semiconductors such as multilayer 2D materials 2,3 . Assuming a thickness of χ=5 nm and a saturation velocity 4 ν=4×10 6 cm/s gives the collector delay time 19 c=χ/(2ν)=6.25 × 10 14 s. The alpha cut-off frequency f was estimated by f=1/(2(e+c)) with ignored base transit time b. With a current of about 4.7×10 6 A/cm 2 , f of about 1.06 THz is obtained (Fig. R1b), which is consistent with the theoretical predictions, demonstrating that the graphene-base heterojunction transistor has a potential application in the THz operation 1 .
Overall, the gap between theoretical and experimental results can be improved by contact engineering (such as using heavily-doped semiconductor and ion implantation) and interface engineering (Supplementary Discussion 1). A heavily-doped emitter is needed for higher current and conductance, and a heavily-doped collector or a thin 2D material collector should also be used to reduce the collector delay time.

Response:
We thank the reviewer very much for kind suggestion. Compared with the previous Gr-on-Si junctions, our work first reported the Si-membrane-on-Gr junction. The ideality factor and the Schottky barrier height of our Si-Gr junction are consistent with the results of previous Gr-on-Si junctions reported in literature [5][6][7][8][9][10][11][12][13] as listed in Table R1. In order to determine the Schottky barrier height, a Si-Gr junction with Au electrodes was fabricated and the traditional temperature-dependent I-V measurements of the forward current were carried out (Fig. R2a). The currents at a small forward bias of 0.1 V (0.1 V at n-Si side) were considered to avoid the effect of series resistance.
The following model was used to fit the relationship between current and temperature: is elementary charge, qBn is Schottky barrier height, V is voltage,  is the ideality factor of 1.85, and k is the Boltzmann constant. qBn was fitted to be ~0.64 eV at room temperature ( Fig. R2b). As temperature decreases, the thermal emission becomes weaker and the tunnel current starts to emerge which shows a weak temperature dependence. 3. I'm not convinced that the contact between probe and Si is ohmic. The resistivity of the Si is too high (6 ohm.cm) to make ohmic contact just by landing probes on it.
The I-V presented in the supplementary is linear but current level is too low compared to graphene if I assume same area. The authors need to fabricate proper ohmic contact to study the device operation. Otherwise, the device operation is limited by the contact resistance between probe and Si.

Response:
Thank you for your valuable suggestion. According to this suggestion, we fabricated Au electrodes after the surface of Si was etched using RIE (Fig. R3). A transfer length method test 14 was carried out to investigate the Ohmic contact between Au and RIE-etched Si surface (Fig. R4a). The I-V characteristics between adjacent Au electrodes indicate an Ohmic contact (Fig. R4b). More details are discussed as follows.
The resistance and distance have a linear relationship as fitted (dash line in Si membrane in the transistor was estimated to be c/S=402 Ω, which is consistent with the total series resistance (~426 Ω) for the Si-Gr emitter at 5 V (Fig. 2a). This contact resistance leads to an improved on-current of 692 A/cm 2 at 5 V, compared with 470 A/cm 2 at 10 V without Au electrode.  As the reviewer pointed out, the previous current is relatively small. This is because that the Si is lightly-doped and the current conduction route has a long distance and a small cross section when the surface of the top Si layer of the SOI wafer is examined by probes directly.

Response:
Thank you for the valuable comments. Indeed, the tunnel emitter mentioned by the reviewer showed a larger current at small voltages 2,3 . However, the on-current of our Schottky emitter can now reach 692 A/cm 2 at 5 V which is the largest reported current value compared with the tunneling counterparts 2,3 . According to the reviewer's suggestion, we will not emphasize on the "largest" current, and only state the fact that a current of 692 A/cm 2 is achieved at 5 V.
Besides the large current, our Schottky junction also has a small capacitance (~28 nF/cm 2 ) since Si is lightly-doped. In contrast, the capacitances for the tunnel emitters are at least one-order-of magnitude higher (~325 nF/cm 2 for 13.6 nm h-BN 2 and ~2.66 μF/cm 2 for 3 nm AlN 3 ). The alpha cut-off frequency f is not only proportional to the emitter current and conductance, but also inversely proportional to the emitter capacitance. Our Schottky emitter has an advantage in f at any current value compared with the tunnel emitter 3 . f is estimated to be above 1 GHz for our Schottky emitter, while it is less than 1MHz for the tunnel emitter 3 (Fig. 2d).

Response:
We are sorry that the previous statement was not clear and led to misunderstanding.
The "prime" is actually a "comma", and "Ve, V" and "Vc, V" in the previous figures have been changed to "Ve (V)" and "Vc (V)", respectively.
We have changed all the "comma" to "brackets" to separate a variable and its unit in all the figures in the revised manuscript and supplementary information.

Response:
Thank you very much for the valuable comments. According to this suggestion, we have added current component discussions, more electrical characteristics and other explanations, which have improved the manuscript significantly. The details are discussed as follows.
Concerning the current gain, the current components are illustrated in Fig. R5.
Graphene (Gr) is connected to ground. Ie is emitted from the Si-Gr emitter. Part of Ie is collected at the Ge collector forming the effective collector current Ic', while the other part flows to ground forming Ib'. A leakage current Ileak of the collector junction also contributes to the collector current Ic and base current Ib, so that Ic= Ic'+Ileak, Ib= Ib'Ileak.
If the gain  is calculated without removing the effect of Ileak, it will be overestimated.
We calculated the gain after removing the effect of Ileak as =Ic'/Ie=(IcIleak)/Ie and Ileak=Ic (Ve =0).   Concerning the symbol for current definition, Fig. R8 shows I-V characteristics for the transistor with a heavily-doped germanium, in which the emitter and collector are exchanged so that electrons are emitted from n + -Ge (Vc < 0) and collected at Si (Ve > 0). In this paper, the current and voltage at the Ge (Si) side is denoted as Vc (Ve) and Ic (Ie), respectively, to avoid misunderstanding (especially considering a "reverse working mode" shown in Supplementary Figs. 12, 13).

Response:
Thanks for the valuable comment. As pointed out by the reviewer, InP HBT technologies have already achieved great progress toward THz operation [15][16][17][18] . In this study, the terahertz operation using the Si-Gr emitter in an ideal case was estimated based on the experimental results (Fig. R1). When the current is larger than 4.7×10 6 A/cm 2 , the alpha cut-off frequency f can exceed 1 THz, thanks to the negligible base transit time. If a current gain approaching 1 is assumed, fT is above 1 THz, which is larger than the experimental result of an InP HBT. As suggested by Di Lecce et al. 1 , the cut-off frequency of an HBT is ultimately limited by the transit time in the base, while a graphene-base heterojunction transistor has no such limitation. This is a key advantage over the HBT, however further effort is needed for its development. 8. The authors also mentioned that it would require 1μm thick emitter to achieve THz operation. Would it increase the emitter transit time?

Response:
Thanks for the valuable comment. Because the current mechanism of the emitter junction is not the drift mechanism, the emitter transit time is usually ignored and not mentioned 19,20 . For the collector junction in a normal n-p-n BJT, when it is reversely biased, an electron must go through the space charge region of the collector by drift, and a "transit time" is needed. However, for the emitter junction, when it is forwardly biased, the carriers go through the space charge region by diffusion. When the forward bias has a slight change, the delay time mainly comes from the time for the free carriers to charge the space charge region which is mainly e as discussed in the main text. After optimized. The paper would also benefit from some improvements in the presentation of the acquired data and in the discussion. There are several points to be addressed, so I recommend publication of the paper after a major revision.

Response:
Thank you very much for your positive comments.
1. The importance of the hot-electron character of the device is stressed in the title, and from the lecture of the text the reader gets the impression that the main message is performance improvement by moving from tunnel emitters to Schottky emitters.
Personally, I would not put too much weight to these two aspects, for the following two

Response:
Thanks for your valuable comments. As the reviewer pointed out, graphene itself is thin enough to give a negligible base transition time for a GBT, and this work can be considered as the first experimental demonstration of the theoretical work by Di Lecce et al 1 . According to this suggestion, the title of the paper has been changed to "A vertical silicon-graphene-germanium transistor". In the revised manuscript, in the introduction of a graphene-base transistor, it is emphasized that the graphene base is almost transparent to the electron transport leading to a negligible b because of the atomic thickness.
According to the reviewer's suggestion, a main claim of the paper has been modified as that the graphene-base heterojunction transistor (GBHT) theoretically proposed by Di Lecce et al. has been experimentally realized in a silicon-graphene-germanium transistor, and its potential for THz operation has been confirmed based on experiments and modeling. In the revised introduction, it is stressed that pioneering theoretical study on the GBHT has been done with a device structure of silicongraphene-silicon 1,21 , and theoretically the THz operation can be realized when the collector current is larger than 10 6 A/cm 2 . One of main contributions of our work is that we have demonstrated the first vertical silicon-graphene-germanium (Si-Gr-Ge) transistor, and with further engineering, the vertical semiconductor-graphenesemiconductor transistor is expected to be one of the most promising devices for ultrahigh frequency operation.

Response:
Thanks for the valuable comments. As the reviewer pointed out, the quality of the interface may influence the performance of the device. In this study, the Si membrane and Ge substrate were made by HF-based process, and their surfaces are hydrogenterminated. The XPS analyses were carried out to examine the surfaces of Si and Ge that were just cleaned by HF, and no obvious oxide was detected. This result indicates that such surfaces should be free of oxide even a transfer process is used, which is consistent with that reported by Kiefer et al. 22 On the other hand, the surface of the transferred graphene is contaminated by PMMA residue as reported previously 23,24 . This is expected to influence the series resistance and ideality factor of the emitter junction, and thus the on-current and highfrequency performance of the transistor. The strategy to improve the quality of the interface is discussed in Supplementary Discussion 1. Supplementary Fig. 14.

Corresponding modification has been made in
-Lines 80-89 (Si emitter), with reference to Fig

Response:
Thanks very much for your valuable comment. According to the reviewer's suggestion, we made new Si-graphene junctions with Au electrodes (Fig. 1), and traditional temperature-dependent I-V measurements of the forward current were carried out to determine the Schottky barrier height. To fit the Schottky barrier height, currents at a small forward bias of 0.1 V (0.1 V at n-Si side) were considered to avoid the effect of series resistance (Fig. R9a). The following model was used to fit relationship between current and temperature 19 : ln(I/T 2 )=Cq(BnV/)/k(1/T), where I is current, T is temperature, C is a constant, q is elementary charge, qBn is Schottky barrier height, V is voltage,  is the ideality factor of 1.85, and k is the Boltzmann constant. qBn was fitted to be ~0.64 eV at room temperature (Fig. R9b). As the temperature decreases, the thermal emission becomes weaker and the tunnel current starts to emerge which shows a weak temperature dependence. The value of qBn is consistent with the reported results (0.3-0.9 eV) 13 , demonstrating that no obvious oxide exists on the surfaces of Si and Ge as discussed above. The value of qBn is different from the previous result which was obtained from the analysis of the leakage current.
It is known that the leakage current is easily affected by the edge current 19 , which may result in the difference. The Schottky barrier height of Gr-Ge junction was also evaluated using the above method, and the obtained qBn is ~0.26 eV (Fig. R10).  The Schottky barrier height was fitted to be qBn=0.26 eV at a small forward bias of 0.1 V (0.1 V at n-Ge side) which is lower than that of the Si-Gr emitter junction.

Corresponding modification has been made in Page 6 (Lines 5-9) and Fig 2b
in the revised manuscript as well as Supplementary Fig. 3 and Supplementary Fig.   6 in the revised Supplementary Information.
-Lines 101-107, on the cut-off frequency. The cut-off frequency is obtained from the emitter characteristics assuming the demonstrated ballistic transport (current gain close to 1) at sufficiently high emitter-base voltages. While this is correct, please note that the fabricated device has high input and low output impedance, that is, its power gain is much lower than 1 when Si is considered as the input and Ge as the output terminal. On the other hand, Fig. 3b indicates that reversing the role of these terminals would also lead to a working device: due to quantum capacitance effects in graphene,

Response:
As the reviewer pointed out, there was a power gain issue of the previous transistor with a heavily-doped n + Ge collector (resistivity: ~0.01 cm), and a "reverse working mode" was suggested. We thank the reviewer very much for the valuable comments and suggestions. According to this comment, first we used a new n + -Ge collector (resistivity: ~0.1 cm) to rebuild the transistor, and a working region with power gain larger than 1 can be found. The electrical characteristics of the Si-Gr-Ge transistors were updated accordingly. Figure R11a-d and e-h show the performances of the transistor using a lightly-doped n-Ge collector (resistivity: ~1 cm) and a heavilydoped n + -Ge collector (resistivity: ~0.1 cm), respectively. The input conductance ge=dIe/dVe as well as the current gain α=dIc/dIe can be calculated from the input characteristics (Fig. R11f). For the transistor with a heavilydoped n + -Ge collector, the output characteristics with different voltages Ve are also shown in Fig. R12, from which the output conductance gc=dIc/dVc at different voltages Ve can be calculated. The power gain can thus be expressed as Ap=(dIcdVc)/(dIedVe)= (dVc/dIc)(dIe/dVe)(dIc/dIe) 2 =ge/gcα 2 . When Vc=3 V and Ve=5 V, the power gain Ap=1.5; and when Vc=3 V and Ve=4 V, the power gain Ap=1.6. Therefore, the region with power gain larger than 1 can be found.

Fig. R12
Output characteristics (Ic-Vc) of the transistor with a heavily-doped n + -Ge collector with Ve from 0 to 5 V.
Next, we explored the reverse working mode (Ge was used as input with a bias Vc>0, and Si as output with a bias Ve<0) of the transistor with a heavily-doped n + -Ge collector. The transistor was in the common base mode. Graphene was connected to ground. As shown in the input (Ic-Vc) and transfer (Ie-Vc) characteristics (Fig. R13a), when Vc increases, Ic increases, and Ie increases accordingly because of the quantum capacitance effect of graphene. This effect can also be observed in the output characteristics (Fig. R13b). The highest current gain dIe/dIc is about 77% when Vc=1.55 V and Ve=4 V. Ic in a the logarithmic coordinates and b the linear coordinates for the transistor in the reverse working mode (Fig. R13). Data for different collector bias Ve is denoted by different colors: black, red, blue, pink and green are for Ve=0, 1,2,3,4 V respectively. In particular, how was the ideal Si-Gr emitter capacitance at 1000 A/cm 2 calculated?

Response:
Thanks very much for your valuable comment. According to the reviewer's suggestion, more details about the estimation are given below. The Schottky barrier height is updated to be 0.64 V as measured. A heavily-doped n + -Si emitter was used in this estimation as the work of Di Lecce et al 1 .
As shown in  (Fig. R15b), which is consistent with the theoretical predictions, demonstrating that the graphene-base heterojunction transistor has a potential application in the THz operation 1 .
Since a heavily-doped Si is used to perform THz operation, the capacitance is about 19 F/cm 2 when f is 1 THz with a current of 4.7×10 6 A/cm 2 . The advantage of a Schottky emitter for THz operation over a tunnel emitter was discussed by Di Lecce et al. 1 . As predicted by simulations, for a tunnel emitter to realize THz operation, the emitter potential barrier which is between the emitter metal and the emitter-to-base tunneling layer should be lower than 0.4 eV, which remains an engineering issue.

Corresponding modification has been made in Page 13 (Lines 5-9 in red color)
and Supplementary Fig. 14 in the revised Supplementary Information.

Response:
In the revised manuscript, Reference [25] has been cited when the graphene-base transistor is introduced, and the advantage of the thickness of graphene to reduce the base transit time has been discussed.
Corresponding modification has been made in Page 2 (Lines 21-22 in pink color) in the revised manuscript.
-Line 186 (Fig. 4) and the related discussion. The Figure is valid for small voltages, much lower than the several volts that must be applied between graphene and the terminals to achieve high frequency operation.

Response:
Thanks for your valuable comments. Fig. R16 shows the energy band diagram of the Si-Gr-Ge transistor when a large bias Vbe is applied to the emitter junction. Not that most of the bias is applied to the series resistance of the junction. The energy band diagram with a small voltage bias is used to illustrate the quantum capacitance effect in the transistor which also exists when the voltage is large (Fig. 4).  4. The paper is in general clearly and nicely written, but it should be linguistically improved in a few places. For example:

Response:
Thank you very much for your positive comments.
-Lines 37-38: "atomic thickness /.../ of graphene will benefit /../ the base resistance". Actually, it degrades the base resistance, because the base is so thin. So this sentence should be split into two: one about the thickness and the current gain, and the other about the mobility and the base resistance.

Response:
Thank you very much for the valuable suggestion. The corresponding sentences have been modified as below in the revised manuscript: "Because of the atomic thickness, the graphene base is almost transparent to electron transport leading to a negligible b. At the same time, the remarkably high carrier mobility of graphene will benefit the base resistance compared with a thin bulk material."

(Lines 1-3) in pink color in the revised manuscript.
-Line 128: "Previously those scattered electrons which could not cross the collector barrier can now tunnel under the barrier". The word "previously " is superfluous or misplaced (BTW, "which" should be preceded by a comma).

Response:
According to the reviewer's kind suggestions, we have changed the statement as follows: "At the collector junction interface, around the top of the barrier, the tunneling distance of an electron decreases dramatically. As a result, electrons which originally cannot cross the collector barrier can now tunnel through the barrier, which increases the current gain." Corresponding modification has been made in Page 9 (Lines 1-4 in red color) in the revised manuscript.
Reviewer #3: The paper describes a novel silicon-grapheme-germanium hot electron transistor with record performance. Before the paper can be accepted, the following issues need to be addressed:

Response:
Thank you very much for your positive comments.

Response:
As the reviewer pointed out, the statement is not carefully formulated. The sentences have been modified as: "When the device is turned on with a positive voltage Vbe applied to the emitter junction (when graphene is grounded, Ve<0), electrons are emitted from the emitter, go through the emitter junction with a barrier height q1, the graphene base, then the collector junction with a lower barrier height q2, and eventually are collected by the collector." Corresponding modification has been made in Page 4 (Lines 12-16 in blue color) in the revised manuscript.
2/ The barrier of 1.47 eV seems to be rather high in my opinion. How do we have to interpret this in terms of electron affinity and work function of silicon and grapheme?
In addition, barrier seems to be dropping at higher temperatures. Can you explain this degradation of the leakage current for higher temperatures?

Response:
We thank the reviewer very much for valuable comment. More studies have been done about the measurement of the Schottky barrier height. It is well known that the leakage current is easily affected by the edge current 19 , which may influence the Schottky barrier height. Therefore, we have made new Si-graphene junctions with Au electrodes (Fig. 1), and instead of measuring the leakage current, traditional temperature-dependent I-V measurements of the forward current were carried out to determine the Schottky barrier height 19 . To fit the Schottky barrier height, the currents at a small forward bias of 0.1 V (0.1 V at n-Si side) were used to avoid the effect of series resistance (Fig. R18a). The following model was used to fit the relationship between current and temperature: ln(I/T 2 )=Cq(BnV/)/k(1/T), where I is current, T is temperature, C is a constant, q is elementary charge, qBn is Schottky barrier height, V is voltage,  is the ideality factor of 1.85, and k is the Boltzmann constant. qBn was fitted to be ~0.64 eV at room temperature (Fig. R18b). As temperature T decreases (and 1000/T increases in Fig. R18b), the absolute value of the slope of the fitted data decreases and thus the corresponding calculated Schottky barrier height decreases. This phenomenon indicates that the temperature dependence of the current becomes weaker. It may be caused by the fact that, as temperature decreases, the thermal emission becomes weaker, and the tunnel current starts to emerge, which shows a weak temperature dependence 19 . should also modify the band structure of the transistor in Fig. 1e. It has perhaps also consequences for the barrier height.

Response:
Thank you for the valuable comments. We have carried out more analyses on the space charge region and band structure. When no bias is applied, the width of the space charge region W can be calculated as W=(20/qNd(ψbiVt)) 0.5 , where Nd is the doping concentration of Si of about 1×10 15 cm 3 , ψbi=BnVtln(Nc/Nd) is the potential barrier height at the Si side, Vt is the thermal voltage, and Nc is the effective density of states in conduction band of Si. As shown above, the measured Schottky barrier height qBn is 0.64 eV. The calculated width of the depletion region is W= 684 nm, and therefore the 880 nm Si membrane is not fully depleted. Fig. R19 has been modified to better illustrate the band structure. Since W is less than 1 μm with the updated qBn, the sentence about that "the space-charge region of a Schottky junction can reach several micrometers in width" has been deleted in the revised manuscript. Corresponding modification has been made in Fig. 1 in the revised manuscript.
4/ I do not understand the statement on p. 12 that ohmic contacts were obtained due to the generation current at the etched surface. It is also not clear from the supplementary material (Fig. 1). In Fig. 1c, one can indeed observe that after prolonged etching (7 min) ohmic behavior is found. But I do not agree with the explanation or at least I do not understand it. It is known that it is easier to make a good ohmic contact to a damaged surface, by defect-assisted leakage, so this could be the reason. At the same time, it is known that dry etching damage can lead to the introduction of shallow donors, so that the doping density can be changed. By the way, I presume that the silicon membrane is p-type (the SOI film?). I would suggest to verify the resistivity of the membrane if possible to check whether there are some changes in the net doping density.
This will also have a drastic impact on the fully or partially depleted nature of the 880 nm film.

Response:
Thank you very much for pointing out the possible mechanism of Ohmic contact and valuable suggestions to investigate the doping concentration of Si. According to your comment and a reference 14   According to your suggestion, the resistivity of n-Si after RIE was investigated by a transfer length method test (Fig. R20) 14 . The I-V characteristics between adjacent Au electrodes indicate an Ohmic contact. The resistance and distance have a linear relationship as fitted (dash line in Fig. 20c). The slope equals to /(Wt), where  is the resistivity of Si, W is the width of Au electrode (93 m), and t is the thickness of the Si trip (about 880 nm) (Supplementary Fig. 15).  was fitted to be 4.5 Ωcm which is similar to the original SOI wafer (1-6 Ωcm), indicating that the RIE processes does not change the doping concentration of Si obviously.
capacitance is between 30 and 50 nF/cm 2 . At the forward bias of 0.30 V, the capacitance-frequency characteristics (Fig. R1b) shows that the capacitance is stable up to 100 kHz, and the measurement result is affected by interface and series resistance at frequency beyond 100 kHz 1 . Thus, the frequency used to measure the capacitance is selected to be 100 kHz. 2. The estimation of cutoff frequency above 1 GHz is oversimplified as the authors ignore the contribution of base-collector capacitance and significantly large series resistance present in the reported device. The authors need to report the experimental value of base-collector capacitance and series resistance to provide a realistic estimation of cutoff frequency of the presented device.

Response:
Thank you very much for the valuable comment. As the reviewer pointed out, an estimation of the alpha cut-off frequency f with more details and experimental evidence will further benefit the conclusion in the paper. According to your suggestion, f has been analyzed and estimated considering the experimental result of emitter capacitance, collector capacitance and series resistance as below.
Firstly, it should be emphasized that the f estimated here is the intrinsic f. Only delay time contributed by junction capacitance will be included, and the effect of the parasitic electrode capacitance should be excluded as discussed in a reference 2 . In our device, the large area Au and graphene on top of the 30-nm-thick Al2O3-Ge substrate (Gr/Au-Al2O3-Ge) ( Fig. 1) will induce large parasitic capacitance which is orders-ofmagnitude larger than the Gr-Ge collector junction capacitance. It is not useful for implementing device functions, and in a device for production, it can be reduced by for example, mesa Ge structure together with thick insulating layer and small base electrode. In the following estimation of f, this parasitic capacitance is excluded.

Estimation of e
The intrinsic alpha cut-off frequency f can be expressed as f=1/[2(e+b+c)].
The base transit time b is ignored (b=0) thanks to the atomically thin thickness of graphene. The emitter charging time e is calculated ase=Ce/ge where Ce=41 nF/cm 2 is the peak emitter capacitance (Fig. R1a) and ge is the emitter conductance which can be achieved from the I-V characteristics of the emitter junction (347 S/cm 2 at 5 V) (Fig.   2a), leading to a e=Ce/ge=118 ps at 5 V.

Estimation of cc
The collector delay timec can be expressed as c=ct+cc where ct is the collector (depleted region) transit time and cc is the collector charging time. We will first estimate cc with experimental results. cc is estimated by cc=rcCc where Cc is the collector capacitance and rc is the collector series resistance. As discussed, to estimate the intrinsic f, the effect of the parasitic electrode capacitance should be excluded, thus Cc cannot be directly measured from the Gr-Ge junction in the device, because the result will contain the parasitic electrode capacitance. At the same time, rc cannot be directly measured from the I-V characteristics of the Gr-Ge junction either, because the measured series resistance of the junction will contain not only the collector series resistance, but also the base series resistance.
To determine Cc and rc, we have fabricated Au(50nm)/Ti(5nm)-Ge Schottky junctions (Fig. R2a) using the same n + -Ge substrate as the device in Fig. 3e, where the area of Au/Ti is the same as the Gr-n + -Ge junction area, i.e. the area of the window of Al2O3 (Fig. 1b). We measure the series resistance (Fig. R2b) and the capacitance at a reverse bias of 4 V (Fig. R2c,d) of the Au/Ti-Ge junctions, and these results will be rc (about 5 Ω) and Cc (1.7 pF), leading to a cc=rcCc=8.5 ps. The reason is discussed below.

Fig. R2
Series resistance and capacitance measurement of a Au/Ti-Ge Schottky junction. a I-V characteristic of the junction. The Au/Ti is connected to the ground. Inset: an optical image of the junction (scale bar: 20 m). b Differential resistance r calculated from the I-V characteristic at a forward bias, which tends to the series resistance about 5 Ω. c The C-V characteristic of the junction at reverse bias at a frequency f of 100 kHz. As the reverse bias increases, C decreases because the depleted region width increases. d The C-f characteristic of the junction at a reverse bias of  V. The capacitance is stable up to 100 kHz. 100 kHz is selected.
1. The series resistance of the Au/Ti-Ge Schottky junction is the sum of a) the resistance integrated over the quasi-neutral region (between the depletion-layer edge and bottom Au electrode), b) the spreading resistance in the substrate, and c) the resistance due to the bottom Au ohmic contact with the substrate, if the resistance of Au(50nm)/Ti(5 nm) is ignored 3 . For the Gr-Ge junction, the collector resistance rc is also the sum of the above 3 items. Since the same Ge substrate with bottom Au electrode is used for the Au-Ge junction and the Gr-Ge junction and the junction areas are also the same, the latter two items (b and c) should be the same for Gr-Ge junction and Au-Ge junction. The width of the depletion-layer is different for the two junctions, but only with a difference less than 1 μm. The thickness of the Ge substrate is about 500 μm, thus the distance between the depletion-layer edge and bottom Au electrode is almost the same, and the first item (a) should also be the same. Based on the above analysis, the measured series resistance of the Au/Ti-Ge Schottky junction is the same as the Ge series resistance in the Gr-Ge junction rc.
2. At a reverse bias V, the capacitance of a Schottky junction can be calculated 3 as C=Ge0/W, W=(2Ge0/qNd(ψbi+VVt)) 0.5 , ψbi=BnVn, Vn=Vtln(Nc/Nd), where Ge is the relative dielectric constant of Ge, 0 is the permittivity in vacuum, W is the width of the depletion region, q is elementary charge, Nd is the doping concentration of Ge, ψbi is the potential barrier height at the Ge side when no bias is applied, Vt is the thermal voltage, qBn is the Schottky barrier height, Vn is the Fermi potential of Ge, and Nc is the effective density of states in conduction band. The same Ge substrate is used for Au-Ge and Gr-Ge junctions, thus the difference of the capacitance comes from Bn which is 0.52 V for Ti-Ge 4 and 0.22 V for Gr-Ge ( Supplementary Fig. 7). However, at a large reverse bias, for example V=4 V, the difference of Bn (0.30 V) can be ignored, thus the difference of W can be ignored. The measured capacitance of Au/Ti-Ge junction is the intrinsic capacitance Cc of the Gr-Ge collector junction.

Estimation of f
Based on the above experimental results and analysis, the intrinsic alpha cut-off frequency f has been updated in Fig. R3