Energy harvesting optical modulators with sub-attojoule per bit electrical energy consumption

The light input to a semiconductor optical modulator can constitute an electrical energy supply through the photovoltaic effect, which is unexploited in conventional modulators. In this work, we leverage this effect to demonstrate a silicon modulator with sub-aJ/bit electrical energy consumption at sub-GHz speeds, relevant for massively parallel input/output systems such as neural interfaces. We use the parasitic photovoltaic current to self-charge the modulator and a single transistor to modulate the stored charge. This way, the electrical driver only needs to charge the nano-scale gate of the transistor, with attojoule-scale energy dissipation. We implement this ‘photovoltaic modulator’ in a monolithic CMOS platform. This work demonstrates how close integration and co-design of electronics and photonics offers a path to optical switching with as few as 500 injected electrons and electrical energy consumption as low as 20 zJ/bit, achieved only by recovering the absorbed optical energy that is wasted in conventional modulation.

The authors demonstrate a "photovoltaic" ring modulator operating with record electrical energy efficiency by operating a silicon P-N diode in weak forward bias. The novelty of their work lies in the "recycling" of the photocurrent in the silicon diode to self-bias or charge a transistor, which in turn can switch the driving voltage of the diode with some electrical gain. The trade-off is a reduction in electrical bandwidth due to large output resistance of the transistor, which limits the device to operation in the MHz regime. A few comments and questions from me below: • What is the waveguide loss for the SiGe waveguide used in the modulator? • It may be helpful to point out that Ipc in Fig 2(c) and (d) has an opposite sign compared with I in Fig 2(b). • In Fig. 3, figures (d) through (g) are not very clear in terms of which line corresponds to which photocurrent. Also, better visual clarity between the dots and x's would be appreciated.
• In comparing rings with loaded Q = 6000 and 12000, the authors comment that an increase in extinction ratio is caused by a steeper slope in the ring transfer function. The change in Q should also affect the responsivity of the ring, which would in turn change the photocurrent. Is this also considered in the simulations?
• What Vgs bias is used for the experimental results in Fig. 4? Was the bias chosen such that the electric gain is the maximum, or for optimal modulation efficiency? • It seems that insertion loss in this paper is defined as the "1" level of the modulator. Is this consistent with all the results tabulated in the supplementary document? Sometimes, people will quote the insertion loss as the average power.
Overall, the paper is very well written and the supplementary information contains very thorough background review and supporting data. The underlying concept is very interesting, although applications may be ultimately limited by the relatively slower speeds, and the fact that the energy associated with the modulator may be dwarfed by other things such as laser power. Nonetheless, I believe this paper to be an excellent contribution to Nature Communications.
Reviewer #2 (Remarks to the Author): The manuscript by de Cea et al. describes Si modulators that use a photovoltaic (PV) effect to reduce the energy consumption of modulation. The proposal is interesting, but the manuscript should be revised before the it can be acceptable for publication. Below are my comments.
1. For the material to exhibit a PV effect at the wavelength of operation, it must be optically absorbing to some extent. Absorption and voltage-dependent absorption are parasitic for a phasemodulator and contributes to additional chirp. Would the authors please discuss more about this trade-off between loss and PV in the main manuscript? Should a device designer try to eliminate parasitic losses, in which case this proposed effect would not exist? How much absorption can be tolerated?
2.In Fig. 1, among the proposed applications, some are not reasonable for Si waveguides. Please revise or clarify these use cases. For brain (optogenetic) applications (b), the operation wavelength is often blue or green, so Si waveguides and modulators are too absorptive and are not used. For single photon applications (c), is having absorption in the modulator better than striving to reduce loss (see Q.1)? The MRI application in (d) might be fine, but it may require higher extinction ratios than demonstrated in this article.
3. On page 3, it is stated that parasitic absorption and free carrier generation is universal. Does this also occur in insulator materials like LiNbO3 or BaTiO3? To my knowledge, there is almost no parasitic absorption. The bandgap is also very large. The authors' claim may only be most applicable to semiconductors. Would the authors please clarify to substantiate their claim (please include references), or revise? 4. Surface passivation would reduce the photocurrent. It is not clear how general the results in this article can be adopted by others in the field who use different processes to fabricate the devices. 5. Despite the authors stating that the parasitic absorption of Si in the infrared could be used for the PV modulation, the device demonstrated has Ge to increase the absorption near 1300nm wavelength! The introduction and motivation for this article should be revised to not over-extend the proposal, claims, and the results. 6. Would the authors please clarify the device insertion loss quoted in Fig. 4? The quoted IL is 2.4dB-3.34dB. Does this also include material loss? 7. The achieved ER is < 0.6 dB, which is very low and is not usually acceptable for optical communications. Page 9 states a 1dB ER is sufficient for error free communication. What is reference for this claim? In optical communication, the bit error rate (BER) is the important metric. A minimum BER of 10^-3 is needed for forward error correction (FEC). What is the achieved BER here? The use of FEC will significantly increase overall power consumption, thus eliminating any benefits of this PV modulation scheme.
8. There is indeed a great deal of interest in minimizing the energy consumption of optical transmitters to fJ/bit levels or less. But this energy consumption specification includes the laser source and the driver electronics, which are often the dominant source of energy consumption and not the modulator itself. As the power dissipation of a laser is static, modulating at high bit rates has the advantage of reducing the energy per bit (= power / bit rate). In this article, the power consumption of the laser has been neglected. It would be instructive to quote the total energy per bit (including the laser source). 9. While this article has an interesting proposal, it seems to have missed the jugular when one focuses on energy consumption. I would recommend reframing the motivation of the article and presenting the work in a more balanced way that describes both the advantages and limitations of the work.
The authors demonstrate a "photovoltaic" ring modulator operating with record electrical energy efficiency by operating a silicon P-N diode in weak forward bias. The novelty of their work lies in the "recycling" of the photocurrent in the silicon diode to self-bias or charge a transistor, which in turn can switch the driving voltage of the diode with some electrical gain. The trade-off is a reduction in electrical bandwidth due to large output resistance of the transistor, which limits the device to operation in the MHz regime. A few comments and questions from me below: • What is the waveguide loss for the SiGe waveguide used in the modulator?

Response:
The total waveguide loss can be extracted from the quality factor and extinction ratio of the ring resonator. For our devices, the ring waveguide loss is about 6.34 cm -1 (27.5 dB/cm).
From preliminary measurements characterizing the device, 0.6 cm -1 (2.6 dB/cm) of the total loss correspond to photocurrent-generating absorption. The remaining losses (5.74 cm -1 = 24.9 dB/cm) can be mainly attributed to free carrier absorption.
This information has been added in Supplementary Section S2.
• It may be helpful to point out that Ipc in Fig 2(c) and (d) has an opposite sign compared with I in Fig 2(b). Fig 2(b) showing the I and V definitions has been added to clarify the current signs. • In comparing rings with loaded Q = 6000 and 12000, the authors comment that an increase in extinction ratio is caused by a steeper slope in the ring transfer function. The change in Q should also affect the responsivity of the ring, which would in turn change the photocurrent. Is this also considered in the simulations?

Response: An inset in
Response: It is true that, in general, the Q factor of the ring will affect the responsivity of the device. However, our results are given as a function of generated photocurrent. Therefore, the conversion from input optical power to generated photocurrent (through the responsivity R) is already incorporated here.
• What Vgs bias is used for the experimental results in Fig. 4? Was the bias chosen such that the electric gain is the maximum, or for optimal modulation efficiency?
Response: For operation of the modulator, we ultimately care about maximizing the amplitude of the generated modulation signal. Thus, Vgs was chosen to maximize modulation efficiency. A sentence has been added pointing to this fact (line 237 in the main manuscript).
• It seems that insertion loss in this paper is defined as the "1" level of the modulator. Is this consistent with all the results tabulated in the supplementary document? Sometimes, people will quote the insertion loss as the average power.

Response:
To the best of our knowledge, all quoted insertion losses in tables 1 and 2 are defined as the "1" level of the modulator.
Overall, the paper is very well written and the supplementary information contains very thorough background review and supporting data. The underlying concept is very interesting, although applications may be ultimately limited by the relatively slower speeds, and the fact that the energy associated with the modulator may be dwarfed by other things such as laser power. Nonetheless, I believe this paper to be an excellent contribution to Nature Communications.

Reviewer #2:
The manuscript by de Cea et al. describes Si modulators that use a photovoltaic (PV) effect to reduce the energy consumption of modulation. The proposal is interesting, but the manuscript should be revised before the it can be acceptable for publication. Below are my comments.
1. For the material to exhibit a PV effect at the wavelength of operation, it must be optically absorbing to some extent. Absorption and voltage-dependent absorption are parasitic for a phase-modulator and contributes to additional chirp. Would the authors please discuss more about this trade-off between loss and PV in the main manuscript? Should a device designer try to eliminate parasitic losses, in which case this proposed effect would not exist? How much absorption can be tolerated?

Response:
The reviewer is absolutely correct in pointing out that some extent of optical loss is required for our scheme to work.
As discussed in Supplementary Section S2, it is not possible to entirely eliminate these parasitic effects in semiconductor devices. This has been recognized by several researchers and reported in a variety of publications (indicated in Supplementary Section S2), which observe these effects in fabrication processes specifically tailored for photonics. The main point of our work is recognizing the presence of this photocurrent and, instead of just treating it as a parasitic (i.e, undesired) effect, use it to our advantage. It is also worth pointing out that there have been other works that take advantage of parasitic optical absorption in semiconductor optical waveguides. For instance, the photocurrent has been used to While for a conventional modulator configuration it is desirable to eliminate absorption, our work demonstrates how, in specific situations, having optical absorption (and even adding it intentionally) can enable devices and systems with better performance than state-of-the-art solutions. Examples include the applications proposed in the manuscript (Fig. 1). We believe that, if a low data rate link with low power consumption is desired, using our photovoltaic modulator is a promising approach and could be advantageous when compared to the use of electrical links or a scheme with a conventional modulatorwhere the parasitic photocurrent could be smaller but wasted and not used to effectively bootstrap an electrical amplifier as it is here.
2.In Fig. 1, among the proposed applications, some are not reasonable for Si waveguides. Please revise or clarify these use cases. For brain (optogenetic) applications (b), the operation wavelength is often blue or green, so Si waveguides and modulators are too absorptive and are not used. For single photon applications (c), is having absorption in the modulator better than striving to reduce loss (see Q.1)? The MRI application in (d) might be fine, but it may require higher extinction ratios than demonstrated in this article.
Response: In all of the examples provided, the photovoltaic modulator is used to replace the electrical readout link between the sensor and a remote computational processor. We will address each of the applications separately: 1. Electrophysiological sensing and neural interfaces: The neural interfaces we are exploring include a wide range of devices used to monitor neural/electrophysiological signals, and the PV modulator is used to establish an optical link between the sensing site and the processor. This does not include optogenetic techniques, which are used to stimulate genetically modified neurons through light.
Neural interfaces with both wired and wireless links have been reported in the literature to establish communication between the sensor and the processor. To date, such demonstrations have a relatively high power dissipation the limits the number of parallel sensing sites that can be supported without affecting the brain or any other tissue under study: Our demonstrated solution with a total (electrical + optical) power of 7.5 µW (and an energy consumption of 7.5 pJ/bit at 1 MHz) could enable the readout of thousands of parallel sensors while still complying with the power dissipation limits set by the brain tissue.
2. Single photon applications: the purpose of the photovoltaic modulator in this application is not to transduce or transport the single photons (this is done by the superconducting nanowire single photon detector or SNSPD), but to substitute the electrical link that is usually employed to read out the signal generated by the SNSPD by an optical link. In this scenario, the PV modulator converts the electrical signal generated by the SNSPD to the optical domain. The reviewer is kindly pointed to [de Cea et al., Sci. Rep. 10, 9470 (2020)] for a more detailed discussion of the benefits of the use of optical readout in cryogenic applications.
As discussed in [de Cea et al., Sci. Rep. 10, 9470 (2020)], a cryogenic optical readout scheme will be beneficial as long as its power dissipation is lower than the electrical link alternative. The PV modulator could represent a significant improvement compared to previous demonstrations of optical readout because it eliminates DC electrical power dissipation. As long as this reduction in power consumption is larger than the added absorption loss, it will be beneficial to use our PV modulator approach.
We have modified the first three paragraphs of the introduction to clarify our envisioned applications and clarify that the purpose of the PV modulator is to enable an optical link between the sensing site and the main processor.
3. On page 3, it is stated that parasitic absorption and free carrier generation is universal. Does this also occur in insulator materials like LiNbO3 or BaTiO3? To my knowledge, there is almost no parasitic absorption. The bandgap is also very large. The authors' claim may only be most applicable to semiconductors. Would the authors please clarify to substantiate their claim (please include references), or revise?

Response:
The sentence in page 3 (line 74) has been modified to make it clear that we are referring to semiconductor-based optical modulators only. As the reviewer correctly points out, free carrier generation in insulator materials should be negligible.

Surface passivation would reduce the photocurrent. It is not clear how general the results in this article
can be adopted by others in the field who use different processes to fabricate the devices.
Response: It is true that the quality of surface passivation will affect the strength of surface state absorption (SSA) and the generated photocurrent. Nevertheless, we believe that SSA is present in any modern fabrication processes, even for those with the highest passivation quality. Two facts support this claim: 2. The silicon surface passivation in our CMOS process (45RF SOI from GlobalFoundries) is highly optimized. Since ours is a microelectronics process that is designed to yield billions of highperformance transistors, high quality surface passivation is needed to ensure the best, most consistent transistor performance. As we report in the Supplemental Material the measured photocurrent in our all-silicon modulators is consistent with the expected photoresponse in the most advanced photonics processes.

5.
Despite the authors stating that the parasitic absorption of Si in the infrared could be used for the PV modulation, the device demonstrated has Ge to increase the absorption near 1300nm wavelength! The introduction and motivation for this article should be revised to not over-extend the proposal, claims, and the results.

Response:
The experimental results shown in the main part of the manuscript correspond to a device with embedded SiGe (with a Germanium concentration between 20% and 30 %, not pure Ge). Nevertheless, our approach is extendable to any semiconductor-based modulator that shows parasitic free carrier absorption, which includes Si and InP. We show experimental data demonstrating the photovoltaic modulation approach working for a Si-only modulator in Supplementary Section S7 (Supplementary Section S6 in the original manuscript), showing a responsivity of 0.85 mA/W at a 1550 nm wavelength. While this responsivity is lower than the 34 mA/W measured in the SiGe modulator, the generated photocurrents are enough to obtain a useful shift in the resonance wavelength of the device (as shown in Supplementary Figure S7).
The analysis and simulations carried out in the "Operating principle and device modeling" section of the main manuscript are general and apply to any modulator that has a parasitic photocurrent in response of an input optical power, and are not particular to the Si resonator with or without embedded SiGe.
6. Would the authors please clarify the device insertion loss quoted in Fig. 4? The quoted IL is 2.4dB-3.34dB. Does this also include material loss?

Response:
We use the standard definition of insertion loss in communications applications, which defines the insertion loss as the ratio between the output power of the '1' bit and the output power of the device when it is out of resonance. This definition does not account for the losses of coupling in and out of the chip, but accounts for any material loss in the ring resonator.
The total waveguide loss can be extracted from the quality factor and extinction ratio of the ring resonator. For our devices, the ring waveguide loss is about 6.34 cm -1 (27.5 dB/cm).
From preliminary measurements characterizing the device, 0.6 cm -1 (2.6 dB/cm) of the total loss correspond to photocurrent-generating absorption. The remaining losses (5.74 cm -1 = 24.9 dB/cm) can be attributed to free carrier absorption.
This information has been added in Supplementary Section S2.
7. The achieved ER is < 0.6 dB, which is very low and is not usually acceptable for optical communications. Page 9 states a 1dB ER is sufficient for error free communication. What is reference for this claim? In optical communication, the bit error rate (BER) is the important metric. A minimum BER of 10^-3 is needed for forward error correction (FEC). What is the achieved BER here? The use of FEC will significantly increase overall power consumption, thus eliminating any benefits of this PV modulation scheme.

Response:
In typical high bandwidth optical communications applications the reviewer's comment is correct: 0.6 dB ER is usually not acceptable (in fact, most standards require an ER > 2 dB).
Nevertheless, operation at the bandwidths that our target applications require significantly change the performance tradeoffs. The low receiver bandwidth substantially decreases the equivalent noise power of the receiver chain detecting the modulated signal generated by our PV modulator. This allows for the achievement of close to error-free transmission without the need of FEC at the ER levels we demonstrate in our work.
We can estimate the required ER based assuming reference systems that are either shot-noise limited (Scenario 1) or receiver noise limited (Scenario 2).

Scenario 1: Shot-noise limited system
One hypothetical scenario is to consider a shot-noise limited system, i.e, a system where the dominant source of noise is that associated with the power of the optical signal itself. In this case, the noise variance of the current at the detector is given by: Where q is the electron charge, B is the signal bandwidth and I is the photocurrent generated by the input optical signal, = * . In the latter, R is the responsivity of the photodetector.
If we assume that, at the receiver site, we have a '1' current level 1 and a '0' current level 0 , we can write: Above, 1 is the optical power at the receiver site corresponding to the '1' bit. Given the expression above, it is easy to evaluate the achievable SNR (and from it the BER) for different ER, received powers and bandwidths.
The figure below shows the achievable BER as a function of ER for a 100 MHz bandwidth signal, assuming a photodetector with responsivity R = 0.5 A/W, for different received powers: Clearly, BER in the order of 0.5 dB are enough to achieve error-free communication (without the need for FEC) for received powers above 1 μW. For reference, our experimental results with a 7.5 μW input optical power and ≈ 3 dB IL correspond to 1 = 3.75 .

Scenario 2: Receiver noise limited system
In this case, which is the most likely in practical receiver implementations, the electrical noise in the receiver dominates the total noise in the system.
In this case, the BER is given by: Where is the input equivalent current noise of the receiver circuit. We therefore need to obtain the input equivalent noise of the receiver to evaluate the SNR and the BER from it. Of course, that will depend on the implementation of the receiver, and can vary depending on the specifics of the receiver.
We can look for commercially available receivers to obtain some representative values of so that we can evaluate BER.
1. Newport 125 MHz optical receivers (https://www.newport.com/f/125-mhz-photoreceivers): According to the specifications, these have a noise equivalent power (NEP) = 2.5 pW/sqrt(Hz). If, as we did in the case of the shot-noise limited system, we plot the BER as a function of ER for different received '1' powers, we get the following: Clearly, for received powers above 2 μW, error-free communication can be established with ER < 1 dB. IN the case of our experimental demonstrations, the received '1' power was 3.75 μW.
Note: The NEP of the receiver analyzed here is representative of a great variety of commercially available optical receivers. For example, a 250 MHz receiver from Thorlabs (https://www.thorlabs.us/thorproduct.cfm?partnumber=FPD510-FS-NIR) has a 3.2 pW/sqrt(Hz).

Texas Instruments 125 MHz Transimpedance Amplifier (TIA)
(https://www.ti.com/lit/ds/symlink/opa857.pdf?ts=1604249163803&ref_url=https%253A%252F %252Fwww.ti.com%252Fproduct%252FOPA857) The datasheet specifies a total rms noise of 12 nA integrated over a 135 MHz bandwidth. Assuming a photodetector with responsivity R = 0.5 A/W, the BER as a function of ER is the following: Once again, for received powers above 2 μW, error-free communication can be established with ER < 1 dB.
Note 1: In this case, since we are only considering the TIA noise, we are assuming that the dark current of the photodetector is negligible.
Note 2: The noise power of the TIA analyzed here is representative of a great variety of commercially available TIAs. For example, Analog Devices has a TIA with 220 MHz bandwidth and 4.8 pA/sqrt(Hz) input current noise (https://www.analog.com/en/products/ltc6561.html).
In conclusion, while it is true that in conventional high-speed optical communications an ER of 1 dB is not enough for low error-rate communication, operating at low bandwidths enables error-free communications with ER on the order of 0.5 dB for received powers above ≈ 2 μW, which we have demonstrated experimentally. A supplementary section (Supplementary Section S6) has been added reproducing (in a shorter format) the discussion here.
8. There is indeed a great deal of interest in minimizing the energy consumption of optical transmitters to fJ/bit levels or less. But this energy consumption specification includes the laser source and the driver electronics, which are often the dominant source of energy consumption and not the modulator itself. As the power dissipation of a laser is static, modulating at high bit rates has the advantage of reducing the energy per bit (= power / bit rate). In this article, the power consumption of the laser has been neglected. It would be instructive to quote the total energy per bit (including the laser source).

Response:
The preliminary list of target applications ( Fig. 1) require data rates in the Mbps range or below, and therefore there is no immediate benefit to having a modulator that could operate at much faster speeds: it is simply not needed.
In this case, it is more beneficial to have an optical modulator capable of operating with the minimal overall power dissipation (including laser power). This is what we discuss and quote in the last paragraph of the discussion section: the total power dissipation of 8 μW includes the laser power. Additionally, we have added a sentence at the end of the first paragraph of the discussion section quoting the optical energy dissipation per bit.
As discussed in response to point 2, even when accounting for the required optical power, the total power dissipation of our solution is lower than that of state-of-the-art solutions.
9. While this article has an interesting proposal, it seems to have missed the jugular when one focuses on energy consumption. I would recommend reframing the motivation of the article and presenting the work in a more balanced way that describes both the advantages and limitations of the work.
Response: As we have discussed in detail in our responses above, we believe that our photovoltaic modulator offers significant advantages for low power, low data rate applications compared to state-ofthe-art solutions: -When compared to electrical solutions, our scheme could provide significant gains in the total power dissipated for the readout of massively parallel data streams. In particular: o For neural interface applications, the 7.5 μW total (electrical + optical) power and 7.5 pJ/bit energy dissipation demonstrated here represents a significant improvement over reported state-of-the-art electrical interfaces. Thus, our solution employing an optical link with a PV modulator could enable the readout of thousands of parallel sensors while still complying with the power dissipation limits set by the brain or any other biological tissue. o For MRI applications, power dissipation is not so critical. Instead, the ability to read out signals with improved robustness to electromagnetic interference is highly advantageous as it reduces the need for bulky and expensive electrical shielding.
The table included below summarizes the bandwidth and power dissipation of state-of-the-art approaches to the target applications described in our work. As can be seen, our proposed solution can provide significant gains in power dissipation per channel, which ultimately translates into an increase in the number of data streams that can be read out in parallel.