Abstract
Non-volatile multilevel optical memory is an urgent needed artificial component in neuromorphic computing. In this paper, based on ferroelectric based electrostatic doping (Fe-ED) and optical readout due to plasma dispersion effect, we propose an electrically programmable, multi-level non-volatile photonics memory cell, which can be fabricated by standard complementary-metal-oxide-semiconductor (CMOS) compatible processes. Hf0.5Zr0.5O2 (HZO) film is chosen as the ferroelectric ED layer and combines with polysilicon layers for an enhanced amplitude modulation between the carrier accumulation and the confined optical field. Insertion loss below 0.4 dB in erasing state and the maximum recording depth of 9.8 dB are obtained, meanwhile maintaining an extremely low dynamic energy consumption as 1.0–8.4 pJ/level. Those features make this memory a promising candidate for artificial optical synapse in neuromorphic photonics and parallel computing.
© 2022 Optica Publishing Group under the terms of the Optica Open Access Publishing Agreement
Non-volatile memories have been recognized as a key technology to build human brain inspired massive parallelism, promoting the development of in-memory computing [1]. Such memories are essential to break the performance limitations in conventional computation and storage separated von Neumann architectures, showing great potential in data processing and energy efficiency [2]. Furthermore, multilevel memory units are the fundamental artificial components for mimicking the behavior of synapses in neuromorphic computing [3–5]. Among those memories, electron-optical memory (EOM) promises a combination of both photonic and electronic circuits integrated into the same chip and addresses the gap of electro-optic data storage [6]. In fact, optical interconnection breaks the data bandwidth limitations faced by copper lines along with its potential integration with electronic circuits in the same chip and the compatibility with the mature and cost-effective manufacturing infrastructure of the microelectronic industry, in which silicon photonics have been a leading technology solution in this field. With the rapidly growing demand of computational capability from artificial intelligence (AI), 5G communications and internet of things (IOT), both silicon and silicon nitride (SiN) based EOM has become an urgently needed device in optoelectronics hybrid system for retaining electrical data with optical readout mechanism [7].
Many efforts have been made to achieve integrated optical memories. Phase change materials (PCMs), such as Ge2Sb2Te5 based memories, are integrated on both silicon and SiN waveguide for multiple synaptic strengths writing and erasing [8–10]. Since the multilevel crystallization and amorphization of PCMs are electrical or optical heating processes, the power consumption and thermal crosstalk to adjacent memories cannot be neglected when facing high-density integration [11]. By contrast, charge induced optical absorption memories with voltage pulse controlling, such as charge trapping [12] and floating gate [13,14] approaches, show extremely low power consumption (10–100 fJ/bit). Among all these energy-efficient devices, ferroelectric thin film is recommended to be a promising functional candidate, developing electric logic and non-volatile multilevel memory devices such as ferroelectric field effect transistor (FeFET) [15], ferroelectric tunnel junction (FTJ) [16]. Moreover, the ferroelectric based electrostatic doping (Fe-ED) in nanosheet field effect transistor (NSFET) has been reported [17]. In addition, ferroelectric devices exhibit endurance for more than 1011 switching cycles [18], which ensures a robust reliability for system operation. However, the feature of ferroelectric based multilevel state data storage is still missing for electric controlled EOM, which may be due to the week interaction between electric signal and optical signal.
In this paper, a Fe-ED optical readout memory is initially proposed. Thin ferroelectric film Hf0.5Zr0.5O2 (HZO) has been chosen as the ferroelectric doping layer. And a tapered polysilicon/HZO/polysilicon (semiconductor/ferroelectric/semiconductor, SFS) stack is used for coupling the optical mode and enhancing light-matter interactions. The memory performance including multilevel electronical information writing/erasing and broadband optical reading will be evaluated at telecommunication wavelengths. In addition, within the CMOS electrical compatible driving voltage, the dynamic energy consumption for switching is estimated. This ferroelectric based electron-optical memory exhibits great potential in energy efficient neuromorphic photonics applications.
1. Concept and methodology
The concept of an EOM is shown schematically in Fig. 1(a). A SiN waveguide integrated with the tapered SFS stack waveguide acts as a memory unit. As the functional core of the EOM, the tapered three-tier stack offers vertical coupling of the optical mode adiabatically between the SFS waveguide and the SiN waveguide, as well as an enhanced electro-optical absorption for the coupled light. The fabrication processes and corresponding steps are as follows. The bottom polysilicon is firstly deposited on the SiO2-capped SiN waveguide substrate by low-pressure chemical vapor deposition (LPCVD). Then a thin HZO film is deposited by atomic layer deposition (ALD). Finally, the wafer is covered with polysilicon by the same process as the bottom polysilicon. Before the further patterning, an anneal is needed for the film crystallization, which is helpful to improve the ferroelectric properties of HZO and reduce the optical absorption of polysilicon.
As shown in Fig. 1(b), the cross-sectional view of this memory is depicted in detail. A ferroelectric layer HZO with a thickness of 10 nm is sandwiched with two 100 nm thick polysilicon layers. This tapered stack is separated by an optimized 200 nm gap [19] from the SiN waveguide of 700 nm width and 400 nm thickness. On the top of the stack, a gate electrically contacts the Ni/Au top electrodes for voltage control. To prevent the overlap with the optical mode, all the electrodes including the grounds are arranged away from the waveguide. In order to obtain significant amplitude modulation depth, the SFS waveguide should be designed to couple the energy as much as possible. Figure 1(c) further shows the top view of SFS waveguide. The tip-width and end-width of the taper sections are 100 nm and 500 nm respectively, and the transforming lengths of the front and back taper are 30 µm which is sufficiently adiabatic to transfer the power between the SiN and SFS waveguide. In the straight waveguide section of 200 µm, the length determines the modulation depth. These geometric parameters are optimized for evaluating the modulation depth of the memory.
Once the information is electrically written into the memory, it can be read optically by measuring the inserting loss between the input and output port of the underlying SiN waveguide. The HZO film is recommended to be a promising ferroelectric candidate material for its fully integration with silicon-based integrated circuit technology [20]. In this memory unit, the residual polarization charges induced by voltage pulse cause nonvolatile ED, which can be used to record information in optical form indirectly based on the electro-optical absorption principle [21]. When the confined optical mode propagates inside the central SiN waveguide, the evanescent field in polysilicon film will response according to the electrical signal due to plasma-dispersion effect [22], by changing the absorption coefficient along with the changing electrical signal. Specifically, the covered polysilicon has a larger refractive index than that of SiN waveguide, generating an enhanced light-matter interaction between optical transmission filed and ED absorption region.
The multilevel writing principle of SFS waveguide is performed in Fig. 2. As shown, random residual polarization charges can be produced by applying appropriate voltage pulses. To guarantee the repeatability and accuracy, the writing process should be carried out after the reset process [5], which can also be applied to the charge erasing process. Consequently, it is acquirable for a full range of residual charge from 0 to Pr with a dynamic range of ΔPr. As expected, the double sandwiched polysilicon layers play an important role in both electrical- and optical-domain [23,24]. On the one hand, the deposited polysilicon layers act as conventional channel of the field effect transistor. As the energy band diagrams depicted in Fig. 3, the prepared Boron-doped polysilicon films are chosen for investigation. Due to the residual polarization charge, the p-doped polysilicon layers will form inversion layer and accumulation layer at the two sides of interfaces with HZO. Figure 3(a) shows that electrostatics induced inversion electrons concentrates at top polysilicon side, while accumulation holes at bottom polysilicon side, after a reset pulse (-Vs) applied. While, as shown in Fig. 3(b), the writing pluses generate accumulation holes at top polysilicon side and inversion electrons at the opposite side. All the non-volatile ED charges could be erased by applying a pulse slightly larger than the coercive voltage Vc. Figure 3(c) shows the band diagrams of erasing state. Due to the same body doping type and level, the band keep flat without remarkable carrier concentration region occurring. On the other hand, the deposited polysilicon is not only a well-performed transistor gate layer in dynamic random-access memory (DRAM) [25], but also a high-index and low-loss optical waveguide materials. In standard complementary metal-oxide-semiconductor (CMOS) process, the propagation losses of polysilicon waveguide have been improved to 6.2 dB/cm at 1550 nm [26].
Furthermore, the optical behavior of SFS waveguide can be associated with the electrical properties discussed above. According to plasma-dispersion effect, the free carrier-induced optical refractive index changes (Δn) and extinction coefficient changes (Δk) of polysilicon due to the spatial redistribution of electrons (ΔN) and holes (ΔP) changes are used, which could be expressed by the Drude-plasma model as [27],
2. Modeling and optimization
In order to evaluate the performance of the proposed EOM, multiple physics modules including electrostatic, semiconductor and wave-optics module are involved. All the 2D and 3D simulations are carried using the finite element method (FEM) in COMSOL Multiphysics platform [28].
Firstly, the hysteresis and nonlinear behavior of HZO film is modeled by domain wall model, which has been verified well with typical bulk ferroelectric materials [29]. While for the ultra-thin high κ HZO nano-films, the charge-voltage relationship should be corrected by adding the capacitive charges stored in the ferroelectric capacitor [30],
where QFe is the total charge, PFe is the polarization charge, ɛFe and tFe are the permittivity and thickness of ferroelectric film, and Va is the voltage applied on the electrode. Figure 4(a) shows the cross-sectional TEM image of TaN/HZO/TaN capacitor. The thickness of HZO is 10 nm. As shown in Fig. 4(b), the measured hysteresis loop, performed by an aixACCT TF Analyzer 2000 measurement system, can be matched well by the simulated P-E curves. The ferroelectric parameters used for modeling are listed in Table 1. After that, the polarizations can be further predicted by using Langevin anhysteretic expression [29]. Figure 4(c) shows the predictions from 4 V to 2 V with a step of 0.5 V. The Pr values at different voltages could be extracted directly in the P-E curves. Depending on this model, the carrier concentrations are simulated to convert the multilevel Pr into complex refractive index changes Δn and Δk as a function of applied voltage pulses Va.According to the theory of ED, the screening length of accumulation and inversion carriers is typically short (<10 nm) [23], which brings much difficulty to photonic devices for phase or amplitude modulations. To address this issue, the tapered Fe-ED SFS waveguide is intentionally introduced. Due to the adiabatic coupling, the optical transmission mode adds gradually into the front tapered section and propagates along the straight SFS waveguide, and then drops into the SiN waveguide in the back tapered section. The redistributed optical field provides an opportunity to interact fully with ED absorption region.
Based on the analysis above, it is necessary for the adiabatic coupling efficiency of the tapered SFS waveguide to be optimized. The main semiconductor and optical parameters used are listed in the second and third part of Table 1. Figure 5(a) shows the confinement factors [31] of the fundamental TE mode in the SiN waveguide (red curve) and in the SFS waveguide (blue curve) as a function of the width of the tapered SFS waveguide. As expected, the taper adiabatically transforms the mode from SiN waveguide into the SFS waveguide. Figure 5(b) displays the optical transmittance as a function of the taper length. The results were simulated using a 3D beam envelopes (BME) method. As we see, a taper with a length larger than 20 µm is sufficient to transfer the optical energy into the bottom SiN waveguide without significant scattering losses (<0.1 dB). From Fig. 5(a) and (b), a taper with width linearly changing from 100 nm to 500 nm and the transforming length of 30 µm is chosen. Furthermore, the vertical-coupled mechanism provides an additional access to enhance the enhanced light-matter interaction between optical transmission filed and ED absorption region to the utmost extent, which is depicted in Fig. 5(c) and (d). As shown in Fig. 5(d), a fundamental electric-field profile of TE mode is exhibited. Simultaneously, under applied voltage pulse of 2 V, the lossy region induced by the free carriers, with maximum extinction coefficient k of 0.11 for electrons doping and 0.07 for holes doping at the same carrier concentration 7.41×1020 cm−3, locates nearly at the maximum intensity of the electric-field, which represents an enhancement of the interaction between them.
4. Results and discussion
The exploration of multilevel optical memories is the basis for mimicking the synaptic weights in an optical neural network. In this section, based on the optimized design, the performance of the proposed EOM will be comprehensively studied. Figure 6 shows the transient response of HZO capacitor in accordance with the random driving signal with a pulse width of 2 µs. Following the operation principle which has been discussed above, the response curves show that Pr of 0.172 C/m2 is maintained after a positive 2 V pulse. Along with the new writing signals, the written Pr increases to 0.274 C/m2 for 4 V pulse and then decreases to 0.249 C/m2 for 3 V pulse. And it is also found that the Pr could be erased by applying 1.4 V pulse. Those non-volatile random writing behaviors are further applied in SFS waveguide.
By solving the Poisson and drift-diffusion equations, the carrier concentration distributions across the ED region are numerically calculated. In the Fig. 7, the electrons, and holes distribution in reset, writing and erasing states are performed successively. After a negative −4 V reset voltage pulse is applied, the inversion electrons with maximum concentration of 1.27×1021 cm−3 appear at the top polysilicon layer, and accumulation holes with the equivalent maximum concentration at the bottom polysilicon layer. Compared with the reset state, the writing states show a reversal ED type due to the reversal polarization direction. As the writing pulses decrease from 4 V to 1.5 V, the maximum carrier concentration, both for electrons and holes, is dramatically drops from 1.28×1021 cm−3 to 1.77×1020 cm−3. While for the erasing state, the residual charges are mostly removed from the HZO film. Although a few charges still exist, its impact on the insertion loss (IL) can be ignored. With the help of the Eq. (1) and (2), the spatial distribution of complex refractive indexes can be refreshed, which is the key mechanism to combine the optical properties with electrical response. For an electrically controlled SiN-based EOM with HZO on top of it, the optical propagation losses within the whole hybrid waveguide response instantaneously with the recorded electrical data. Figure 8(a) shows the electro-absorption losses of TE modes. When total 11 levels of electrical pulses are applied, the corresponding recorded optical memory states calculated by the 3D BME method, starts from the ground state (erasing state) of 0.4 dB at 1.4 V to the highest level of 9.8 dB at 4 V. Meanwhile, the memory states are calculated by the fast 2D eigenmode expansion (EME) method [33], which is also presented and agrees well with the BME results. Furthermore, as shown in Fig. 9, attributed to the tapered adiabatic coupling and transmission, our proposed optical module exhibits a broadband optical response of 300 nm for TE mode, providing a more practical memory technology in telecommunication wavelengths.
In addition, to evaluate such ultralow energy consumption for such EOM device based HZO- stack more specifically, the dynamic energy consumption for each state can be calculated using the following equation,
5. Conclusion
To summarize, a novel non-volatile and multilevel ferroelectric based electron-optical memory is proposed and evaluated. Based on plasma dispersion effect, the multi-level memory could be randomly written by voltage pulse stepwise and read in optical form through the variation of free carrier concentration. Due to the natural low loss and elaborate configuration of polysilicon and HZO, the hybrid memory with a tapered HZO-based polysilicon waveguide shows an insertion loss below 0.4 dB in erasing state and the maximum recording depth of 9.8 dB for TE mode, meanwhile maintaining an extremely low energy consumption as 1.0–8.4 pJ/level and broadband optical reading at telecommunication wavelengths. Low loss SiN-based electro-optical memory combined with ferroelectric layer shows potential in overcoming the bottleneck of von Neumann, which will open novel routes to the development of new high-density, low-power-consumption non-volatile memory architectures in neuromorphic photonics and parallel computing.
Funding
National Key Research and Development Program of China (2018YFB2200500); National Natural Science Foundation of China (62004145, 62025402, 62090033, 91964202, 92064003, 61874081, 62004149); Key Research Project of Zhejiang Lab (2021MD0AC01).
Disclosures
The authors declare no conflicts of interest.
Data availability
Data underlying the results presented in this paper are not publicly available at this time but may be obtained from the authors upon reasonable request.
References
1. G. W. Burr, R. M. Shelby, A. Sebastian, S. Kim, S. Kim, S. Sidler, K. Virwani, M. Ishii, P. Narayanan, A. Fumarola, L. L. Sanches, I. Boybat, M. Le Gallo, K. Moon, J. Woo, H. Hwang, and Y. Leblebici, “Neuromorphic computing using non-volatile memory,” Adv. Phys-X 2(1), 89–124 (2017). [CrossRef]
2. N. Verma, H. Jia, H. Valavi, Y. Tang, M. Ozatay, L.-Y. Chen, B. Zhang, and P. Deaville, “In-Memory Computing: Advances and Prospects,” IEEE Solid-State Circ. Mag. 11(3), 43–55 (2019). [CrossRef]
3. Y. Park, M.-K. Kim, and J.-S. Lee, “Emerging memory devices for artificial synapses,” J. Mater. Chem. C 8(27), 9163–9183 (2020). [CrossRef]
4. Z. Cheng, C. Ríos, W. H. P. Pernice, C. D. Wright, and H. Bhaskaran, “On-chip photonic synapse,” Sci. Adv. 3(9), e1700160 (2017). [CrossRef]
5. S. Oh, H. Hwang, and I. K. Yoo, “Ferroelectric materials for neuromorphic computing,” APL Mater. 7(9), 091109 (2019). [CrossRef]
6. B. J. Shastri, A. N. Tait, T. Ferreira de Lima, W. H. P. Pernice, H. Bhaskaran, C. D. Wright, and P. R. Prucnal, “Photonics for artificial intelligence and neuromorphic computing,” Nat. Photonics 15(2), 102–114 (2021). [CrossRef]
7. Y. Hao, S. Xiang, G. Han, J. Zhang, X. Ma, Z. Zhu, X. Guo, Y. Zhang, Y. Han, Z. Song, Y. Liu, L. Yang, H. Zhou, J. Shi, W. Zhang, M. Xu, W. Zhao, B. Pan, Y. Huang, Q. Liu, Y. Cai, J. Zhu, X. Ou, T. You, H. Wu, B. Gao, Z. Zhang, G. Guo, Y. Chen, Y. Liu, X. Chen, C. Xue, X. Wang, L. Zhao, X. Zou, L. Yan, and M. Li, “Recent progress of integrated circuits and optoelectronic chips,” Sci. China Inf. Sci. 64(10), 201401 (2021). [CrossRef]
8. C. Ríos, M. Stegmaier, P. Hosseini, D. Wang, T. Scherer, C. D. Wright, H. Bhaskaran, and W. H. P. Pernice, “Integrated all-photonic non-volatile multi-level memory,” Nat. Photonics 9(11), 725–732 (2015). [CrossRef]
9. J. Feldmann, M. Stegmaier, N. Gruhler, C. Rios, H. Bhaskaran, C. D. Wright, and W. H. P. Pernice, “Calculating with light using a chip-scale all-optical abacus,” Nat. Commun. 8(1), 1256 (2017). [CrossRef]
10. C. Rios, P. Hosseini, C. D. Wright, H. Bhaskaran, and W. H. Pernice, “On-chip photonic memory elements employing phase-change materials,” Adv. Mater. 26(9), 1372–1377 (2014). [CrossRef]
11. C. Peng and M. Mansuripur, “Thermal cross-track cross talk in phase-change optical disk data storage,” J. Appl. Phys. 88(3), 1214–1220 (2000). [CrossRef]
12. I. Olivares, J. Parra, and P. Sanchis, “Non-Volatile Photonic Memory Based on a SAHAS Configuration,” IEEE Photonics J. 13(2), 1–8 (2021). [CrossRef]
13. J.-F. Song, X.-S. Luo, A. E.-J. Lim, C. Li, Q. Fang, T.-Y. Liow, L.-X. Jia, X.-G. Tu, Y. Huang, H.-F. Zhou, and G.-Q. Lo, “Integrated photonics with programmable non-volatile memory,” Sci Rep 6(1), 22616 (2016). [CrossRef]
14. J. Parra, I. Olivares, A. Brimont, and P. Sanchis, “Non-volatile epsilon-near-zero readout memory,” Opt. Lett. 44(16), 3932–3935 (2019). [CrossRef]
15. J. Hoffman, X. Pan, J. W. Reiner, F. J. Walker, J. P. Han, C. H. Ahn, and T. P. Ma, “Ferroelectric Field Effect Transistors for Memory Applications,” Adv. Mater. 22(26-27), 2957–2961 (2010). [CrossRef]
16. V. Garcia and M. Bibes, “Ferroelectric tunnel junctions for information storage and processing,” Nat Commun 5(1), 4289 (2014). [CrossRef]
17. S. Zheng, J. Zhou, H. Agarwal, J. Tang, H. Zhang, N. Liu, Y. Liu, G. Han, and Y. Hao, “Proposal of Ferroelectric Based Electrostatic Doping for Nanoscale Devices,” IEEE Electron Device Lett. 42(4), 605–608 (2021). [CrossRef]
18. R. Cao, Q. Liu, M. Liu, B. Song, D. Shang, Y. Yang, Q. Luo, S. Wu, Y. Li, Y. Wang, and H. Lv, “Improvement of Endurance in HZO-Based Ferroelectric Capacitor Using Ru Electrode,” IEEE Electron Device Lett. 40(11), 1744–1747 (2019). [CrossRef]
19. P. K. Singaravelu, G. Devarapu, S. A. Schulz, Q. Wilmart, S. Malhouitre, and S. Olivier, “Low-loss, compact, spot-size-converter based vertical couplers for photonic integrated circuits,” J. Phys. D: Appl. Phys. 52(21), 214001 (2019). [CrossRef]
20. S. J. Kim, J. Mohan, S. R. Summerfelt, and J. Kim, “Ferroelectric Hf0.5Zr0.5O2 Thin Films: A Review of Recent Advances,” JOM 71(1), 246–255 (2019). [CrossRef]
21. S. Pal, P. K. Tiwari, and S. Gupta, “A Proposal for an Electrostatic Doping-Assisted Electroabsorption Modulator for Intrachip Communication,” IEEE Trans. Electron Devices 66(5), 2269–2275 (2019). [CrossRef]
22. J. Liu, M. Beals, A. Pomerene, S. Bernardis, R. Sun, J. Cheng, L. C. Kimerling, and J. Michel, “Waveguide-integrated, ultralow-energy GeSi electro-absorption modulators,” Nat. Photonics 2(7), 433–437 (2008). [CrossRef]
23. G. Gupta, B. Rajasekharan, and R. J. E. Hueting, “Electrostatic Doping in Semiconductor Devices,” IEEE Trans. Electron Devices 64(8), 3044–3055 (2017). [CrossRef]
24. T. J. Seok, N. Quack, S. Han, R. S. Muller, and M. C. Wu, “Large-scale broadband digital silicon photonic switches with vertical adiabatic couplers,” Optica 3(1), 64 (2016). [CrossRef]
25. J. Lin, K. Huang, and B. Jheng, “Performances of a Capacitorless 1T-DRAM Using Polycrystalline Silicon Thin-Film Transistors With Trenched Body,” IEEE Electron Device Lett. 29(11), 1222–1225 (2008). [CrossRef]
26. J. S. Orcutt, S. D. Tang, S. Kramer, K. Mehta, H. Li, V. Stojanović, and R. J. Ram, “Low-loss polysilicon waveguides fabricated in an emulated high-volume electronics process,” Opt. Express 20(7), 7243–7254 (2012). [CrossRef]
27. R. Soref and B. Bennett, “Electrooptical effects in silicon,” IEEE J. Quantum Electron. 23(1), 123–129 (1987). [CrossRef]
28. B. Rahman and J. Davies, “Finite-element solution of integrated optical waveguides,” J. Lightwave Technol. 2(5), 682–688 (1984). [CrossRef]
29. R. C. Smith and Z. Ounaies, “A Domain Wall Model for Hysteresis in Piezoelectric Materials,” J. Intel. Mat. Syst. Str. 11(1), 62–79 (2000). [CrossRef]
30. K. Ni, M. Jerry, J. A. Smith, and S. Datta, “A Circuit Compatible Accurate Compact Model for Ferroelectric-FETs,” in Proc. IEEE Symp. VLSI Technol. (IEEE, 2018), pp. 131–132.
31. A. W. Elshaari and S. F. Preble, “10Gb/s broadband silicon electro-optic absorption modulator,” Opt. Commun. 283(14), 2829–2834 (2010). [CrossRef]
32. Z. J. Yan, R. Xu, Y. Y. Wang, S. Chen, Y. L. Fan, and Z. M. Jiang, “Thin HfO2 films grown on Si(100) by atomic oxygen assisted molecular beam epitaxy,” Appl. Phys. Lett. 85(1), 85–87 (2004). [CrossRef]
33. J. Willems, J. Haes, and R. Baets, “The bidirectional mode expansion method for two-dimensional waveguides: the TM case,” Opt Quant Electron 27(10), 995–1007 (1995). [CrossRef]
34. A. Emboras, I. Goykhman, B. Desiatov, N. Mazurski, L. Stern, J. Shappir, and U. Levy, “Nanoscale plasmonic memristor with optical readout functionality,” Nano Lett. 13(12), 6151–6155 (2013). [CrossRef]
35. C. Ríos, N. Youngblood, Z. Cheng, M. L. Gallo, W. H. P. Pernice, C. D. Wright, A. Sebastian, and H. haskaran, “In-memory computing on a photonic platform,” Sci. Adv. 5(2), (2019).