Abstract
Neuromorphic photonics has recently emerged as a promising hardware accelerator, with significant potential speed and energy advantages over digital electronics for machine learning algorithms, such as neural networks of various types. Integrated photonic networks are particularly powerful in performing analog computing of matrixvector multiplication (MVM) as they afford unparalleled speed and bandwidth density for data transmission. Incorporating nonvolatile phasechange materials in integrated photonic devices enables indispensable programming and inmemory computing capabilities for onchip optical computing. Here, we demonstrate a multimode photonic computing core consisting of an array of programable mode converters based on onwaveguide metasurfaces made of phasechange materials. The programmable converters utilize the refractive index change of the phasechange material Ge_{2}Sb_{2}Te_{5} during phase transition to control the waveguide spatial modes with a very high precision of up to 64 levels in modal contrast. This contrast is used to represent the matrix elements, with 6bit resolution and both positive and negative values, to perform MVM computation in neural network algorithms. We demonstrate a prototypical optical convolutional neural network that can perform image processing and recognition tasks with high accuracy. With a broad operation bandwidth and a compact device footprint, the demonstrated multimode photonic core is promising toward largescale photonic neural networks with ultrahigh computation throughputs.
Introduction
The unmet gap between the rate of energy efficiency improvement of current digital electronics and the fastgrowing load of computation by emerging applications such as machine learning and artificial intelligence^{1,2} has once again brought optical computing into focus^{3,4,5,6}. Integrated photonics provides a scalable hardware platform to realize largescale optical networks on a chip, which affords an enormous bandwidth density that is unreachable for electronics^{7,8,9}. To use integrated photonics for optical computing, programmable photonic components and nonlinear elements are indispensable building blocks. Phasechange materials (PCM) recently emerged as an ideal material system to realize optical programmability^{10,11,12}. The optical properties of PCMs change dramatically during the phase transition, which can be electrically or optically controlled. Harnessing this has allowed for embodiments of programmable optical switches, couplers, lens, and metamaterials to be demonstrated^{13,14,15,16,17,18,19,20,21}. The phase change in the chalcogenide family of GeSbTe alloys is nonvolatile, requiring no sustaining power supply to retain the programmed state or stored information^{19,20,21,22,23,24,25,26}. Their use in programmable photonic devices thus can have a significant advantage in power consumption over electrooptic^{27,28,29} or thermooptic methods^{30,31,32}. Photonic devices incorporating those nonvolatile PCMs thus can realize optical memories and perform inmemory computing simply by measuring the transmission of the optical input data through the programmed device^{33,34,35}. Proliferating these phasechange photonic devices in a scalable network, prototypes of optical neural networks (ONN) have been proposed and demonstrated^{35,36,37,38}.
Here, we report a programmable waveguide mode converter based on a phasegradient metasurface made of phasechange material Ge_{2}Sb_{2}Te_{5} (GST). This phasechange metasurface mode converter (PMMC) utilizes GST’s large refractive index change during its phase transition to control the conversion of the waveguide’s two spatial modes (TE_{0} and TE_{1} modes). The PMMC can be programmed to control the waveguide mode contrast precisely at 64 distinguishable levels, which is used to represent the weight parameters with 6bit precision in MVM computation. We build a 2 × 2 array of PMMCs and implement them as programmable kernels to realize a multimode optical convolutional neural network (OCNN). By performing image processing tasks such as edge detection and pattern recognition, we demonstrate the OCNN’s viability and potential in largescale optical computing.
Results
Highprecision programmable phasechange mode converter
The design of the PMMC is based on the principle of a phasegradient metasurface but replacing noble metals with phasechange materials^{39}. Fig. 1a shows a 3D schematic of the design, which consists of a linear array of GST nanoantennae directly integrated on a silicon nitride (SiN) waveguide. Each GST nanoantenna scatters the waveguide mode and causes a phase shift Φ, which depends on its geometry (e.g., width), as well as the refractive index of its material (Fig. 2b). A linear array of such nanoantennae with tapering widths thus produces a spatial gradient of the scattering phases dΦ/dx, which is equivalent to a wavevector k_{g}. If the phasegradient metasurface is designed such that k_{g} matches the wavevector difference between two spatial modes of the waveguide: k_{mode1}−k_{mode2}, it satisfies the phasematching condition and facilitates the conversion between the two modes. Such phasegradient metasurfaces for waveguide mode conversion were realized using noble metals or dielectrics materials and thus lacked tunability. Here, we use GST, which has a large change in its optical properties when a phase transition happens. When the GST is in the amorphous phase (aGST), its refractive index n is ~4.7 (representative value in the literature, the same hereafter)^{40}. In contrast, when it is turned to the crystalline phase (cGST), n increases to ~7.5 with a drastic change of 2.8 over the whole measured spectral range from 1540 nm to 1580 nm (See Supplementary Fig. 1a for more detailed information). This change will significantly modify the scattered phase of each GST nanoantenna (Fig. 2b) so as to modify the metasurface’s function. Fig. 1c plots the simulated phase of the scattered fields inside the waveguide by a single nanoantenna of 30nmthick GST as a function of its width and for aGST and cGST phases. Since cGST has a much larger n, the scattered phase shows a much stronger dependence on the width than the aGST phase. By controlling the geometry of the GST nanoantennae and the interval between adjacent ones in the array, a welldefined phase gradient dΦ/dx is established (see Supplementary Note 2 and 3 for details). The entire metasurface consists of an array of 25 nanoantennae with tapering widths from 510 nm to 84 nm (shaded region in Fig. 1c) and is patterned on a SiN waveguide 1.8 µm wide and 330 nm thick. The waveguide supports two transverseelectric modes: the fundamental TE_{0} mode and the firstorder TE_{1} mode. We design the metasurface, in the cGST phase, to have a uniform dΦ = 2.5° for every dx = 400 nm to satisfy the generalized phasematching condition, \(k_0(n_{{\mathrm{TE0}}}  n_{{\mathrm{TE1}}}) = N \cdot d{\varPhi}/dx\), where k_{0} is the freespace wavevector, \(n_{{\mathrm{TE0}}}\) and \(n_{{\mathrm{TE1}}}\) are the effective index of the TE_{0} and TE_{1} modes, respectively, and N is the number of interactions between the guided modes and the metasurface. The cGST metasurface thus can efficiently convert the TE_{0} mode to the TE_{1} mode, as shown by the finitedifference timedomain (FDTD) simulation result in Fig. 1d. When the GST is transitioned to the aGST phase, as shown in Fig. 1c, the dΦ/dx is much reduced and thus insufficient for the phasematching condition so that mode conversion between TE_{0} and TE_{1} modes does not occur, which is clearly seen in Fig. 1e. Therefore, the GST phasegradient metasurface, as designed here, functions as a programmable waveguide mode converter controlled by the tunable material phase of the GST.
Fig. 2a–c shows the scanning electron microscope images of the complete PMMC device. The 30 nm thick GST film is deposited by sputtering on Si_{3}N_{4} on an oxidized silicon substrate. It is then patterned into metasurface with electron beam lithography and plasma etching, and conformally encapsulated with a 218nmthick layer of Al_{2}O_{3} deposited by atomic layer deposition. The photonic circuits of Si_{3}N_{4}, including multimode waveguides, directional couplers and grating couplers, are patterned with standard processes^{17}. A pair of asymmetric directional couplers (Fig. 2c) is designed to function as mode selectors to selectively couple only the TE_{1} mode component in the multimode waveguide with the TE_{0} mode component in the singlemode waveguide (See Supplementary Note 5 for details). Fig. 2a depicts the measurement and control scheme. To program the PMMC, we use optical pulses to control the phase of the GST film for simplicity^{41}. Previously, electrical control using integrated microheaters has been demonstrated by a number of groups, including us^{17,26,42,43,44}. When operating the PMMC, an optical signal is input in the TE_{0} mode to the PMMC and converted to TE_{1} mode with a proportion controlled by the state of the GST metasurface. At the output of the PMMC, the TE_{1} component is separated by the mode selector and coupled out at the second port while the TE_{0} component remains in and outputs from the multimode waveguide. The output powers of both modes are measured to determine their respective transmission coefficients. Fig. 2d shows the transmission spectrum of the PMMC when the metasuface is set to be either in the fully aGST or cGST phases. The insertion losses of the input and output fibers and grating couplers have been accounted for by calibration measurements. In the aGST phase, the device is in the onstate for the TE_{0} mode with a high transmission T^{on} over a broad wavelength range (1540–1580 nm). The lowest insertion loss is 0.9 dB at 1575 nm wavelength. A small portion (<−10 dB) of the TE_{1} mode is generated due to the asymmetric perturbation induced by the metasurface even though the aGST phase has a low refractive index. The situation changes dramatically when the metasurface is transitioned to the cGST phase and converts the TE_{0} mode to the TE_{1} mode effectively. In this offstate for the TE_{0} mode, its transmission T^{off} is <−15 dB over the entire measured bandwidth. The corresponding switching extinction ratio, defined as \({\Delta}T/T^{{\mathrm{off}}} = \left( {T^{{\mathrm{on}}}  T^{{\mathrm{off}}}} \right)/T^{{\mathrm{off}}}\), is ~16 dB or 4000%, which is more than 10fold improvement compared to previously reported switch devices using GST^{24,43,45}. This large switching ratio stems from the phase engineering approach to effectively use GST’s large refractive index change during its phasetransition, as opposed to only using the absorption coefficient change, to facilitate scattering into a different mode that is filtered. The total area of the GST in the metasurface is only 1.3 µm^{2}, significantly smaller than that in prior devices, and thus in principle, our device consumes less energy to switch. As expected from energy conservation, the TE_{1} mode is switched in the opposite way to the TE_{0} mode. From aGST to cGST phase, the TE_{1} transmission increases from ~−10 dB to ~−6.5 dB, with the insertion loss due to cGST’s absorption. Another important parameter to quantify a mode converter’s performance is the mode purity in the multimode waveguide, defined as \(\beta _{{\mathrm{TE0(TE1)}}} = P_{{\mathrm{TE0(TE1)}}}/\left( {P_{{\mathrm{TE0}}} + P_{{\mathrm{TE1}}}} \right)\), where P_{TE0} (P_{TE1}) is the power in the TE_{0} (TE_{1}) mode. The PMMC shows very high performance in controlling mode purity. As shown in Fig. 2e, when switching the GST from aGST to cGST phase, the PMMC efficiently converts TE_{0} mode to TE_{1} mode, changing the mode purity from β_{TE0} > 80% to β_{TE1} > 85% over a broad bandwidth, showing an excellent agreement with the numerical simulation results.
PMMC photonic kernel
The phase composition of the GST in the metasurface can be continuously tuned by partial phase transition so that the PMMC can be continuously programmed to multiple intermediate levels of phase purity values. We program the PMMC with a sequence of 50nslong control pulses to “quench” the GST progressively from the fully cGST phase toward the fully aGST phase. As a result, the TE_{1} mode purity β_{TE1} increases stepwise. Since the mode selector separates the two modes, we can measure their power and calculate the difference to determine the mode contrast \({\varGamma} = \beta _{{\mathrm{TE0}}}  \beta _{{\mathrm{TE1}}}\), which is used as a programming parameter. Fig. 2f demonstrates the multilevel programmability of the PMMC, in which Γ is sequentially set to 64 distinguishable levels between −0.73 to +0.67 at 1555 nm. Since the theoretical range of Γ is \((  1,1)\), it is an ideal parameter to represent the elements in the matrix w, with both positive and negative values, in a multiplyaccumulate (MAC) operation: x → x ∙ w + b, where b is the bias parameter. MAC is the constitutional step of matrixvector multiplication (MVM) in all neural network algorithms. The PMMC allows storing w by programming Γ in the GST metasurface as a nonvolatile memory. Inmemory MAC computing can be performed with the PMMC by a measurement of the transmitted power when the input data x is encoded in the power of the input optical signal. The lower inset of Fig. 2f shows the histograms of 20 repeated programming operations to set the PMMC mode contrast at two adjacent levels (levels 30 and 31), respectively. The wellseparated histograms clearly demonstrate the device’s programming resolution and accuracy. The demonstrated 64level programmability of the PMMC—the highest to the best of our knowledge for phasechange photonic devices^{24}—corresponds to 6bit resolution in setting w, which is critical to the training and inference precision of the neural network^{46,47}.
We harness the PMMC’s highprecision programmability and inmemory computing capability to demonstrate an optical convolutional neural network (OCNN)^{28,29,30,48}. A typical CNN consists of an input layer and an output layer, which are connected by multiple hidden layers in between. The hidden layers usually consist of a series of convolutional layers followed by pooling layers and fully connected layers at the end. We design a prototype OCNN using a small network of PMMCs to implement patchkernel matrix multiplication to compute convolution. Fig. 3a illustrates the operation principle of the OCNN for image processing, where an input grayscale image of dimensions n × n is convolved with a kernel of dimensions k × k to compute an activation map of dimension (n–k+1) × (n–k+1), assuming the convolution stride is 1. When operating the OCNN, we group the input image into (n–k+1)^{2} patches (the shaded area in the upper panel of Fig. 3a) with the same dimensions as the convolution kernel, k^{2}. Each patch corresponds to the receptive field of an element in the activation map accordingly. Thus, a convolution operation requires (n–k+1)^{2} × k^{2} MAC operations in total, which is a high load of computation and can most benefit from optical computing’s speed and energy advantages.
To compute the convolution, (n–k+1)^{2} patch matrices of the input image are optically fed into the photonic kernel sequentially while the kernel elements, that is, the PMMCs, are programmed to fixed values. At each timeframe of the computation, the corresponding patch matrix is reshaped into a single column of data with the length k^{2}. The data is input into the optical system in k^{2} channels as sequences of incoherent optical pulses, whose power amplitude is controlled by a variable optical attenuator (VOA) to encode the value of each pixel value X_{ij} in the greyscale image. The corresponding element W_{ij} of the kernel matrix is programmed as the mode contrast Γ of each PMMC. The resulting transmitted power of TE_{0} and TE_{1} modes are then summed incoherently using two photodetectors. Their output difference is calculated electronically and used in postprocessing steps. As a result, the output will correspond to a time series of patchkernel MVM with the amplitude encoding the values of the computation results, which is the activation map of convolution. Since the modal contrast Γ of our PMMCs can assume both positive and negative values, it can represent the kernel matrix elements without the need of an additional offset, which otherwise would take additional steps to set in each computation cycle.
Convolutional edge detection with PMMC core
Experimentally, we build a smallscale, fourchannel system with four PMMCs to represent a 2 × 2 kernel matrix, as shown in the optical image in Fig. 3b. As a demonstration, we perform the convolution of a 256 × 256 8bit grayscale image of a cameraman (Fig. 3c) to detect its edge features. As shown in Fig. 3b, the TE_{0} mode output coming from all the PMMCs is combined using onchip Yjunctions, while the TE_{1} mode output power is combined offchip because the same ports are used to program the PMMCs optically. Because combining four incoherent sources using Yjunctions will inherently reduce the power by a factor of 4, we rescale the measured TE_{0} mode power by this factor when calculating the power differences between two modes. To detect vertical and horizontal edges, kernel matrices as in the right column of Fig. 3d, e are used, and so are the PMMCs programmed. Take the vertical edge detection for example, the kernel is set to be \(\left[ {\begin{array}{*{20}{c}} {  1} & 1 \\ {  1} & 1 \end{array}} \right]\) so to compute the discrete firstorder derivative, \(X_{i + 1,j} + X_{i + 1,j + 1}  X_{i,j}  X_{i,j + 1}\), where i, j are the indices of the input image matrix. Each kernel element W_{ij} is stored as the mode contrast value Γ in the corresponding PMMC, with W_{ij} = 1(−1) corresponding to the fully aGST (cGST) phase (see Supplementary Note 9 for a more detailed description about the operation procedure). The computed images after convolution without any postprocessing are shown in the left column of Fig. 3d, e, for horizontal and vertical edge detection, respectively. The two images are then added to produce the right image in Fig. 3b, which highlights silhouettes of the objects with sharp edges such as the cameraman and the buildings in the original image, while suppressing smooth features such as the sky and the water. The optically computed edge detection image agrees very well with the calculated result using conventional image processing algorithms (see Supplementary Fig. 16). This result verifies the capacity and fidelity of optical convolution performed with the PMMCbased photonic kernel, which is a prerequisite for an OCNN.
OCNN for image recognition
Beyond the convolution layer, the MAC computation performed with optical signals and the PMMC network can also be applied to the pooling (average pooling) and the fully connected layers, where the PMMCs are used as weight banks instead, to realize a complete OCNN. In our experiment, we sequentially reuse the PMMC array in both convolution and fully connected layers to demonstrate an OCNN and perform proofofconcept imaging recognition tasks of distinguishing handwritten numbers “1” and “2” from the MNIST database. Fig. 4a illustrates the architecture and processes of the OCNN. The 28 × 28 pixels, 8bit grayscale images of number “1” or “2” are fed into the input layer as optical signals. The data is then convolved with two 2 × 2 photonic kernels K_{1} and K_{2} to generate two 27 × 27 images of activation maps. After adding a bias b_{1} and applying the nonlinear ReLu function, the output images are sent to an average pooling layer with a subsampling factor of 27, which reduces the images a 2 × 1 vector. This vector is then fed into the fully connected layer with a 2 × 2 photonic weight bank K_{3} programmed in the PMMC array, added with a bias b_{2} and applied the standard sigmoid function. The final output is a vector that gives the identified class of the input image, that is, \(\left[ {\begin{array}{*{20}{c}} 1 & 0 \end{array}} \right]^T\) corresponds to the number “1” and \(\left[ {\begin{array}{*{20}{c}} 0 & 1 \end{array}} \right]^T\) corresponds to the number “2”. In this OCNN, the MVM computations such as the convolution and the fully connected layers are all performed optically with the PMMCs, whereas bias and nonlinear functions are realized electronically.
Before using the OCNN, we first train all the parameters in the layers with the standard backpropagation algorithm using the gradient descent method^{49}. The training set consists of 11,000 images of the handwritten number “1” or “2” from MNIST training images. The training yields values for each element in the convolutional kernels and the weight bank, as shown in Fig. 4b, d. We then program the PMMC array to represent these elements. In Fig. 4c, we show the raw data of the convolutional activation maps encoded in time series of optical signals, which are the output from the PMMC array after the input image convolves with the photonic kernel K_{1} and K_{2}. Since each photonic processing layer results in electrical signals output from the photodetectors, electronic postprocessing is performed to add bias and apply nonlinear function and pooling. The resultant data is recoded into optical signals and fed to the next photonic layer. Further experimental details are included in the Supplementary Note 9. We evaluate the system’s performance after training on a recognition test set, which consists of 100 randomly chosen “1” or “2” images (55 number “1” and 45 number “2”) from the MNIST testing image database. Fig. 4e shows the result that our OCNN correctly identified 91 out of 100 cases (9% error rate), which compares squarely with the result of a computer (10% error rate). The slight difference is mainly caused by the small deviation of the experimentally programmed values in the matrices (K_{1}, K_{2} and K_{3}) from the trained values, which occurs when the system’s conditions drift during operation. This result successfully demonstrates the OCNN’s viability and accuracy in performing standard neural network algorithms.
Discussion
In summary, we have demonstrated a compact programmable waveguide mode converter using GSTbased phasegradient metasurface with high programming resolution, efficiency and broadband operation. We have built a photonic kernel based on an array of such PMMC devices and implemented an optical convolutional neural network to perform image processing and recognition tasks. Our results show that phasechange photonic devices, such as the PMMC demonstrated here, can enable robust and flexible programmability and realize a plethora of unique optical functionalities that are scalable for largescale optical computing and neuromorphic photonics. Although optical computation in this work is performed at a low speed of ~1 kHz by using lowspeed VOAs to encode data into optical signals, stateoftheart integrated photonic transmitters and photodetectors can drive the system at a speed of many 10 s of Gbits/sec^{50,51}. Using wavelength division multiplexing (WDM) can further increase the number of parallel computation. The 2 × 2 array prototype system demonstrated in this work performs optical computation incoherently in a broadband. It can be scaled up toward a large network using a photonic crossbar array architecture^{52,53,54,55,56,57} (see Supplementary Note 10 for details of such a design), and compares favorably with other photonic computing schemes using coherent methods^{30} or optical resonators^{28,58,59}. The feasible size (n × m) of such crossbar arrays will not be limited by the insertion loss of the PMMC (~7 dB for TE_{1} mode, Fig. 2d); rather it will be limited by the directional couplers with coupling efficiency of 1/n, as is needed to equally combine signals from n units. Scaling up to a large network thus faces the challenge of diminishing optical power unless with onchip optical amplification, which is not yet available. Still, an OCNN system using the PMMC device can afford an extremely high areal computing density (defined as MAC operations per time per unit area) because of its compact footprint of ~80 × 20 µm^{2} (Fig. 2a, including the mode selector). For example, assuming a moderate datarate of 10 Gbits/sec and 4 WDM wavelengths in parallel per channel, the computing density will reach an upperbound value of 25 TOPS/mm^{2} (Teraoperations per second per mm^{2}), which is significantly higher than that of digital electronic accelerators such as GPUs and tensor processing units (TPUs)^{60,61}. Using silicon instead of silicon nitride can further reduce the device footprint to increase the computing density^{62}. Besides MAC operation, the equally important computing processes of applying nonlinear functions and pooling can also be achieved optically by using elements such as nonlinear optical resonators, modulators, and amplifiers^{27,28,63}. Alternatively, a hybrid photonicelectronic system may optimally balance energyefficiency and speed advantages of photonic systems, while realizing flexible nonlinearity, connectivity, and training precision using microelectronics^{26,64,65}. With these advances and after overcoming the scaling challenge, the photonic neural network accelerator will be very promising for AI in data centers where massive optical interconnects have already been deployed.
Data availability
The data that support the findings of this study are available from the corresponding author upon reasonable request.
Code availability
No custom computer code or mathematical algorithm is used to generate the results that are reported in this study.
References
Marr, B., Degnan, B., Hasler, P. & Anderson, D. Scaling Energy Per Operation via an Asynchronous Pipeline. IEEE Transactions on Very Large Scale. Integration (VLSI) Systems. 21, 147–151 (IEEE, 2013).
Jones, N. How to stop data centres from gobbling up the world’s electricity. Nature 561, 163–166 (2018).
Athale, R. & Psaltis, D. Optical computing: past and future. Opt. Photon. News 27, 32–39 (2016).
Solli, D. R. & Jalali, B. Analog optical computing. Nat. Photon. 9, 704–706 (2015).
Prucnal, P. R. & Shastri, B. J. Neuromorphic Photonics (CRC Press, 2017).
Caulfield, H. J. & Dolev, S. Why future supercomputing requires optics. Nat. Photon. 4, 261–263 (2010).
Zhang, C., Zhang, S., Peters, J. D. & Bowers, J. E. 8 × 8 × 40 Gbps fully integrated silicon photonic network on chip. Optica 3, 785–786 (2016).
Shen, Y. W. et al. Silicon photonics for extreme scale systems. J. Lightwave Technol. 37, 245–259 (2019).
Wade, M. et al. In 2018 European Conference on Optical Communication (ECOC). 1–3 (IEEE, 2018).
Wuttig, M., Bhaskaran, H. & Taubner, T. Phasechange materials for nonvolatile photonic applications. Nat. Photon. 11, 465–476 (2017).
Yang, Z. & Ramanathan, S. Breakthroughs in photonics 2014: phase change materials for photonics. IEEE Photon. J. 7, 1–5 (2015).
Zhang, W., Mazzarello, R., Wuttig, M. & Ma, E. Designing crystallization in phasechange materials for universal memory and neuroinspired computing. Nat. Rev. Mater. 4, 150–168 (2019).
Briggs, R. M., Pryce, I. M. & Atwater, H. A. Compact silicon photonic waveguide modulator based on the vanadium dioxide metalinsulator phase transition. Opt. Express 18, 11192–11201 (2010).
Wang, Q. et al. Optically reconfigurable metasurfaces and photonic devices based on phase change materials. Nat. Photon. 10, 60–U75 (2016).
Chu, C. H. et al. Active dielectric metasurface based on phase‐change medium. Laser Photon. Rev. 10, 986–994 (2016).
Yin, X. et al. Beam switching and bifocal zoom lensing using active plasmonic metasurfaces. Light. Sci. Appl. 6, e17016 (2017).
Wu, C. et al. Lowloss integrated photonic switch using subwavelength patterned phase change material. ACS Photon. 6, 87–92 (2018).
Cheng, Z. et al. Devicelevel photonic memories and logic applications using phasechange materials. Adv. Mater. 30, e1802435 (2018).
Xu, P., Zheng, J., Doylend, J. K. & Majumdar, A. Lowloss and broadband nonvolatile phasechange directional coupler switches. ACS Photon. 6, 553–557 (2019).
Zhang, Y. et al. Broadband transparent optical phase change materials for highperformance nonvolatile photonics. Nat. Commun. 10, 4279 (2019).
de Galarreta, C. R. et al. Nonvolatile reconfigurable phasechange metadevices for beam steering in the near infrared. Adv. Funct. Mater. 28, 1704993 (2018).
Stegmaier, M., Ríos, C., Bhaskaran, H., Wright, C. D. & Pernice, W. H. P. Nonvolatile alloptical 1 × 2 switch for chipscale photonic networks. Adv. Opt. Mater. 5, 1600346 (2017).
Zhang, Q. et al. Broadband nonvolatile photonic switching based on optical phase change materials: beyond the classical figureofmerit. Opt. Lett. 43, 94 (2018).
Li, X. et al. Fast and reliable storage using a 5bit, nonvolatile photonic memory cell. Optica 6, 1–6 (2019).
Martins, T. et al. Fiberintegrated phasechange reconfigurable optical attenuator. Appl. Photon. 4, 111301 (2019).
Zheng, J. et al. Nonvolatile electrically reconfigurable integrated photonic switch enabled by a silicon PIN diode heater. Adv. Mater. 32, e2001218 (2020).
George, J. K. et al. Neuromorphic photonics with electroabsorption modulators. Opt. Express 27, 5181–5191 (2019).
Tait, A. N. et al. Silicon photonic modulator neuron. Phys. Rev. Appl. 11, 064043 (2019).
Hamerly, R., Bernstein, L., Sludds, A., Soljacic, M. & Englund, D. Largescale optical neural networks based on photoelectric multiplication. Phys. Rev. X 9, 021032 (2019).
Shen, Y. et al. Deep learning with coherent nanophotonic circuits. Nat. Photon. 11, 441–446 (2017).
Sun, J., Timurdogan, E., Yaacobi, A., Hosseini, E. S. & Watts, M. R. Largescale nanophotonic phased array. Nature 493, 195–199 (2013).
Ribeiro, A., Ruocco, A., Vanacker, L. & Bogaerts, W. Demonstration of a 4x4port universal linear circuit. Optica 3, 1348–1357 (2016).
Bocker, R. P. Matrix multiplication using incoherent optical techniques. Appl. Opt. 13, 1670–1676 (1974).
Ríos, C. et al. Inmemory computing on a photonic platform. Sci. Adv. 5, eaau5759 (2019).
Chakraborty, I., Saha, G. & Roy, K. Photonic inmemory computing primitive for spiking neural networks using phasechange materials. Phys. Rev. Appl. 11, 014063 (2019).
Caulfield, H. J., Kinser, J. & Rogers, S. K. Optical neural networks. Proc. IEEE 77, 1573–1583 (1989).
Feldmann, J. et al. Parallel convolution processing using an integrated photonic tensor core. arXiv preprint arXiv:2002.00281 (2020).
Feldmann, J., Youngblood, N., Wright, C. D., Bhaskaran, H. & Pernice, W. H. P. Alloptical spiking neurosynaptic networks with selflearning capabilities. Nature 569, 208–214 (2019).
Li, Z. et al. Controlling propagation and coupling of waveguide modes using phasegradient metasurfaces. Nat. Nanotechnol. 12, 675–683 (2017).
Park, J.W. et al. Optical properties of pseudobinary GeTe, Ge2Sb2Te5, GeSb2Te4, GeSb4Te7, and Sb2Te3 from ellipsometry and density functional theory. Phys. Rev. B 80, 115209 (2009).
Liu, Y., Aziz, M. M., Shalini, A., Wright, C. D. & Hicken, R. J. Crystallization of Ge2Sb2Te5 films by amplified femtosecond optical pulses. J. Appl. Phys. 112, 123526 (2012).
Farmakidis, N. et al. Plasmonic nanogap enhanced phasechange devices with dual electricaloptical functionality. Sci. Adv. 5, eaaw2687 (2019).
Zhang, H. et al. Miniature multilevel optical memristive switch using phase change material. ACS Photon. 6, 2205–2212 (2019).
RodriguezHernandez, G., Hosseini, P., Ríos, C., Wright, C. D. & Bhaskaran, H. Mixedmode electrooptical operation of Ge2Sb2Te5 nanoscale crossbar devices. Adv. Electron. Mater. 3, 1700079 (2017).
Ríos, C. et al. Integrated allphotonic nonvolatile multilevel memory. Nat. Photon. 9, 725–732 (2015).
Giannopoulos, I. et al. In 2018 IEEE International Electron Devices Meeting (IEDM). 27.27.21–27.27.24 (IEEE, 2018).
Le Gallo, M. et al. Mixedprecision inmemory computing. Nat. Electron. 1, 246–253 (2018).
Nahmias, M. A. et al. Photonic multiplyaccumulate operations for neural networks. IEEE J. Sel. Top. Quant. 26, 1–18 (2020).
Convolution Neural Network  simple code  simple to use (MATLAB Central File Exchange, 2020).
Xiong, C. et al. Monolithic 56 Gb/s silicon photonic pulseamplitude modulation transmitter. Optica 3, 1060–1065 (2016).
Moazeni, S. et al. A 40Gb/s PAM4 Transmitter Based on a RingResonator Optical DAC in 45nm SOI CMOS. IEEE J. SolidState Circuits 52, 3503–3516 (2017).
Sawchuk, A. A. & Jenkins, B. K. In Optical Computing. 143–153 (International Society for Optics and Photonics 1986).
Joshi, A. et al. In 2009 3rd ACM/IEEE International Symposium on NetworksonChip. 124133 (IEEE, 2009).
Khope, A. S. P. et al. Multiwavelength selective crossbar switch. Opt. Express 27, 5203–5216 (2019).
Ohno, S., Toprasertpong, K., Takagi, S. & Takenaka, M. Si microring resonator crossbar arrays for deep learning accelerator. Jpn J. Appl. Phys. 59, SGGE04 (2020).
Nahmias, M. A. et al. Photonic multiplyaccumulate operations for neural networks. IEEE J. Sel. Top. Quant. Electron. 26, 1–18 (2019).
Han, S., Seok, T. J., Quack, N., Yoo, B.W. & Wu, M. C. Largescale silicon photonic switches with movable directional couplers. Optica 2, 370–375 (2015).
Tait, A. N., Chang, J., Shastri, B. J., Nahmias, M. A. & Prucnal, P. R. Demonstration of WDM weighted addition for principal component analysis. Opt. Express 23, 12758 (2015).
Tait, A. N., Nahmias, M. A., Shastri, B. J. & Prucnal, P. R. Broadcast and weight: an integrated network for scalable photonic spike processing. J. Lightwave Technol. 32, 4029–4041 (2014).
Silver, D. et al. A general reinforcement learning algorithm that masters chess, shogi, and Go through selfplay. Science 362, 1140–114 (2018).
Jouppi, N. P. et al. In Proceedings of the 44th Annual International Symposium on Computer Architecture  ISCA ‘17 112 (ACM Press, New York, New York, USA, 2017).
Li, X. et al. Experimental investigation of silicon and silicon nitride platforms for phasechange photonic inmemory computing. Optica 7, 218–225 (2020).
Gayen, D. K., Chattopadhyay, T., Pal, R. K. & Roy, J. N. Alloptical Multiplication with the help of Semiconductor Optical Amplifier—assisted Sagnac Switch. J. Comput. Electron. 9, 57–67 (2010).
Atabaki, A. H. et al. Integrating photonics with silicon nanoelectronics for the next generation of systems on a chip. Nature 556, 349–354 (2018).
Bangari, V. et al. Digital electronics and analog photonics for convolutional neural networks (DEAPCNNs). IEEE J. Sel. Top. Quant. 26, 1–13 (2020).
Acknowledgements
We acknowledge the funding support provided by the ONR MURI (Award No. N000141712661). Part of this work was conducted at the Washington Nanofabrication Facility/Molecular Analysis Facility, a National Nanotechnology Coordinated Infrastructure (NNCI) site at the University of Washington, which is supported in part by funds from the National Science Foundation (awards NNCI1542101, 1337840, and 0335765), the National Institutes of Health, the Molecular Engineering & Sciences Institute, the Clean Energy Institute, the Washington Research Foundation, the M. J. Murdock Charitable Trust, Altatech, ClassOne Technology, GCE Market, and Google and SPTS.
Author information
Authors and Affiliations
Contributions
C.W. and M.L. conceived the research. C.W. fabricated the devices, performed the measurements and analysed the data. H.Y. and I.T. deposited the Ge_{2}Sb_{2}Te_{5} thin films and characterized the optical properties. S.L. and R.P. assisted the fabrication and characterization. C.W., I.T., and M.L. cowrote the manuscript.
Corresponding author
Ethics declarations
Competing interests
The authors declare no competing interests.
Additional information
Peer review information Nature Communications thanks Wolfram Pernice and the other, anonymous, reviewer(s) for their contribution to the peer review of this work.
Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Supplementary information
Rights and permissions
Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made. The images or other third party material in this article are included in the article’s Creative Commons license, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons license and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this license, visit http://creativecommons.org/licenses/by/4.0/.
About this article
Cite this article
Wu, C., Yu, H., Lee, S. et al. Programmable phasechange metasurfaces on waveguides for multimode photonic convolutional neural network. Nat Commun 12, 96 (2021). https://doi.org/10.1038/s4146702020365z
Received:
Accepted:
Published:
DOI: https://doi.org/10.1038/s4146702020365z
This article is cited by

Integrated optical memristors
Nature Photonics (2023)

Lithographyfree reconfigurable integrated photonic processor
Nature Photonics (2023)

Inmemory photonic dotproduct engine with electrically programmable weight banks
Nature Communications (2023)

Compact optical convolution processing unit based on multimode interference
Nature Communications (2023)

Nonvolatile electrically programmable integrated photonics with a 5bit operation
Nature Communications (2023)
Comments
By submitting a comment you agree to abide by our Terms and Community Guidelines. If you find something abusive or that does not comply with our terms or guidelines please flag it as inappropriate.