Open Access
Add to BibSonomy
Abstract
For the advantages of the faster computation speed and lower energy consumption, all-optical computation has attracted great attention compared with the traditional electric computation method. Optical switches are the critical elementary units of optical computation devices. However, the traditional optical switches have two shortcomings, expending the outside energy to keep the switch state and the weak multi-level adjustable ability, which greatly restrict the realization of the large-scale photonic integrated circuits and optical spiking neural networks. In this paper, we use a subwavelength grating slot-ridge (SWGSR) waveguides on the silicon platform to design a nonvolatile multi-level adjustable optical switch based on the phase change material Ge2Sb2Te5 (GST). Changing the phase state of GST can modulate the transmission of the optical switch, and the change of the optical transmittance of the optical switch is about 70%, which is much higher than that of previous optical switches. As no static power is required to maintain the phase state, it can find promising applications in optical switch matrices and reconfigurable optical spiking neural networks.
© 2022 Optica Publishing Group under the terms of the Optica Open Access Publishing Agreement
Corrections
29 September 2022: A typographical correction was made to the author affiliations. A correction was made to the funding section.
1. Introduction
Traditional electronic computer systems suffer from the bottleneck of computation speed and energy consumption, which cannot satisfy the demand for high-speed and large-scale communication systems [1]. To achieve efficient computing and low energy consumption, high-performance optical signal processing systems such as neuromorphic computing networks are presented [2–4]. Usually, the neuromorphic computing networks approximately consist 1011 artificial neurons [5], and every artificial neuron is designed by an adjustable optical switch. Therefore, the ability of a single optical switch will greatly affect the computation efficiency, and energy consumption of the optical neuromorphic networks. In the practical research process and engineering application, the volatile optical switch using the thermo-optic effect and the electro-optic effect requires continuous external energy to maintain the switching state, which increases the energy consumption and has larger footprints thus restricting the photonics neuromorphic computing networks development [6].
Optical phase-change materials (O-PCMs) have an exceptional optical property contrast between the covalent-bonded amorphous state and the resonant-bonded crystalline state [7], which includes chalcogen-based alloys (e.g., Ge2Sb2Te5 (GST) and Ge2Sb2Se4Te1 (GSST)), antimony-based chalcogenides (e.g., Sb2S3 and Sb2Se3), vanadium dioxide (VO2), and liquid crystal. They are used in reconfigurable photonic applications such as optical switches [8–16], optical routers [17,18], and meta-surface [19–24]. Optical switches based on O-PCMs require no static power to remain in the switch state and have sub-nanosecond modulation speed [25], which is regarded as the key element unit for designing optical neuromorphic networks. Compared with traditional electronic switches, optical switches based on O-PCMs can sustain for 104 seconds after changing the switching state and there is no need to maintain it using static energy. Therefore, it can greatly reduce energy consumption, especially in application scenarios that require frequent switching of switching states. Compared with other phase-change materials (e.g., VO2 and liquid crystal.), GST does not have the fast phase transition speed. For example, the phase transition speed of the VO2, GST, and liquid crystal is nanosecond, several ten nanoseconds and microsecond, respectively. But the optical loss of the VO2 is larger than GST. Therefore, GST is used as the active material because it is far more stable and has the balance between modulation speed and optical loss compared with other PCMs [26–31]. However, the changing of the optical transmittance in different phase states is small due to the weak interaction between the waveguide mode and PCMs when the GST film is straightly sputtered on the top of the silicon ridge waveguide. Therefore, researchers have designed different waveguide structures such as Mach-Zehnder (MZ) switches [32–35], ring coupler switches [36–38], and subwavelength grating (SWG) waveguide [11,26,35] to enhance the GST modulation effect on the optical transmittance. The optical transmittance of these structures using the interference effect and coupler effect is more sensitive to the PCMs’ phase state. But they have large footprints which seriously hinder the large-scale integration of the optical neuromorphic networks. To solve this problem, some reports propose to design special waveguides such as slot-ridge waveguides [16] and surface plasmon waveguides [10–12] to enhance the interaction effect between the waveguide and the PCMs. But the largest optical transmittance difference between the aGST and cGST of these devices is about 40%, this small modulation range prejudices to design the multi-level adjustable optical switches and restricts the development of photonics neuromorphic computing networks.
In order to realize the larger optical transmittance difference between the different phase state of the PCMs, we propose the SWGSR waveguides on the silicon platform and design a nonvolatile multi-level adjustable optical switch based on the phase change material GST. The section of our waveguides is like the slot-ridge waveguides, it can enhance the coupler effect between the optical field and the PCMs. Meanwhile, we design the period structure along the transmitting direction to reserve the energy in the amorphous state and scatter the energy in the cGST. The optical transmittance difference between the aGST state and cGST is about 70%, which is almost twice the optical transmittance difference of other present structures.
2. Waveguide structure and analysis
The 3D structure diagram of our device is shown in Fig. 1(a) and the 2D structure diagram is shown in Fig. 1(b). The original geometrical parameters of the cross-sectional in the slot-ridge waveguide area are shown in Fig. 1(b). The silicon waveguide is fabricated in slot-ridge shape on the silicon-on-insulator (SOI) wafer which height is 340 nm, and the height of the slab (hslab) and ridge (hridge) is 170 nm and 170 nm, respectively. The width of the ridge (wridge) and the slot (wslot) are 300 nm and 100 nm, respectively. The GST film is sputtered deposition on the slot-ridge waveguide, the thickness and width of GST are set to be 50 nm and 25 nm, respectively. The complex refractive indices for the aGST and cGST are 4.6 + 0.18i and 7.2 + 1.9i using spectroscopic ellipsometry [39]. Using three-dimension Finite-Difference Time-Domain (3D FDTD) theory to simulate the output transmission and optical field of the SWGSR waveguide. In the simulation, the mesh size is set to be 20 nm, and the boundary condition is the perfect match layer (PML). The wavelength of the input signal in all simulation models is 1550 nm and we ignore the effect of material dispersion in the simulation calculation process. Because the slot divides the ridge waveguide into two waveguides, the electric fields of the two waveguides are close to each other and overlap in the slot region, which results in strong electric field distribution in the slot region.
Fig. 1. (a) The scheme of the nonvolatile multi-level adjustable optical switch based on a phase change material. (b) The section diagram of the nonvolatile multi-level adjustable optical switch.
The electric field distributions of the traditional ridge waveguide and the slot-ridge waveguide are shown in Fig. 2(a) and Fig. 2(b). Therefore, the slot-ridge waveguide can enhance the interaction between the electric field and the external environment, is widely used in optical sensing, and induce the waveguide nonlinear effect. Due to the enhancing interaction effect between the electric field and the PCMs, researchers designed an on-chip photonic synapse based on slot-ridge waveguides with PCMs for in-memory computing [16]. Compared with the synapse based on the traditional ridge waveguide with GST, the optical transmittance changing of the synapse based on the slot-ridge waveguide with GST is larger, and the optical transmittance difference between the amorphous state (aGST) and crystalline state (cGST) can reach over than 40% by optimizing the slot-ridge waveguide geometrical parameters [12]. Figure 2(c)-(f) show the fundamental mode electric field distributions simulation results of the ridge waveguide and the slot-ridge waveguide when the GST is aGST and cGST. as shown in Fig. 2 (c), there is little electric energy coupling into the GST region when the GST is in the aGST state, the most energy is localized in silicon waveguides, which results in a high value of the optical transmittance. But this phenomenon also contributes to that the changing of the GST phase state cannot affect the optical transmittance observably. Figure 2 (e) shows the fundamental mode electric field distributions simulation results when the GST is in the cGST state. We can find that the maximum intensity of the electric field is localized in the GST region, and the electric field gradually spreads outwards because the GST film is too thin that cannot transmit the complete guide mode. The result that the optical transmittance difference being small when the GST is in a different phase state which restricts the multi-level adjustable capacity. Benefitting from the strong electric field distribution in the slot region, the slot-ridge waveguide can improve the optical transmittance difference when the GST is in a different phase state. We can find that there is a strong interaction between the electric field and the GST film in the slot region in Fig. 2(d) when the GST is in the aGST state, which contributes to that changing of the GST phase state can affect the optical transmittance observably. Meanwhile, due to the reduction of the width of the silicon ridge waveguide, more energy will also leak out from the silicon region, which leads to a small optical transmittance value when the GST is in the cGST, which is shown in Fig. 2(f). We can observe this significant advantage by analyzing the changing of the fundamental mode effective mode index. The effective complex refractive indices (n + ik) of ridge waveguides with aGST and cGST were 3.15 + 0.03i and 4.17 + 1.51i while those of slot-ridge waveguides with aGST and cGST were 3.03 + 0.02i and 3.77 + 1.60i. The difference of the imaginary part Δk refers to the change of the optical transmittance in the waveguides, which were 1.48 and 1.58 for ridge waveguides and slot-ridge waveguides, respectively. This means that the change of the optical transmittance of slot-ridge waveguides is larger than that of ridge waveguides.
Fig. 2. (a) The fundamental mode electric field distribution of the ridge waveguide. (b) The fundamental mode electric field distribution of the slot-ridge waveguide. The electric field intensity of the optical mode profiles of the ridge waveguide with (c) aGST and (e) cGST. The electric field intensity of the optical mode profiles of the slot-ridge waveguide with (d) aGST and (f) cGST. The black dash line is the outline of the cross-section geometry of the SWGSR waveguide.
However, the slot-ridge waveguide gets a larger change of the optical transmittance by improving the difference of the imaginary part Δk of the mode effective index, the difference of the real part Δn has been ignored. In fact, the difference of the real part Δn of the mode effective index has a great impact on the optical transmittance when the light transmits through a periodic structure [40–43]. So, two parameters can affect the waveguide optical transmittance for a periodic structure. Therefore, if we can use both the real and imaginary parts of the mode effective index to modulate the optical transmittance of the waveguide, the optical transmittance difference of the waveguides will be much larger when the GST is in a different phase state. For this purpose, we propose the SWGSR waveguide to let the fundamental mode transmit with low loss when the GST is in the aGST state and with high loss when the GST is in the cGST state, the optical transmittance difference will be larger than using a uniform slot-ridge waveguide. Figure 3 has shown the simulation results of the two structures. We choose the longitudinal central section of the slot-ridge region of different waveguides to analyze the modulation effect on the optical transmittance. Figure 3(a) and (b) are the optical transmit situations in the uniform slot-ridge waveguide when the GST is in the aGST state and the cGST state. We can find that the transmission energy has reduced obviously, but this structure has a large transmission loss when the GST is in the aGST state. Finally, the optical transmittance of the slot-ridge waveguide is 54.1% and 5.1% when the GST is in the aGST state and cGST state, respectively. However, if the slot-ridge waveguide is not uniform but has the period structure along the transmittance direction, the fundamental mode with the different real parts of the mode effective indexs will have a different coupling efficiency with the periodic waveguide structure, which results in different optical transmittance. Therefore, we design the waveguide structure as shown in Fig. 1(a). The SWGSR waveguide consists of a silicon slab layer and a periodic slot-ridge waveguide with a period of 200 nm and a duty cycle (η) of 50%. The optical transmit situations in the SWGSR waveguide when the GST is in the aGST state and the cGST state are shown in Fig. 3(c) and (d). Finally, the optical transmittance of the SWGSR waveguide is 89.4% and 30.3% when the GST is in the aGST state and cGST state, respectively. The difference in the optical transmittance is 49% and 59.1% for the slot-ridge waveguide and SWGSR waveguide, respectively. This means that the structure we proposed has a larger change in the optical transmittance and is more suitable for designing the nonvolatile multi-level adjustable optical switch.
Fig. 3. (a)-(b) The electric field distribution of the uniform slot-ridge waveguide along the propagation direction when the GST is in aGST state and cGST state. (c)-(d) The electric field distribution of the SWGSR waveguide along the propagation direction when the GST is in aGST state and cGST state.
3. SWGSR geometric parameters optimization
The geometry of the waveguide cross-section, such as hslab, hridge, wridge, wslot, will affect the real and imaginary parts of the effective mode index of the fundamental mode, which in turn affects the optical transmittance of the SWGSR waveguide. Meanwhile, the period and η will affect the coupling efficiency between the optical mode and the period structure, which results in the modulation effect on the optical transmittance. So, the optical transmittance of the SWGSR waveguide is related to many parameters, we will determine the optical switch with the best multi-level adjustable capacity by sweeping these parameters in this section. For our SWGSR waveguide structure, the optical transmittance is monitored to analyze and quantify the optimization effect. The change in the optical transmittance ΔT is obtained using the 3D FDTD simulation method to quantify the multi-level adjustable capacity of our optical switch. Both amorphous and crystalline phase states of the GST were employed to attach different weight coefficients to the input optical signal. In the simulation, the initial value of hslab, hridge, wridge, wslot, period and η is 100 nm, 240 nm, 300 nm, 100 nm, 200 nm, and 50%, respectively. And the period numbers are set to 30 initially. It should be noted that the sum of the height of the slab layer and the slot-ridge region is always kept at 340nm.
First, we change the value of the wridge to observe the change of the optical transmittance difference when the GST is in a different phase state. The relationship between the change of the wridge and the optical transmittance difference is shown in Fig. 4(a). We can find that the optical transmittance of the SWGSR waveguide decreases slowly with the increase of wridge, when the GST is in the aGST state and the cGST state. And the optical transmittance difference reaches the maximum when the wridge is 140 nm, which is 67.6%. The reason for the gradual decreasing trend of the optical transmittance difference over the entire modulation range is that the mode field cannot exist stably in the slot-ridge region when the value of wridge is relatively small but exists in the slab layer, which reduces the interaction with the GST material. Therefore, the transmission loss caused by the GST material will reduce. In addition, we can find that the optical transmittance difference is also accompanied by a local increase during its gradual decrease process. It is because the real part of the fundamental mode effective index also changes gradually, and the coupling efficiency between the electric field and the periodic structure also exhibits a periodic change in the process of increasing the wridge. In the next simulations, we set the wridge to 140nm, and Fig. 4(b) shows the relationship between the change of the hslab and the optical transmittance difference. We can find that the optical transmittance of the SWGSR waveguide is very low regardless of whether the GST is in the aGST state or in the cGST state when the hslab is less than 60 nm. It is because the waveguide with a slab height of less than 60 nm cannot restrict the electric field, on the premise that the wridge is 140 nm, which results in the corresponding fundamental mode having a large mode loss. When the height of the slab layer increases so that the waveguide can restrict the electric field, the optical transmittance of the SWGSR waveguide will gradually increase when the GST is in aGST state or cGST state. The increase in the height of the slab layer makes more optical energy distributed in the slab layer, which reduces the interaction between the optical field and the GST, thereby reducing the transmission loss caused by GST materials loss. Finally, the optical transmittance difference reaches the maximum when the wridge is 140 nm and hslab is 130 nm, which is 70.7%. Next, we changed the value of wslot to observe the optical transmittance change of the SWGSR waveguide when the GST is in different phase states. It can be seen from Fig. 4(c) that the optical transmittance difference of the SWGSR waveguide will gradually decrease with the increase of the slot width when the GST is in different phase states. The optical transmittance difference reaches the maximum when the wridge is 140 nm, hslab is 130 nm and wslot is 100 nm, which is 70.7%.
Fig. 4. The relationship between the optical transmittance and the sweeping parameters. The purple arrow points to the ordinate axis used by the corresponding curve.
Since the optical transmittance of the SWGSR waveguide is related to the period of the subwavelength grating, adjusting the period can also significantly affect the optical transmittance difference when the GST is in a different phase, as shown in Fig. 4(d). With the increase of the period, the transmittance of the SWGSR waveguide first decreases and then increases when the GST is in the aGST state. This is because the effective index of the fundamental mode propagating in the waveguide is determined after we determine the value of wridge, hslab, and wslot. And there will be little optical energy that can transmit through the SWGSR waveguide when the relationship between the real part of the mode effective index and the grating period satisfies the Bragg reflection condition. Another detail in Fig. 4(d) can also support this point. The optical transmittance of the SWGSR waveguide with the aGST PCMs film is larger than that with the cGST PCMs film. It is because that the optical loss of the cGST is much larger than that of the aGST. However, the optical transmittance difference of the SWGSR waveguide when the GST is in a different phase state is a minus value when the period is 310 nm. This is precisely because the relationship between the real part of the fundamental mode effective index and the grating period satisfies the Bragg reflection condition when the phase change material is aGST, which contributes to the significant decrease of the optical transmittance. The optical transmittance difference reaches the maximum when the wridge is 140 nm, hslab is 130 nm, wslot is 100 nm and period is 200 nm, which is 70.7%. Likewise, changing η of the subwavelength grating also affects the average refractive index of the grating and then affects the coupling efficiency between the mode field and the grating. Therefore, we can see in Fig. 4(e) that the optical transmittance difference of the SWGSR waveguide also has a crest when the GST is in a different phase. The optical transmittance difference reaches the maximum when the wridge is 140 nm, hslab is 130 nm, wslot is 100 nm, the period is 200 nm, and η is 50%, which is 70.7%. Finally, we simulate and analyze the effect of the period numbers on the optical transmittance of the SWGSR waveguide. Since the optical switch has a certain transmission loss, the optical transmittance of the SWGSR waveguide will gradually decrease when the GST is in the aGST state and the cGST state with the increase of the transmission distance. However, the fundamental modes propagating in the SWGSR waveguide have different transmission losses when the GST is in the aGST state and the cGST state, the rate of decrease in the optical transmittance of the SWGST waveguide is also different when the GST film is in a different phase state. We can see from Fig. 4(f) that the optical transmittance difference reaches the maximum when the wridge is 140 nm, hslab is 130 nm, wslot is 100 nm, the period is 200 nm, η is 50% and period numbers are 30, which is 70.7%.
We choose the structure having the largest optical transmittance difference of the SWGSR waveguide when the GST is in a different phase state for further mode field analysis. The electric field distributions are shown in Fig. 5. Figure 5(a1) and (b1) represent the input electric field distribution of the SWGSR waveguide when the GST is in the aGST state and the cGST state, respectively. Most of the electric field energy is restricted in the slab layer when the GST is in the aGST state, which reduces the interaction of the electric field with the GST film, thereby reducing the transmission loss of the electric field. When the GST is in the cGST state, the electric field energy begins to diffuse from the slab layer to the slot-ridge region, which increases the interaction between the electric field and the GST film, resulting in greater mode loss. Figure 5(a2) and (b2) represent the output electric field distribution of the SWGSR waveguide when the GST is in the aGST state and the cGST state, respectively. The electric field intensity of the optical field does not attenuate greatly after transmitting through the SWGSR waveguide when the GST is in the aGST state. However, the electric field intensity attenuates greatly when the GST is in the cGST state. The electric field transmitting situations in the SWGSR waveguide when the GST is in the aGST state and the cGST state are shown in Fig. 5(a3) and (b3). It can be seen from Fig. 5(a3) and (b3) that the optical transmittance difference of the SWGSR waveguide is obvious when the GST is in different phase states. Therefore, the SWGSR waveguide can transmit the fundamental mode with low loss when the GST is in the aGST state and with high loss when the GST is in the cGST state, and then gets a larger optical transmittance difference than using a uniform slot-ridge waveguide.
Fig. 5. (a1) - (a2) The input and output electric field distributions when the GST is in aGST state. (a3) The electric field distribution along the propagation direction when the GST is in aGST state. (b1) - (b2) The input and output electric field distribution when the GST is in cGST state. (b3) The electric field distribution along the propagation direction when the GST is in cGST state. The black dash line is the outline of the cross-section geometry of the SWGSR waveguide.
We also compare the ridge waveguide, slot-ridge waveguide, SWG waveguide, and SWGSR waveguide with equal waveguide cross-section dimensions, i.e., the silicon width of the ridge waveguide and SWG waveguide is set to 380 nm, which is equal to the sum of twice the ridge width (140 nm) and slot width (100 nm) of the slot waveguide and SWGSR waveguide. Other waveguide geometric parameters are silicon height of 340 nm (ridge, slot-ridge, SWG, SWGSR), slot width of 100 nm (slot, SWGSR), period of 200 nm (SWG, SWGSR), and duty of 50% (SWG, SWGSR). As listed in Tab. 1, one can see that the SWGSR waveguide features the largest optical transmittance difference, which benefits from designing a nonvolatile multi-level adjustable optical switch.
Table 1. Transmittance of different structure
4. Multi-level modulation and image recognition test
To determine the performance of our photonic synapses after the parameter optimization, we performed a recognition test using our photonic synapses based on SWGSR structure with the simulated properties. First, we introduced a crystallization level parameter, which led to multi-level weights needed in the recognition test. The crystallization level resulted from the mixed state of amorphous and crystallized molecules with different proportions during crystallization [44]. Theoretically, the effective permittivity ɛeff(p) is approximately defined by the effective-medium theory:
Fig. 6. The multi-level adjustable optical switch controls the optical transmittance at 65 distinct levels, approximately corresponding to a 6-bit programming resolution.
The height of the box plot represents the optical transmittance at different levels of crystallinity, and the curve represents the fitted curve of the optical transmittance. This curve is given by the function:
where b, k1, k2, and k3 and is estimated to be 0.77028, -0.00566, -2.17638 × 104, and 2.08128 × 106, respectively. In order to examine the advantages of our device in neural network computing, we designed a back-propagation neural network (BPNN) based on this nonvolatile multi-level adjustable optical switch and conducted a test of handwritten digit recognition. The optimized parameters, weights and bias, are obtained by calculating gradient to realize the value of cost function is minimum and the value of accuracy is maximum. In the optimization, the optical transmittance of the switch plays the part of applying weights, which refer to the connection strength between the neurons. In our simulations, the weights of different crystallization levels are given by: Tmin and Tmax represent the minimum and maximum optical transmittance, which are easily estimated from the simulation results in Fig. 6. Therefore, the maximum weight is set to be 1, and the minimum weight is set to be 0. The weight levels are set to 65 to ensure that the optical transmittance difference corresponding to the adjacent weight level exceeds 1%, which ensures the accuracy of identification and reliability in the actual application. And in practical applications, the repeatability of the intermediate states and switch cycles (on/off) are related to the material homogeneity of GST. During the continuous phase transition process, GST may not be able to maintain the material homogeneity as the original state, which affects the material homogeneity of GST. For the recognition test, we use the standard mixed national institute of standards and technology database (MNIST) handwritten digits as the dataset [45]. The size of the perceptron networks is 784 × 300 × 10, which means that the number of neurons in the hidden layers is 300. And the number of training epochs and learning rates are 40 and 0.1, respectively. The optimizer uses stochastic gradient descent (SGD) and the model consists of one flatten layer and two dense layers. The activation function of the first layer of dense is the sigmoid function, and the activation function of the second layer of dense is the softmax function which is used for classification. The initial weights of the two dense layers are initialized with the glorot uniform method and the initial bias is initialized with a zero matrix. Figure 7 (a) is the schematic diagram of our neural network, which is a three-layer perceptron. Each neuron in each layer is connected to all neurons or input data in the previous layer. The recognition accuracies of the neural network are based on the test dataset after the network is trained on the training dataset. The recognition accuracies of the MNIST dataset using the synapses based on the SWGSR waveguide are shown in Fig. 7 (b). For the synapses based on the SWGSR waveguide, the highest accuracy of the MNIST dataset is 96.96% after 40 computational iterations for the test dataset, which provides reliable support for the accuracy of handwritten digital recognition. The information of the hidden layer is shown in Table 2.Fig. 7. (a) The schematic diagram of our neural network, which consists of one flatten layer and two dense layers. (b) the recognition test accuracy of our synapses based on the SWGSR waveguide.
Table 2. Information of the neural network
5. Potential fabrication and limitations of the proposed device
In this section, we furnish a brief presentation on the fabrication technology of the nonvolatile multi-level adjustable optical switch based on the GST. The fabrication process flow was illustrated in Fig. 8. The device was fabricated on an SOI wafer with 340 nm-thick top silicon layer and 2 µm-thick buried silicon oxide layer. Firstly, clean the SOI wafer and add the photoresist coating on the silicon layer. The slot-ridge waveguide was patterned by electron beam lithography (EBL) process and then etched by inductively coupled plasma (ICP) process after the developing process. Then, the SiO2 cladding layer was deposited by plasma enhanced chemical vapor deposition (PECVD). After these processes, a 6 µm-width sputtering window is opened by lithography and wet-etching with buffered oxide etch (BOE). Finally, 50 nm GST was sputtered on the top of the slot-ridge waveguide and followed by 50 nm indium tin oxide (ITO), which prevented the oxidation of the 50 nm GST. Following these processes, the fabrication was finally realized after the liftoff process. And two gold/titanium (Au/Ti) electrode is fabricated on the ITO layer. The phase transition of GST can be provoked by optical or electrical (charge injection or Joule heating) schemes. The electrically induced phase transition scheme has the faster modulation speed and can easily control every GST cell independently. Therefore, our designed nonvolatile optical switch relies on the Joule heating in the ITO layer that is electrically driven Au/Ti electrodes formed on its two ends [46,36]. When the ITO is heated and increases the environment temperature of the GST, GST will absorbate the Joule heat and change its crystalline level.
6. Conclusion
In this paper, we propose a nonvolatile multi-level adjustable optical switch with the SWGSR structure. By utilizing the mode loss of the slot-ridge waveguide and the coupling loss between the mode and the periodic structure, we design a nonvolatile optical switch with a large transmittance adjustable range (77.7% with the aGST state and 7% with the cGST state). The optical transmittance difference between the aGST state and cGST state is about 70%, which is almost twice the optical transmittance difference of other present structures. The simulation results show that our multi-level adjustable optical switch controls the optical transmittance at 65 distinct levels, approximately corresponding to 6-bit programming resolution, with the optical transmittance difference corresponding to the adjacent weight level exceeding 1%, which ensures the accuracy of identification and the reliability in practical applications. The previous nonvolatile multi-level optical switches just have 5-bit programming resolution with the optical transmittance difference corresponding to the adjacent weight level exceeding 1%. In addition, a recognition test is performed using the BPNN, which is based on our nonvolatile multi-level optical switch, and for the MNIST handwritten digits, a 96.96% recognition accuracy is achieved. This work has great potential for synthesizing photonic synapses with more weight levels and higher resolution between adjacent weight levels, which benefits realizing the higher efficient in-memory computing neuromorphic networks in the future.
Funding
National Key Research and Development Program of China (2019YFB2203604); National Natural Science Foundation of China (NSFC) (62125503); Key R&D Program of Hubei Province of China (2020BAB001, 2020BAA007); Key R&D Program of Guangdong Province (2018B030325002); Science and Technology Innovation Commission of Shenzhen (JCYJ20200109114018750); Fundamental Research Funds for the Central Universities (2019kfyRCPY037).
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. M. Hochberg and T. Baehr-Jones, “Towards fabless silicon photonics,” Nat. Photonics 4(8), 492–494 (2010). [CrossRef]
2. Y. Shen, C. N. Harris, S. Skirlo, M. Prabhu, T. Baehr-Jones, M. Hochberg, X. Sun, S. Zhao, H. Larochelle, D. Englund, and M. Soljačić, “Deep Learning with Coherent Nanophotonic Circuits,” Nat. Photonics 11(7), 441–446 (2017). [CrossRef]
3. J. Feldmann, N. Youngblood, C. D. Wright, H. Bhaskaran, and W. H. Pernice, “All-optical spiking neurosynaptic networks with self-learning capabilities,” Nature 569(7755), 208–214 (2019). [CrossRef]
4. J. Xiang, Y. Zhang, Y. Zhao, X. Guo, and Y. Su, “All-optical silicon microring spiking neuron,” Photonics Res. 10(4), 939–946 (2022). [CrossRef]
5. J. Hasler and B. Marr, “Finding a roadmap to achieve large neuromorphic hardware systems,” Front. Neurosci. 7(118), 1–29 (2013). [CrossRef]
6. J. Zheng, Z. Fang, C. Wu, S. Zhu, P. Xu, J. K. Doylend, S. Deshmukh, E. Pop, S. Dunham, M. Li, and A. Majumdar, “Nonvolatile electrically reconfigurable integrated photonic switch enabled by a silicon PIN diode heater,” Adv. Mater. 32(31), 2001218 (2020). [CrossRef]
7. K. Shportko, S. Kremers, M. Woda, D. Lencer, J. Robertson, and M. Wuttig, “Resonant bonding in crystalline phase-change materials,” Nat. Mater. 7(8), 653–658 (2008). [CrossRef]
8. P. Markov, R. E. Marvel, H. J. Conley, K. J. Miller, R. F. Haglund Jr., and S. M. Weiss, “Optically Monitored Electrical Switching in VO2,” ACS Photonics 2(8), 1175–1182 (2015). [CrossRef]
9. Y. Jung, H. Han, A. Sharma, J. Jeong, S. S. P. Parkin, and J. K. S. Poon, “Integrated Hybrid VO2−Silicon Optical Memory,” ACS Photonics 9(1), 217–223 (2022). [CrossRef]
10. R. Miquel, S. E. Robert, Q. Romain, P. Valerio, and R. Jan, “Active control of surface plasmon waveguides with a phase change material,” ACS Photonics 2(6), 669–674 (2015). [CrossRef]
11. E. Gemo, S. G.-C. Carrillo, C. R. De Galarreta, A. Baldycheva, H. Hayat, N. Youngblood, H. Bhaskaran, W. H. P. Pernice, and C. D. Wright, “Plasmonically-enhanced all-optical integrated phase-change memory,” Opt. Express 27(17), 24724–24737 (2019). [CrossRef]
12. E. Gemo, J. Faneca, S. G.-C. Carrillo, A. Baldycheva, W. H. P. Pernice, H. Bhaskaran, and C. D. Wright, “A plasmonically enhanced route to faster and more energy-efficient phase-change integrated photonic memory and computing devices,” J. Appl. Phys. 129(11), 110902 (2021). [CrossRef]
13. Z. Yifei, C. B. Jeffrey, L. Junying, L. Huashan, D. Qingyang, Y. Anupama, Z. Si, S. Y. Mikhail, F. Zhuoran, Z. Huikai, R. Christopher, R. Paul, B. Bridget, R. Carlos, L. Hongtao, K. Myungkoo, G. Tian, W. Jamie, L. Vladimir, R. Kathleen, and H. Juejun, “Broadband transparent optical phase change materials for high-performance nonvolatile photonics,” Nat. Commun. 10(1), 4278 (2019). [CrossRef]
14. W. Changming, Y. Heshan, L. Huan, Z. Xiaohang, T. Ichiro, and L. Mo, “Low-loss integrated photonic switch using subwavelength patterned phase change material,” ACS Photonics 6(1), 87–92 (2019). [CrossRef]
15. W. Changming, Y. Heshan, L. Seokhyeong, P. Ruoming, T. Ichiro, and L. Mo, “Programmable phase-change metasurfaces on waveguides for multimode photonic convolutional neural network,” Nat. Commun. 12(1), 96 (2021). [CrossRef]
16. Z. Huan, H. Beiju, Z. Zanyun, C. Chuantong, Z. Zan, Z. Hengjie, C. Run, and C. Hongda, “On-chip photonic synapses based on slot-ridge waveguides with PCMs for in-memory computing,” IEEE Photonics J. 13(2), 1–13 (2021). [CrossRef]
17. X. Peipeng, Z. Jiajiu, D. K. Jonathan, and M. Arka, “Low-loss and broadband nonvolatile phase-change directional coupler switches,” ACS Photonics 6(2), 553–557 (2019). [CrossRef]
18. Y. Xing, N. S. Muhammad, Y. Wei, Z. Fengang, L. Liangjun, C. Jianping, and Z. Linjie, “Phase change material enabled 2 ( 2 silicon nonvolatile optical switch,” Opt. Lett. 46(17), 4224–4227 (2021). [CrossRef]
19. S. H. Badri, H. Soofi, and S. SaeidNahaei, “Thermally reconfigurable extraordinary terahertz transmission using vanadium dioxide,” J. Opt. Soc. Am. B 39(6), 1614–1621 (2022). [CrossRef]
20. H. Peiman, C. D. Wright, and H. Bhaskaran, “An optoelectronic framework enabled by low-dimensional phase-change films,” Nature 511(7508), 206–211 (2014). [CrossRef]
21. C. R. de Galarreta, A. M. Alexeev, Y.-Y. Au, M. Lopez-Garcia, M. Klemm, M. Cryan, J. Bertolotti, and C. D. Wright, “Nonvolatile reconfigurable phase-change metadevices for beam steering in the near infrared,” Adv. Funct. Mater. 28(10), 1704993 (2018). [CrossRef]
22. S. G.-C. Carrillo, L. Trimby, Y.-Y. Au, V. K. Nagareddy, G. Rodriguez-Hernandez, P. Hosseini, C. Ríos, H. Bhaskaran, and C. D. Wright, “A nonvolatile phase-change metamaterial color display,” Adv. Opt. Mater. 7(18), 1801782 (2019). [CrossRef]
23. S. Abdollahramezani, O. Hemmatyar, M. Taghinejad, H. Taghinejad, A. Krasnok, A. A. Eftekhar, C. Teichrib, S. Deshmukh, M. A. El-Sayed, E. Pop, M. Wuttig, A. Alù, W. Cai, and A. Adibi, “Electrically driven reprogrammable phase-change metasurface reaching 80% efficiency,” Nat. Commun. 13(1), 1696 (2022). [CrossRef]
24. S. Abdollahramezani, O. Hemmatyar, M. Taghinejad, H. Taghinejad, Y. Kiarashinejad, M. Zandehshahvar, T. Fan, S. Deshmukh, A. A. Eftekhar, W. Cai, E. Pop, M. A. El-Sayed, and A. Adibi, “Dynamic Hybrid Metasurfaces,” Nano Lett. 21(3), 1238–1245 (2021). [CrossRef]
25. M. Wuttig, H. Bhaskaran, and T. Taubner, “Phase-change materials for non-volatile photonic applications,” Nat. Photonics 11(8), 465–476 (2017). [CrossRef]
26. S. H. Badri, M. M. Gilarlue, S. G. Farkoush, and S.-B. Rhee, “Reconfigurable bandpass optical filters based on subwavelength grating waveguides with a Ge2Sb2Te5 cavity,” J. Opt. Soc. Am. B 38(4), 1283–1289 (2021). [CrossRef]
27. S. H. Badri and S. G. Farkoush, “Subwavelength grating waveguide filter based on cladding modulation with a phase-change material grating,” Appl. Opt. 60(10), 2803–2810 (2021). [CrossRef]
28. H. Xu, D. Dai, and Y. Shi, “Silicon integrated nanophotonic devices for on-chip multi-mode interconnects,” Appl. Sci. 10(18), 6365 (2020). [CrossRef]
29. J. Wang, Y. Xuan, M. Qi, H. Huang, Y. Li, M. Li, X. Chen, Z. Sheng, A. Wu, W. Li, X. Wang, S. Zou, and F. Gan, “Broadband and fabrication-tolerant on-chip scalable mode-division multiplexing based on mode-evolution counter-tapered couplers,” Opt. Lett. 40(9), 1956–1959 (2015). [CrossRef]
30. T. H. Pan and S. Y. Tseng, “Short and robust silicon mode (de)multiplexers using shortcuts to adiabaticity,” Opt. Express 23(8), 10405–10412 (2015). [CrossRef]
31. D. Guo and T. Chu, “Silicon mode (de)multiplexers with parameters optimized using shortcuts to adiabaticity,” Opt. Express 25(8), 9160–9170 (2017). [CrossRef]
32. H. Y. Zhang, L. J. Zhou, J. Xu, L. J. Lu, J. P. Chen, and B. M. A. Rahman, “All-optical non-volatile tuning of an AMZI-coupled ring resonator with GST phase-change material,” Opt. Lett. 43(22), 5539–5542 (2018). [CrossRef]
33. N. Dhingra, J. Song, G. J. Saxena, and E. K. Sharma, Member and B. M. A. Rahman, “Design of a compact low-loss phase shifter based on optical phase change material,” IEEE Photonics Technol. Lett. 31(21), 1757–1760 (2019). [CrossRef]
34. J. Faneca, T. D. Bucio, F. Y. Gardes, and A. Baldycheva, “O-band N-rich silicon nitride MZI based on GST,” Appl. Phys. Lett. 116(9), 093502 (2020). [CrossRef]
35. Y. Wang, J. Zhang, H. Jin, and P. Xu, “Reconfigurable and dual-polarization Bragg grating filter with phase change materials,” Appl. Opt. 60(31), 9989–9993 (2021). [CrossRef]
36. H. Taghinejed, S. Abdollahramezani, A. A. Eftekhar, T. Fan, A. H. Hosseinnia, O. Hemmatyar, A. E. Dorche, A. Gallmon, and A. Adibi, “ITO-based microheaters for reversible multi-stage switching of phase-change materials: towards miniaturized beyond-binary reconfigurable integrated photonics,” Opt. Express 29(13), 20449–20462 (2021). [CrossRef]
37. J. Zheng, A. Khanolkar, P. Xu, S. Colburn, S. Deshmukh, J. Myers, J. Frantz, E. Pop, J. Hendrickson, J. Doylend, N. Boechler, and A. Majumdar, “GST-on-silicon hybrid nanophotonic integrated circuits: a non-volatile quasi-continuously reprogrammable platform,” Opt. Express 8(6), 1551–1561 (2018). [CrossRef]
38. N. Youngblood, C. Ríos, E. Gemo, J. Feldmann, Z. Cheng, A. Baldycheva, W. H. P. Pernice, C. D. Wright, and H. Bhaskaran, “Tunable Volatility of Ge2Sb2Te5 in Integrated Photonics,” Adv. Funct. Mater. 29(11), 1807571 (2019). [CrossRef]
39. C. Zhang, M. Zhang, Y. Xie, Y. Shi, R. Kumar, R. R. Panepucci, and D. Dai, “Wavelength-selective 2 × 2 optical switch based on a Ge2Sb2Te5-assisted microring,” Photonics Res. 8(7), 1171–1176 (2020). [CrossRef]
40. S. Y. Kim, S. J. Kim, H. Seo, and M. R. Kim, “Variation of the complex refractive indices with Sb-addition in Ge-Sb-Te alloy and their wavelength dependence,” Opt. Data Storage 3401(98), 112–115 (1998). [CrossRef]
41. Z. Ruan, L. Shen, S. Zheng, and J. Wang, “Subwavelength grating slot (SWGS) waveguide on silicon platform,” Opt. Express 25(15), 18250–18264 (2017). [CrossRef]
42. P. Cheben, R. Halir, J. H. Schmid, H. A. Atwater, and D. R. Smith, “Subwavelength integrated photonics,” Nature 560(7720), 565–572 (2018). [CrossRef]
43. H. Wu, C. Li, L. Song, H.-K. Tsang, J. E. Bowers, and D. Dai, “Ultra-sharp multimode waveguide bends with subwavelength Gratings,” Laser Photonics Rev. 13(2), 1800119 (2019). [CrossRef]
44. Y. Chen, X. Li, Y. Sonnefraud, A. I. Fernández-Domínguez, X. Luo, M. Hong, and S. A. Maier, “Engineering the phase front of light with phase-change material based planar lenses,” Sci. Rep. 5(1), 8660 (2015). [CrossRef]
45. Y. LeCun, C. Cortes, and C. J. Burges, “The MNIST database of handwritten digits,” 1998. [Online]. Available: http://yann.lecun.com/exdb/mnist.
46. O. Hemmatyar, S. Abdollahramezani, I. Zeimpekis, S. Lepeshov, A. Krasnok, A. I. Khan, K. M. Neilson, C. Teichrib, T. Brown, E. Pop, D. W. Hewak, M. Wuttig, A. Alu, O. L. Muskens, and A. Adibi, arXiv preprint arXiv:2107.12159 (2021).
