1. Introduction

The von Neumann architecture remains the foundational paradigm for computer design, characterized by its generality, programmability, and scalability, allowing it to apply to various applications. However, it faces bottlenecks in data transfer speed, parallel computing, and low power consumption [1]. Photonic Integrated Circuits (PICs) represent a novel computer technology that utilizes photons for information transmission and processing. It promises to overcome the bottlenecks in von Neumann's architecture, enabling efficient computation and low energy consumption in a range of optical signal processing systems. Furthermore, Silicon Photonics offers high speed, low energy consumption, large bandwidth, and compatibility with traditional Complementary Metal-Oxide-Semiconductor (CMOS) manufacturing processes, making it one of the most promising PIC technologies [2,3]. Silicon photonics has seen extensive research progress in areas such as integrated optical switches [4], programmable photonic circuits [5–7], and photonic artificial neural networks [8,9].

Integrated optical switches and photonic memory play a crucial role in photonic artificial neural networks. Each artificial neuron utilizes an optical switch as its basic building block [10]. Therefore, the performance of photonic artificial neural networks can be significantly influenced by the power consumption, switching speed, stability, and other characteristics of optical switches [4]. However, most of these silicon devices currently rely on thermal-optic and electro-optic effects to modulate the switch states. Since both effects require a continuous voltage to maintain the switch state, these devices tend to have higher energy consumption [11]. Additionally, it is essential to note that these devices require a relatively large physical space to achieve a wide range of refractive index modulation [12], further elevating the challenge of achieving low power consumption. Therefore, maintaining low power consumption and achieving a compact design for developing photonic neural networks becomes increasingly prominent.

Combining plasmonic slot waveguide (PSW) with phase-change materials (PCMs) effectively addresses the abovementioned issues. PSW consists of two metal blocks with a narrow gap using surface plasmon excitations as signal carriers. This design allows for the optical field localization within a sub-wavelength range, surpassing the diffraction limit while enhancing the interaction between light and matter. Consequently, it significantly reduces the size of photonic integrated circuits and enhances device compactness [13,14]. In recent years, PCMs have found widespread applications in both electronic and photonic fields [15–17]. When rapidly switching between two stable states, PCMs exhibits significant changes in its electrical and optical properties. PCM-based on-chip optical modulators has made significant progress in various studies. Yang et al. proposed a Mach-Zehnder interferometer (MZI) embedded with Sb₂Se₃ phase-change material in an Al₂O₃ layer, achieving outstanding extinction ratios of 20 dB [18]. However, this structure faces difficulties in achieving seamless integration into optical neural networks due to the limitations imposed by the large device size. Meng et al. successfully developed a multi-state electrically programmable low-loss, non-volatile photonic memory using Ge₂Sb₂Se₅, demonstrating advantages such as low insertion loss and broadband transparency [19]. The relatively low refractive index difference Δk of Ge₂Sb₂Se₅, only 0.12 between its crystalline and amorphous states, restricts its ability to achieve significant optical transmittance differences, thereby hindering the attainment of a broader optical weight dynamic range crucial for photonic neural computing. An optical modulator based on an Sb₂S₃-coated silicon photonics platform has performance metrics such as low loss and high extinction ratios [20]. However, its optical transmittance difference seems relatively low. GST is one of the most widely used PCMs for photonic devices. GST demonstrates excellent electrical and optical properties contrast between its aGST and cGST over a broad wavelength range. It exhibits very short switching times, zero static power consumption, and a small footprint, facilitating the development of ultra-compact devices. However, directly sputtering GST thin films onto the top of silicon waveguides results in minimal optical transmission changes between different phases. This is due to the relatively weak interaction between waveguide modes and PCMs [21,22]. Recently, devices combining PCMs with surface plasmon polaritons have been proposed, such as on-chip all-optical memory [23–26] and optical switches [27–29]. Although surface plasmon waveguides can enhance the interaction between waveguides and PCMs, the maximum light transmittance difference between aGST and cGST in these devices is only about 60% [27]. This limited modulation range poses challenges in designing multi-level adjustable optical switches and hinders the further development of photonic neural computing networks.

In this study, to enhance the light transmittance difference of GST between different phases, we designed a periodic segmented GST structure based on PSW to construct a non-volatile multi-level adjustable optical switch. Initially, the plasmonic effect actively enhanced the coupling between the optical field and the GST. Subsequently, a periodic structure along the transmission direction was designed to preserve energy in the aGST and disperse energy in the cGST. This design results in an optical transmittance difference of approximately 85%, achieving a low insertion loss of 0.5dB. The low insertion loss is crucial for applications involving synaptic weights, particularly in large-scale photonic neural network applications. Compared to other recent studies combining PSW with GST [27–29], our device achieves optimal performance with a lower insertion loss and higher transmittance difference at a wavelength of 1550nm. Our crucial innovation lies in the significant reduction of the footprint of GST by utilizing a PSW structure. By employing a periodic segmented structure, we have successfully enhanced the optical transmittance difference. This innovation enables our device to maintain low insertion loss and high extinction ratio while possessing a more extensive optical weight dynamic range, providing vital support for applications such as photonic neural computing.

2. Device design

2.1 Confinement factor density of plasmonic slot waveguide

Silicon waveguides are a highly effective type of dielectric waveguide for transmitting information with minimal loss. However, when operating in the visible and infrared wavelengths, metals can cause significant absorption losses, leading to increased losses overall. This can significantly influence the transmittance contrast of optical switch devices. Hence, emphasis is placed on the geometric structure of PSW in this section to optimize the balance between optical field confinement and transmission losses. The proposed PSW is realized by employing two finite-height metal films placed over SiO₂ with a narrow gap separating them, as illustrated in Fig. 1(a). As the electromagnetic field cannot penetrate the metal, the waveguide mode is strictly confined within the gap between the two metal films. The impact of waveguide dimensions on its performance was studied using the finite-difference software Lumerical MODE Solutions. Au, Si, and SiO₂ using the Palik model, which comprehensively and accurately describes the refractive indices of materials like Au [30]. The heights of the two metals were fixed at different values (70nm, 120nm, 170nm, 220nm), and the gap distance (w) between the two metals was gradually increased from 30nm to 200nm. The relationship between effective refractive index (n_eff) and gap distance w is shown in Fig. 1(b). For all four different metal film thicknesses, n_eff decreases as the gap w increases, gradually approaching the value of a Silicon on Insulator (SOI) waveguide unaffected by plasmonic materials. The narrower-gap PSW has a larger n_eff, indicating that, compared to traditional SOI waveguides, the PSW provides greater optical confinement. A key parameter in the description of optical confinement is the confinement factor (Г), which represents the ratio of power in the gap region (highlighted by the red dashed line area in Fig. 1(a))to the total power in the entire waveguide region [31]:

As shown in Fig. 1(c), power is highly confined within the gap for a narrow slot, and with an increase in metal film thickness, the power confinement within the gap also further increases. However, the high-power confinement in the small gap region comes at the cost of large propagation losses (see Fig. 1(d)). Therefore, there is a trade-off between the confinement factor and propagation losses in the PSW. To better optimize the transmittance contrast performance of the optical switch, we fixed the metal film thickness at 220nm. We later determined the optimal size of the gap w by comparing transmittance differences.

 

Fig. 1. PSW performance analysis with different metal film thicknesses and gaps. (a) Schematic diagram of the PSW. (b) Variation of effective refractive index (n_eff) with gap width w. (c) Relationship between the confinement factor and gap width w. (d) Loss (in dB/µm) as a function of gap width w.

Download Full Size | PDF

2.2 Device structure and working principle

Figure 2(a) depicts the 3D structural design of the non-volatile multi-level adjustable optical switch device combining segmented GST and PSW. Figures 2(b) and (c) present the top and cross-sectional views of the plasmonic slot region with the original geometric parameters. The entire structure is based on a 350nm SOI platform. A silicon-gold taper structure with a small gap (L_gap= 50nm and W_gap= 40nm) is employed to enhance the coupling efficiency between the silicon waveguide and the PSW. The gold with a height of 220nm serves as the device electrode to trigger GST phase change. The width and height of the strip silicon waveguide are fixed at 500nm and 220nm, respectively. The taper length (L_taper) and tip width (W_tip) are set to 600nm and 110nm, respectively. The PSW length (L_slot) and width (W_slot) are 400nm and 130nm, respectively. The GST film with a thickness of 30nm is sputter-deposited on the PSW. The GST segment period and fill factor are optimized to be 100nm and 60%, respectively. In the aGST and cGST, the refractive indices are set as n_amor = 3.9027 + 0.0055i and n_cry= 6.0769 + 0.9040i [32]. Additionally, to ensure stable operation, a protective layer (5nm of silicon dioxide) is coated on top to prevent oxidation of the GST material. The geometric parameters of the silicon cone waveguide and metal funnel structure mentioned above have been optimized using the 3D finite-difference time-domain (FDTD) method. By employing Lumerical frequency domain field monitors, we computed the transmittance difference (ΔT) between cGST and aGST, enabling analysis and quantitative assessment of the optimization effects. Ultimately, we achieved a maximum optical transmittance difference of 85%. A detailed optimization process is provided in the appendix (see Supplement 1).

 

Fig. 2. (a)Schematic diagram of the multi-level adjustable optical switch device combining PSW and segmented GST. (b, c) Top and cross-sectional views of the plasmonic slot region. (d, e) Electric field intensity distribution in the PSW region for aGST and cGST.

Download Full Size | PDF

Two gold electrodes form a PSW in the nanoscale gap region. Broadband modulation of light transmission is achieved through coupling with a silicon waveguide, leveraging the significant refractive index contrast of GST in its aGST and cGST, and employing a segmented structure to enhance mode coupling. The reduction in mode volume enhances the electric field within the PSW while reducing the GST footprint. Figures 2(d) and 2(e) illustrate the electric field distribution of GST in the aGST and cGST within the PSW at an input signal wavelength of 1550 nm, respectively. When GST is in the aGST, the electric field intensity in the slot waveguide is significantly higher than in the cGST due to the lower absorption loss in aGST. This results in a higher optical transmittance. Following the crystallization of GST, the elevated refractive index results in an augmented electromagnetic field penetration into the metallic region, consequently diminishing the horizontal propagation of optical signals. Hence, the modulation level of optical transmission in the device will be collectively influenced by field penetration into the metallic region, absorption in GST under two-phase states, and the mode coupling efficiency between the silicon waveguide and the PSW.

3. Performance comparison between segmented and non-segmented structures

In this section, we compare the performance of the designed segmented GST device structure with the traditional non-segmented GST structure. Simultaneously, we elucidated the mechanism of improving modulation depth using GST-segmented structures. The difference in the imaginary part of the refractive index Δk between aGST and cGST is one of the reasons for the variation in optical transmittance. When light passes through periodic structures, the difference in the real part of the effective refractive index (Δn_eff) of the mode significantly also affects the light transmittance [10,33]. According to the Effective Medium Theory (EMT), the equivalent refractive index of GST-segmented structures can be calculated, where a segmented structure can be approximated as a continuous structure made of equivalent birefringent materials [33]:

The difference in effective refractive index leads to the mode mismatch and dispersion when optical propagates in the waveguide, thereby affecting the transmittance [35]. In the aGST, the segmented structure exhibits a smaller Δn_eff compared to the non-segmented structure when paired with the PSW. This reduction in Δn_eff contributes to mitigating mode coupling losses, thus enhancing transmission in the aGST. Conversely, in the cGST, though the segmented structure still demonstrates a smaller Δn_eff with the PSW compared to the non-segmented structure, the intrinsic high absorption loss of the cGST maintains similarly low transmittance levels for both segmented and non-segmented structures. Therefore, the segmented structure demonstrates a larger transmittance difference between the two states. The optical transmittance difference of the segmented structure reaches 85% at 1550 nm, as shown in Fig. 3(a), while the unsegmented structure only exhibits 63%. We conducted a comparative analysis of the transmittance variations at a wavelength of 1550nm for both the unsegmented structure and the segmented structures varying periods ranging from 2 to 5. Throughout this process, we ensured consistency in the overall volume and segment gap of the GST. As the GST segment period increases, the Δneff between the GST hybrid waveguide and PSW gradually decreases, resulting in an enhanced transmittance in the aGST. However, in the cGST, the transmittance is influenced not only by Δneff but also by the extent of light absorption due to the significant imaginary part of the effective refractive index (k_eff). It is found that a favorable balance between Δn_eff and k_eff enables a significant transmission difference for the 4-period segment, as shown in Fig. 3(a). Therefore, we opted for a 4-period segment in this study. Subsequently, we compared the variation in optical transmittance difference of segmented and unsegmented GST structures as the length of the PSW increased from 300 nm to 1200 nm. As shown in Fig. 3(b), despite the increase in inherent losses in the metal with the lengthening of the PSW, the transmittance difference with segmented structures was significantly higher than with unsegmented ones for the same length of the slot waveguide. Therefore, using segmented structures can significantly improve the transmittance difference rather than just relying on the length of the slot waveguide.

 

Fig. 3. (a)Optical transmittance differences under different segment periods. (b) The transmittance difference of the segmented and unsegmented structures under different lengths of PSW. (c) The electric field intensity distribution of the segmented and non-segmented structures. (d) The extinction ratio and (e) insertion loss for the segmented and non-segmented switch structures.

Download Full Size | PDF

We analyzed the mode fields of the device in both cGST and aGST, respectively. Figure 3(c) illustrates the electric field distribution in the xy plane of the device. Here, a high-intensity electric field is evident in the small gap between the tapered waveguide and the metal funnel, confirming that introducing a small gap between the tapered waveguide and the metal funnel can significantly enhance the coupling efficiency of light from the dielectric waveguide to the PSW [36,37]. Moreover, by comparing the electric field intensity of the output waveguides, regardless of segmented or non-segmented GST, both structures exhibit significant differences in optical transmittance between the aGST and cGST. Specifically, when GST is in the aGST, the transmission loss of the device is lower, while in the cGST, it incurs higher losses when transferring the fundamental mode. However, it is worth noting that when GST is in the aGST, the segmented structure's output waveguide exhibits significantly higher electric field intensity than the non-segmented GST structure. The energy confinement in the metal slot is also stronger in the segmented structure, leading to lower reflection losses than the non-segmented GST structure. Therefore, adopting a segmented GST structure in the PSW can achieve a larger optical transmittance difference than the non-segmented structure.

The extinction ratio (ER) and insertion loss (IL) of the two structures are defined as [38]:

4. Phase change thermal simulation

A three-dimensional finite element model was constructed to investigate the spatial and temporal temperature response during the phase-change process using COMSOL Multiphysics. Commonly employed methods for achieving the phase transition of PCMs include optically and electrically induced phase-change methods. In addressing challenges associated with optical thermal heating, particularly in large-scale integration, the choice was made to utilize Joule heating, specifically electrically induced threshold switching, to trigger the phase change of GST. The metal facilitated the plasmonic effect and the provision of two electrodes for Joule heating, contributing to a more compact device. The simulation process predicted the temperature distribution using the Joule heat physics interface within the electromagnetics framework, incorporating the Solid Heat Transfer and Electric Current modules. In the electric current module, electrical insulating boundary conditions were applied to all external boundaries, excluding the two metal electrodes (i.e., potential and ground). The SiO₂ bottom was maintained at a constant temperature T = 293K within the Solid Heat Transfer module. A convective heat flux boundary condition with a heat transfer coefficient of 5 W/(m²·K) was applied to the device's surface, while adiabatic boundary conditions were employed for all other external boundaries. The basic properties of the materials utilized in the simulation can be found in Table 1.

Table 1. Thermal and electrical properties of the materials used in the FEM simulation.

View Table

Figure 4(a) illustrates the simulated temperature distribution at the bottom switch unit after applying a 10ns-12V amorphization pulse. It is observed that the high thermal conductivity of the metal electrode keeps its temperature primarily at room temperature, and the temperature rise is mainly confined to the GST within the PSW. Additionally, temperature distribution curves at the interface of segmented GST are plotted along the red dashed line for different pulse widths. Clearly, with the gradual increase in pulse width, the temperature distribution curves along the PSW also proportionally rise. The segmented structure establishes temperature gradients within each segment, playing a crucial role in multilevel phase changes. Specifically, when the amorphization pulse width is 6ns, the peak temperature of the middle segment of the PCM just reaches the melting point of GST, indicating that amorphization occurs in the middle of each segmented GST. Subsequently, as the pulse energy and peak temperature increase gradually, amorphization gradually extends from the middle of GST towards both ends. Eventually, when the temperature of the entire GST segment is slightly above the melting point, complete amorphization can be achieved. The thermal distribution of segmented GST at the center cross-section is illustrated separately at the end of crystallization and re-amorphization pulses in Figs. 4(b) and 4(c). From the temperature distribution curves, it can be observed that heat is primarily generated at the bottom of the PSW. During this process, the phase transition from aGST to cGST occurs only at the bottom of the GST film, while the other parts remain in the aGST state. A 30 ns-7 V pulse was applied to metal electrodes to simulate the crystallization process of GST, as shown in Fig. 4(d). The applied pulse voltage is sufficient to raise the temperature of the entire GST above the crystallization temperature (250°C) and still below the melting temperature of GST (630°C). Next, we apply a 10ns-long 12V voltage pulse to the metal electrodes to study the effect of the amorphization pulse on GST phase transition, as depicted in Fig. 4(e). During this process, GST can be heated above the melting temperature and then quenched to an aGST state by a cooling rate of >1°C/ns, emphasizing the importance of maintaining a fast cooling rate for the amorphization process [40].

 

Fig. 4. (a) Temperature distribution simulation using Joule heating, consisting of four segments, each with a 60nm-long GST. The top panel shows the longitudinal temperature distribution along the bottom GST of the PSW when simulating pulse width variations. (b, c) Simulated temperature distribution along the yz plane of the PSW at the end of (d) fully crystallized pulse (i.e., 40ns) and (e) amorphized pulse (i.e., 10ns).

Download Full Size | PDF

5. Multilevel modulation and image recognition testing

To validate the performance of the photon synapse after parameter optimization, we adopted a structure combining PSW and GST and simulated the performance of the photon synapse in a digital recognition task. In this study, we introduced parameters at the crystallization level to meet the requirements for multi-level weight updates in handwritten digit recognition tasks. The crystallization level refers to the incomplete phase change of the PCMs during crystallization, resulting in the coexistence of aGST and cGST molecules, thereby triggering specific properties [41]. We also determined the effective dielectric constant through theoretical definitions [42]:

 

Fig. 5. (a) Multi-level adjustable light transmission versus crystallinity. (b) Structure of simulated ANN for handwritten digit recognition task, consisting of an input layer with 748 neurons, a hidden layer with 300 neurons, and an output layer with 10 neurons, with input images of $28 \times 28$ pixels. (c) Synaptic recognition test accuracy combining PSW and segmented GST structure. (d) Predictive precision before (d) and after 40 iterations of training (e) of the model.

Download Full Size | PDF

The parameters ${y_0}$, ${A_1}$, and the nonlinear fitting factor $- 1/{t_1}$ are estimated to be 1.51018, -0.58086, and 0.956, respectively. In neural networks, the optical transmittance based on the combination of GST and PSW plays the role of applying weights in the application. The weight values determine the strength of signal transmission during the connection process. To reduce errors in synaptic weight calculations caused by nonlinear factors and further improve performance, we use both positive and negative synaptic weights in this recognition test. Therefore, the weight between different neurons is defined as:

The parameter r is the scaling factor, and ${T_0}$ is an adjustable value within the dynamic range between the minimum and maximum transmittance of the optical switch. According to the transmittance of the optical switch in Fig. 5(a), ${T_0}$ is defined as the middle value (0.466) between the maximum and minimum transmittance. Thus, the maximum weight W_max is set to 1, and the minimum weight W_min is set to -1. Figure 5(b) shows the neural network architecture for performing the handwritten digit task. The recognition task is carried out by a fully connected feedforward neural network, which is a three-layer perceptron based on the backpropagation algorithm, consisting of an input layer with 748 neurons, a hidden layer with 300 neurons, and an output layer with 10 neurons. The input images are from the MNIST dataset [43]. Each pixel corresponds to an input neuron. The number of training epochs and learning rate are set to 40 and 0.1, respectively. The activation function for the first dense layer is the sigmoid function, and the activation function for the second dense layer is the softmax function for classification. After training the network on the training dataset, the recognition accuracy of the neural network is evaluated on the test dataset. The accuracy using the dataset with the photon synapse combined with GST and PSW is shown in Fig. 5(c). Finally, utilizing the photon synapse combined with segmented GST and PSW, after 40 iterations on the test dataset, the accuracy reaches up to 95%. To further verify the precision of the model, confusion matrices before and after training are plotted in Figs. 5(d) and (e), demonstrating a precision greater than 94% after training simulation. The high accuracy and precision indicate the excellent performance of the ANN combined with segmented GST and PSW, confirming the potential of constructing a neural network with an all-PCM.

6. Device fabrication and tolerance

6.1 Device fabrication

This section will briefly introduce the manufacturing technology of a non-volatile multi-level reconfigurable optical switch that combines PSW and segmented GST. The manufacturing process flowchart is shown in Fig. 6. Firstly, the photoresist is coated on the SOI platform, and high-precision pattern transfer is achieved through electron beam lithography (EBL). Subsequently, after the development process, silicon conical waveguides are formed through inductively coupled plasma (ICP—RIE) etching. Using a lift-off process to remove excess metal and eventually obtain a 200nm-thick metal film is a feasible fabrication process. The electron beam resist is spin-coated onto the entire sample and patterned using EBL to define windows for electron beam evaporation deposition of a 220nm-thick metal layer, achieving a high-resolution configuration of the PSW. This process selectively covers specific surface areas and enables patterning on minimal features, typically ranging from submicron to nanometer dimensions [44]. Several reports have demonstrated that the resolution of EBL can reach below 10 nm [45], enabling the achievement of minute gaps between silicon and metal. It is worth noting that prior to metal deposition, a 3nm layer of titanium (Ti) needs to be deposited. The adhesive properties of titanium enhance the adhesion between the metal and silica, significantly mitigating difficulties during the lift-off process. Finally, 30nm of GST is sputtered into the PSW, followed by sputtering 5nm of silicon dioxide (SiO₂) to prevent oxidation of the 30nm GST. The entire process ensures high precision and controllability, providing feasibility for manufacturing this optical switch. Therefore, our proposed non-volatile multi-level reconfigurable optical switch, combining PSW and segmented GST, demonstrates repeatability in switch performance, excellent durability, and prolonged retention time. From a manufacturing perspective, the design offers advantages due to its compatibility with CMOS manufacturing processes.

 

Fig. 6. The schematic diagram of the fabrication process.

Download Full Size | PDF

6.2 Device tolerance

To assess the manufacturing tolerance caused by the shape inconsistency of GST due to inaccurate alignment during the secondary exposure process, we investigated the influence of segmented GST shapes on device performance. Concurrently, we also considered the gap W_gap between the metal cone funnel and silicon cone waveguide, as well as the positional deviation δ_Au between the PSW and silicon waveguide resulting from misalignment. Given the relatively frequent occurrence of positional deviations between the PSW and silicon waveguide, we further discuss the influence of such deviations on MNIST prediction simulation results. Specifically, we studied the impact of GST dimensions (ΔW_PCM) and positional deviations (δ_GST) on the optical transmittance difference (ΔT) of the device in both aGST and cGST, as depicted in Fig. 7(a). Figure 7(b) elucidates the effect of ΔW_PCM on ΔT. As each segmented GST's area gradually increases until the original design size (ΔW_PCM = 0), the optical transmittance difference under both phases gradually increases due to the increased average absorption loss of cGST Conversely, when the area of each segmented GST exceeds the design value (positive ΔW_PCM), the optical transmittance difference under both phases shows a decreasing trend due to the increased duty cycle of GST within each period, affecting the coupling efficiency between the mode field and GST. It is noteworthy that, although the optical transmittance under both phases is only 70% at ΔW_PCM= −20nm, within the range of -15nm ≤ ΔW_PCM ≤ 20nm, the average optical transmittance under both phases can still be maintained at above 80%, which is deemed acceptable. Furthermore, Fig. 7(c) presents the impact of the overall positional deviation distance of the segmented GST structure on the device. It can be observed that when δ_GST is within the range of ±20 nm, the average optical transmittance difference under both phases remains at 85%. The gap W_gap affects the absorption loss when the dielectric silicon waveguide mode is converted to the PSW mode. However, when there is a particular gap maintained between the metal and silicon waveguides (W_gap ≤ 80nm), as shown in Fig. 7(d), the optical transmittance difference of the device under both phases can still maintained above 80%. Regarding the positional deviation δ_Au between the PSW structure and the silicon waveguide during the manufacturing process, as illustrated in Fig. 7(e), when δ_Au exceeds 30nm, it increases the absorption loss during the mode transition from the silicon waveguide to the PSW, leading to a decreasing trend in the transmittance difference. However, when δ_Au maintained within the range from -30 to 30 nm, the optical transmittance difference can still be above 80%, which is acceptable. A high optical transmittance difference is required to ensure a wide range of optical weight updates. Our analysis of the positional deviation δ_Au shows that in case δ_Au is 50nm, the optical transmittance difference is only 76%. Therefore, after 40 iterations of training, as shown in Fig. 7(f), its accuracy significantly decreases but remains above 90%. Therefore, we can conclude that the designed device performs well regarding manufacturing tolerance.

 

Fig. 7. (a) Diagram of the four types of tolerance. (b) The ΔT versus the ΔW_PCM. (c) The ΔT versus the δGST. (d) The ΔT versus the ΔW_gap. (e) The ΔT versus the δAu. (f) The influence of δAu on MNIST prediction simulation results.

Download Full Size | PDF

7. Conclusion

In this paper, we propose a multi-level tunable optical switch combining segmented GST and PSW. By utilizing PSW, this design offers extreme light confinement, improves device compactness, and enhances the interaction between light and matter. Additionally, we employ segmented GST in the PSW with periodicity to adjust the average refractive index of the waveguide, influencing the coupling efficiency between the mode field and GST, thus increasing the transmittance. Considering the coupling losses between dielectric waveguides and PSW, as well as the intrinsic losses of GST, we design a non-volatile optical switch with a large tunable range of transmittance. The optical transmittance difference between aGST and cGST reaches a high level of 85%, accompanied by a low insertion loss of 0.5 dB at a wavelength of 1550nm. Furthermore, through recognition tests based on weight updates using our non-volatile multi-level optical switch, we successfully achieved a recognition accuracy of 95% for MNIST handwritten digits. The demonstrated high accuracy confirms that combining segmented GST with PSW in an artificial neural network (ANN) exhibits outstanding performance, laying the foundation for the potential construction of PCMs neural networks and paving the way for achieving more efficient memory computing in the future neuromorphic networks.