Multilevel Photonic Switching in GST-467 for Deep Neural Network Inference

TL;DR

This work introduces GST-467 as a high-contrast optical PCM suitable for multi-level photonic switching in silicon photonics. By implementing a segmented GST-467 topology on an SOI waveguide and optimizing via 3D FDTD, EMA, and multiphysics laser-heating simulations, the authors achieve a sevenfold improvement in the switch's figure of merit and up to 48 resolvable transmission states. The device operates optimally in the telecom C-band (around $\lambda \approx 1550$ nm) with strong extinction ratios and low insertion loss, enabling energy-efficient, all-optical DNN inference that outperforms other PCMs on Fashion-MNIST and EMNIST benchmarks. These findings position GST-467 as a leading material for scalable, low-power photonic computing and neuromorphic hardware.

Abstract

Phase-change materials (PCMs) have emerged as key enablers of non-volatile, ultra-compact photonic switches for energy-efficient deep neural network (DNN) applications. In this work, we investigate the recently discovered $\mathrm{Ge_{4}Sb_{6}Te_{7}}$ (GST-467) as a high-contrast optical PCM and demonstrate its suitability for multi-level photonic computing. The complex refractive indices of amorphous and crystalline GST-467 were experimentally extracted and used to propose a segmented silicon-on-insulator photonic switch optimized at 1550 nm. Three-dimensional FDTD simulations reveal that segmentation significantly enhances the extinction ratio while maintaining low insertion loss, resulting in a more than seven times higher design figure of merit than an unsegmented design. Laser-induced thermo-optical simulations further establish efficient, reversible switching with sub-nJ energy requirements for crystallization and amorphization. Compared with established GST, GSST, and GSS compositions, GST-467 provides the largest transmission contrast and supports up to 48 resolvable optical states. When deployed as multi-level weights in photonic DNN architectures, the GST-467 switch achieves superior classification accuracy on EMNIST and Fashion-MNIST benchmarks. These results position GST-467 as a highly promising PCM for scalable, low-energy photonic computing and neuromorphic hardware.

Figures (9)

• Figure 1: (a) 3D schematic representation of the proposed optical switch. The parameters $l$ and $w$ denote the total length and width of the segmented GST sections, respectively. (b) Top view of the switch layout, where $\mathrm{L_{PCM}}$ and $\mathrm{L_{etch}}$ represent the lengths of the individual GST-467 and etched sections. (c) Cross-sectional view along the vertical y-z plane of the switch structure. The heights of the GST-467 segment and the silicon waveguide are indicated by hPCM and hSi, respectively. An Al2O3 layer serves as the top cladding for the switch, while SiO2 is employed as the cladding material in all other directions.
• Figure 2: Fundamental TE mode profiles of the silicon waveguide at a wavelength of 1550 nm: (a) without GST-467, (b) with amorphous GST-467 (a-GST), and (c) with crystalline GST-467 (c-GST) segments. The presence of the high-index GST cladding layer results in a lateral shift of the mode center toward the GST segment from the center of the silicon core. Furthermore, the mode becomes increasingly confined in the presence of GST, with the highest confinement observed in the c-GST case, as indicated by the elevated effective index values.
• Figure 3: Structure optimization results. (a) 3D surface plot of FOM as a function of $\mathrm{T_{PCM}}$ and $\mathrm{L_{PCM}}$. The maximum FOM is observed for $\mathrm{T_{PCM}}$ = 880 nm and $\mathrm{L_{PCM}}$ = 80 nm. (b) With $\mathrm{T_{PCM}}$ = 880 nm held constant, the $\mathrm{L_{etch}}$ is varied to identify the optimal configuration. The highest FOM is achieved for $\mathrm{L_{PCM}}$ = 80 nm and $\mathrm{L_{etch}}$ = 50 nm. Variation of (c) FOM and (d) ER and IL as a function of the number of GST segments for the fixed $\mathrm{T_{PCM}}$ and $\mathrm{L_{etch}}$ of 880 nm and 50 nm. The FOM attains its maximum value when the number of segments is 11, corresponding to $\mathrm{L_{PCM}}$ = 80 nm. This maximum FOM originates due to the highest ER (48.36 dB) achieved for the optimized dimensions while having a lower IL (0.86 dB).
• Figure 4: Electric field distribution during propagation through the switch structure. Field propagation through the PCM segment is shown for (a) a-GST and (b) c-GST, while propagation through the silicon core is depicted for (c) a-GST and (d) c-GST phases. In the a-GST state, the electric field exhibits strong confinement within the silicon core, with reduced field intensity in the PCM region, indicating efficient guided mode propagation. Conversely, for the c-GST phase, the field rapidly decays along the propagation direction. Additionally, an increase in field intensity within the PCM segment is observed, signifying a lateral shift of the mode center toward the PCM layer. This shift enhances absorption losses due to the high extinction coefficient associated with the crystalline phase, thereby contributing to field dissipation.
• Figure 5: Wavelength-dependent optical performance of the optimized segmented GST-467 switch. Real and imaginary parts of the neff for the (a) amorphous and (b) crystalline states, respectively, showing monotonic decreases in both components with increasing wavelength. (c) Computed FOM across the S, C, and L-bands, exhibiting a pronounced peak near 1550 nm due to the optimal trade-off between low insertion loss and high extinction ratio. (d) Simulated transmission spectra for the amorphous and crystalline phases, where Tamor increases monotonically with wavelength while Tcry reaches a minimum in the C-band. (e) Corresponding IL and ER spectra, revealing that IL decreases steadily whereas ER attains a maximum near 1550 nm, confirming the C-band as the optimal operational window for high-contrast, low-loss switching in the GST-467-based segmented photonic structure.