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

TL;DR

This paper demonstrates that GST-467, a phase-change material, enables efficient, multi-level photonic switches with high contrast and low energy for deep neural network inference, advancing scalable photonic computing.

Contribution

The study introduces GST-467 as a novel high-contrast PCM for multi-level photonic switching, with optimized design and demonstrated application in neural network inference.

Findings

GST-467 supports up to 48 optical states.

Segmentation improves extinction ratio significantly.

Achieves low-energy, reversible switching with high contrast.

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.