Non-volatile, compact, and broadband multimode matrix vector multiplication for photonic neuromorphic computing
Year: 2026
Type: article
Abstract: Matrix-vector multiplication (MVM) is a fundamental operation in artificial neural networks. While conventional electronic implementations of MVMs face bottlenecks in speed and energy efficiency, here, we design an ultra-compact (185 μm × 21 μm) non-volatile photonic MVM engine leveraging mode division multiplexing (MDM) with the fundamental and first-order transverse magnetic modes. The architecture integrates a multimode-interferometer power splitter (0.98 dB insertion loss), a directional-coupler (de)multiplexer (≈ -10 dB crosstalk suppression, 0.87 dB insertion loss), a weighting core comprising four independently programmable weighting paths, each containing five Ge 2 Sb 2 Se 4 Te 1 (GSST) phase-change cells in series, giving twenty cells in total. By electrically programming individual GSST cells between amorphous and crystalline states, the device supports 1296 distinct matrix configurations with an extinction ratio of 11.29 dB and 17.3 dB for TM0 and TM1 modes, respectively. Electro-thermal simulations indicate ultrafast and reversible switching using a 5 V and 33 ns pulse for crystallization, while a 12 V and 5 ns pulse triggers melting and rapid quenching for amorphization. The non-volatile nature of GSST eliminates static power consumption, enabling energy-efficient operation. The complete MVM exhibits a total insertion loss of 2.76 dB and 3.64 dB at 1550 nm for TM0 and TM1 modes, respectively. This device maintains functionality across a 40 nm bandwidth (1540–1580 nm), offering a scalable pathway for energy-efficient, MDM-enabled photonic computing.
Source: Scientific Reports
Authors: Ali Atri, Abbas Zarifkar
Institution: Shiraz University
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: gold
Estimated APC paid: $2,690
OpenAlex ID: W7223443442