Ultrafast Non-Volatile Weyl LuminoMem for Mid-Infrared In-Memory Computing

Abstract: Integrated optoelectronic systems strive to combine the logic/memory density of electronics with the bandwidth of photonics, but monolithic realization is impeded by the inefficient electronic-to-photonic interface. Current architectures rely on separate readout circuitry and modulators, creating bottlenecks in energy and latency, while existing direct transduction methods often compromise on switching speed or non-volatility. Here, we report an ultrafast, non-volatile optoelectronic memory, named LuminoMem, that integrates electrical storage and mid-infrared light emission in a single device. The device utilizes a floating-gate architecture, in which the Weyl semiconductor tellurium serves simultaneously as a charge-trapping storage layer and an emissive medium. This design enables nanosecond-scale electrical programming of non-volatile photoluminescence at 3.4 um, allowing direct optical access to stored states without external modulation. We demonstrate 4-bit (16-level) optical storage capacity and validate the device's performance through neural network simulations that achieve high accuracy on the Fashion-MNIST dataset. By effectively bridging the gap between electronic storage and mid-infrared photonics, the demonstrated mid-infrared LuminoMem provides a hardware foundation for promoting current computation efficiency and potential intelligent platforms that co-integrate computing, memory, and sensing capabilities.

• The paper introduces LuminoMem, a monolithically integrated, ultrafast, non-volatile memory that achieves 4-bit multilevel storage with a switching speed of 70 ns.
• It employs a Gr/h-BN/Te heterostructure and Fowler-Nordheim tunneling to precisely modulate mid-IR photoluminescence for direct optical encoding of logic states.
• Furthermore, the device demonstrates robust endurance, accurate analog weight programming, and significant potential for neuromorphic and hybrid photonic computing applications.

Ultrafast Non-Volatile Weyl LuminoMem for Mid-Infrared In-Memory Computing

Introduction and Motivation

Integrated optoelectronic systems are increasingly fundamental for advancing computation architectures that require the high-density logic and storage intrinsic to electronics and the high bandwidth and parallelism achievable through photonics. However, the integration of these domains is hampered by inefficient, energy-intensive electronic-to-photonic interfaces and fabrication incompatibilities. Conventional system architectures, involving separate electrical readout and optical modulation stages, are bottlenecked by energy overheads, latency, and material/process incompatibilities. Most existing direct transduction strategies compromise either on non-volatility or switching speed, and thus fail to satisfy the requirements for scalable, high-performance, in-memory computing in photonic domains.

Device Architecture and Physical Principles

The work introduces LuminoMem, a monolithically integrated, ultrafast, non-volatile optoelectronic memory, utilizing a van der Waals (vdW) floating-gate architecture with Weyl semiconductor tellurium (Te) as both the charge storage and mid-infrared (MIR) light emissive medium. The device is constructed with a Gr/h-BN/Te heterostructure on a SiO$_2$/Si substrate. Te’s nearly direct 365-meV bandgap enables MIR photoluminescence centered at 3.4 μm—a wavelength of high practical utility for molecular sensing and communication.

Electrical programming exploits Fowler-Nordheim tunneling across the h-BN barrier: positive gate biases inject electrons and quench the majority holes in Te, reducing radiative recombination and MIR PL (program/"OFF" operation), while negative biases eject electrons, accumulate holes, and enhance emission (erase/"ON" operation). The h-BN barrier ensures charge retention and non-volatility, while the NIR readout (1550 nm) is non-destructive, as the photon energy is insufficient to promote stored carriers out of Te. The result is direct, MIR optical encoding and retrieval of logic states, bypassing external modulators and optimizing energy and latency.

Multi-Level Storage and Performance Metrics

LuminoMem supports 4-bit (16-level) analog storage per cell, achieved by fine control of the programming pulse voltage amplitude. The charge accumulation in the floating Te is highly deterministic, leading to clearly separated optical readout levels. The device demonstrates a memory window of up to 80 V, with endurance exceeding 1,000 program/erase cycles and stable data retention over 10$^4$ s, all with negligible photoluminescence degradation. Notably, the switching speed reaches 70 ns—orders of magnitude faster than PCM or ferroelectric optical memories, which are typically limited to the microsecond regime. The speed is fundamentally attributed to the atomically sharp vdW interfaces, the high dielectric strength of few-layer h-BN, and minimized parasitic RC delays.

The non-volatile photoemission states are spectrally stable, with MIR peak positions fixed throughout operation, an essential characteristic for wavelength-division multiplexed photonic systems.

Neuromorphic Computing and In-Memory Photonic Architectures

The analog, linear, and programmable modulation of MIR PL intensity in LuminoMem allows direct emulation of synaptic weights, with potentiation (incremental PL increase) and depression (PL decrease) mapped to rapid electrical pulse sequences. Simulated in hardware-aware crossbar neural networks (using data-driven LUTs of measured device characteristics), arrays of LuminoMem cells are deployed as electronic/photonic synapses for a three-layer BP-ANN, benchmarked on the Fashion-MNIST dataset.

The system attains 86.9% accuracy on pristine test data, with only modest degradation to 82.5% and 76.6% under 10% and 20% additive input noise, respectively—demonstrating robust, fault-tolerant learning suitable for noisy sensory and embedded environments. The stability and endurance of weight programming cycles were confirmed for 20 analog switching levels, suitable for high-precision neural inference and analog in-memory processing.

Implications and Future Prospects

LuminoMem overcomes longstanding barriers in monolithic optoelectronic integration by combining non-volatile, multilevel storage with nanosecond-scale photonic programmability and direct, MIR optical readout. The electrical-to-optical interface is fundamentally streamlined, eliminating redundant conversion chains and enabling scalable, high-density, low-latency photonic computing systems.

From a theoretical perspective, leveraging topological Weyl semiconductors such as Te creates opportunities for further exploiting topologically robust, polarization-tunable, and nonlinear optical transitions, potentially enhancing information throughput and device functionality in MIR neural networks and photonic circuits. The architecture is also adaptable—the PL optical readout may be supplanted by electroluminescence, and the programming/erasing could be fully optical, paving the way for fully photonic in-memory computation platforms.

Practically, the presented design enables integration with CMOS control electronics and wavelength-multiplexed photonic interconnects. Applications extend beyond computation to reconfigurable sensing, environmental and industrial monitoring, and secure communications in MIR atmospheric windows. The demonstrated robustness against noise, endurance, and high-density analog storage directly address requirements for edge AI and intelligent sensing hardware.

Conclusion

LuminoMem constitutes a significant advance in optoelectronic in-memory computing, providing ultrafast, non-volatile, multi-level analog memory and MIR optoelectronic modulation in a single monolithic platform. Its unique combination of speed, endurance, storage density, and optical accessibility suits it for direct integration in future photonic neural networks, MIR sensor-processing arrays, and hybrid electronic-photonic computational systems. Future research directions include all-photonic programming and readout protocols, expanded functionality via Weyl-physics and nonlinear effects, and wafer-scale device uniformity necessary for practical deployment of MIR photonic neural architectures.