Spiking neurons from tunable Gaussian heterojunction transistors

Spiking neural networks (SNNs) are increasingly attractive for artificial intelligence (AI) due to their inherent energy efficiency. This efficiency stems from their ability to exploit spatiotemporal processing, spiking sparsity, and high inter-neuron bandwidth. However, traditional silicon-based complementary metal-oxide-semiconductor (CMOS) technology struggles to fully realize the potential of SNNs. CMOS transistors do not intrinsically mimic the time-dependent conductance of biological ion channels, necessitating complex multi-transistor circuits for each neuron-synapse pair. This complexity significantly reduces the achievable very-large-scale integration (VLSI) density. Existing CMOS-based SNN implementations, even those achieving multiple spiking modes, often require at least 20 transistors per neuron, subject to stringent design constraints and current-based addressing. While digital approaches like IBM TrueNorth and SpiNNaker offer apparent VLSI advantages, their limited chip area necessitates multiplexing several neurons onto each digital core, thus compromising the inherent parallelism of biological networks. To overcome these limitations, researchers are exploring alternative materials and device architectures capable of directly encoding neuromorphic functionality at the device level. Memristors, memtransistors, domain-wall memories, metal-insulator-transition (MIT) devices, multi-gated transistors, and Gaussian synapses have shown promise for synaptic functions, but creating efficient spiking neuron implementations remains a challenge. Existing attempts, based on MIT devices, suffer from low gain and limited output swing. Other approaches, combining memristors with capacitors or CMOS transistors, lack biophysical characteristics or require a large number of circuit elements. Similarly, leaky integrate-and-fire neurons implemented using the magneto-electric effect suffer from continuous energy dissipation, and ferroelectric field-effect transistor-based neurons are limited in the range of spiking behaviors they can exhibit. Photonic implementations show speed and bandwidth advantages, but lack biophysical characteristics. The use of low-dimensional materials, however, offers a potential solution. These materials, exhibiting weak electrostatic screening, enable gate-tunable electronic properties ideal for neuromorphic applications. Specifically, atomically thin semiconducting materials in gate-tunable p-n heterojunctions create an antiambipolar response with Gaussian transfer curves. While this behavior has found use in signal processing, logic devices, and photodetectors, previous single-gated geometries lacked the control needed for efficient neuromorphic functionality. This research addresses this gap by introducing a novel dual-gated approach.

The existing literature highlights the limitations of CMOS-based approaches to building spiking neural networks (SNNs) for neuromorphic computing. The complexity of replicating the temporal dynamics of biological neurons using CMOS transistors leads to inefficient and bulky circuits. Several alternative materials and devices have been proposed, including memristors, memtransistors, and MIT devices, each with its own set of drawbacks. Memristor-based solutions often lack the dynamic range and precision needed for complex neural behaviors. Similarly, memtransistors, while exhibiting some synaptic functionalities, have limitations in implementing the full range of spiking patterns seen in biological systems. Multi-gated transistors offer some improvements in control but still require sophisticated circuit designs. The research reviewed suggests a need for a simpler, more efficient device that can intrinsically emulate the behavior of biological neurons, leading to the exploration of low-dimensional materials and their heterojunctions.

This study focuses on the fabrication and characterization of dual-gated Gaussian heterojunction transistors (GHeTs) based on a mixed-dimensional van der Waals heterostructure of monolayer molybdenum disulfide (MoS2) and solution-processed semiconducting single-walled carbon nanotubes (CNTs). The fabrication process employs a self-aligned technique using photolithography and atomic layer deposition (ALD) to create a semi-vertical device architecture. Monolayer MoS2, chosen for its atomically thin nature, processing stability, and large-area compatibility via CVD, serves as the n-type material. Solution-processed CNTs, selected for their p-type/ambipolar characteristics, conformability, and suitable band alignment with MoS2, form the p-type component. The self-alignment method involves creating undercut profiles in the photoresist to allow for controlled metal evaporation and conformal ALD of the dielectric oxide, resulting in an encapsulated metal electrode with a self-aligned dielectric extension. Atomic force microscopy (AFM) is used to verify the quality and dimensions of the dielectric extension. The fabrication process involves sequential steps: CVD-grown monolayer MoS2 transfer and patterning; self-aligned bottom contact fabrication; additional dielectric growth and patterning as an etch mask; deposition of top contacts; transfer of a CNT network with deterministic overlap on the MoS2; and finally, ALD of a dielectric and top gate patterning. The resulting dual-gated GHeTs are characterized electrically. Initially, independent biasing of top and bottom gates is used to assess the tunability of the Gaussian transfer curves. Rectification ratios and transconductance are measured to determine the influence of each gate on the device's behavior. Next, a dependent biasing scheme is employed to optimize gate modulation. Gaussian fits to the antiambipolar response are used to quantify the peak position, height, and full-width-half-maximum (FWHM) under different biasing conditions. The device is integrated into a circuit to demonstrate its use as a spiking neuron. This involves modeling the Hodgkin-Huxley (HH) model of biological neurons, emulating sodium (Na+) and potassium (K+) ion channels using the GHeT and additional transistors and passive components. The circuit is simulated using the Cadence Virtuoso platform with the Spectre simulator and experimentally implemented. Various spiking behaviors are analyzed, including phasic spiking, delayed spiking, tonic bursting, spike latency, integrator, and phasic spiking. The energy consumption is evaluated. Further simulations explore the impact of various circuit parameters and biasing strategies on different spiking modes.

The research successfully demonstrates a highly tunable dual-gated Gaussian heterojunction transistor (GHeT) fabricated from a mixed-dimensional van der Waals heterostructure of monolayer MoS2 and solution-processed semiconducting single-walled carbon nanotubes (CNTs). The dual-gate geometry allows for unprecedented electrostatic control over the device's Gaussian transfer characteristics. Independent gate operation shows that the top gate modulates the CNT film, while the bottom gate modulates the MoS2, enabling tuning of rectification ratios over two orders of magnitude. Dependent gate operation, where both gates are swept simultaneously with a controlled offset, further enhances control, allowing for tunability of the Gaussian response's peak position, height, and FWHM over three orders of magnitude. The GHeT exhibits a high yield (85% over a 0.5 × 0.5 cm area). Crucially, the study demonstrates the GHeT's application as a spiking neuron, replicating various biological spiking behaviors. Using a simple circuit comprising a single GHeT, two transistors, two capacitors, and two resistors, the researchers achieved constant spiking, mimicking Class-I spiking, and phasic spiking. By making minor modifications to the circuit, they also demonstrated spike latency, integrator, phasic bursting, tonic bursting, and dampened tonic bursting behaviors. The energy consumption of this spiking neuron is approximately 250 nJ per spike, with potential for significant reduction through circuit optimization and integration. The experimental results closely match the simulations, validating the model and its potential for neuromorphic computing.

This study successfully addresses the need for simplified, energy-efficient spiking neuron implementations for neuromorphic computing. The demonstrated dual-gated GHeT offers a significant advancement over previous approaches, providing superior electrostatic control within a smaller device footprint compared to other antiambipolar devices. The ability of the GHeT to mimic the complex transient behavior of biological sodium ion channels using a simple circuit is a critical achievement. Furthermore, the dual-gated programmability, through both independent and dependent biasing schemes, allows for a remarkable versatility in generating a diverse range of biological spiking responses, demonstrating its potential for building complex neuromorphic systems. The compatibility of the GHeT fabrication process with previously demonstrated MoS2 memtransistor-based synapses opens up pathways towards scalable, biomimetic neuromorphic platforms. The tunable Gaussian response of the GHeT extends beyond neuromorphic computing, finding potential applications in natural language processing, machine learning (Bayesian inference), and computer vision, where complex Gaussian functions are crucial components of algorithms. The inherent simplicity of the GHeT compared to CMOS-based implementations promises to significantly accelerate the development of advanced AI technologies.

This research successfully demonstrates a highly tunable dual-gated Gaussian heterojunction transistor (GHeT) based on a MoS2-CNT heterostructure, capable of emulating a wide range of biological neuron spiking behaviors using a simple and efficient circuit. The device's versatility, ease of fabrication, and low energy consumption make it a promising candidate for accelerating the development of advanced neuromorphic computing and AI technologies. Future research directions include further optimization of the device and circuit designs to improve performance, reduce energy consumption, and explore integration with other neuromorphic components to create complex neural networks.

While the study demonstrates the GHeT's capabilities in replicating various spiking patterns, the current implementation is a proof-of-concept. Further research is needed to optimize the device for higher speeds and lower energy consumption. The long-term stability and reliability of the GHeT under continuous operation need to be thoroughly investigated. Additionally, scaling up the fabrication process to produce large arrays of GHeTs for practical applications remains a challenge that needs to be addressed in future work.