Skip to content

Wide‐Bandgap Semiconductor‐Based Neuromorphic Computing

FieldValue
JournalInformation & Functional Materials
Published2026-05-23
AuthorsHongyu Tang, Pengsheng Min, Yang Zhang, Q H Zhang, Wei Zhang
DOI10.1002/ifm2.70020

Wide-bandgap semiconductor neuromorphic hardware has reached 100% accuracy on digit patterns 0 to 9 in a physical SiC-memristor reservoir. A separate AlN photonic circuit recognized gestures and gaits with accuracy above 96%, latency below 10 ns, and energy below 0.34 pJ. The review connects these results to material properties, device physics, and process choices across representative semiconductor families.

This is a literature review rather than a new device experiment. The authors survey representative memristors, two-terminal synaptic devices, synaptic transistors, and artificial neurons. They compare material properties, switching mechanisms, array demonstrations, and application benchmarks from published studies.

Table 1 compares bandgap, thermal conductivity, breakdown field, relative permittivity, and mobility. Device tables collect forming voltage, retention, endurance, on/off ratio, energy per event, optical wavelength, and demonstrated synaptic functions.

These reported values come from different stacks and test conditions. The review identifies material trends and device trade-offs rather than ranking devices under one controlled protocol.

Reference targets include an on/off ratio above 10² for practical memory. They also include synaptic-transistor energy of 1 to 10 fJ, switching from subnanoseconds to microseconds, endurance above 10⁹ cycles, and retention above 10 years at room temperature. These targets are not achieved by every reviewed device.

Material choice sets the available switching mechanism and optical, thermal, and polarization controls.

GaN and AlN rely on spontaneous and piezoelectric polarization, charge trapping, and interface barriers. AlScN adds switchable ferroelectric polarization that changes carrier accumulation and transport barriers.

BN devices formed metal-ion or carbon filaments through defects and grain boundaries. h-BN also supported tunneling in atomically thin layers.

SiC devices combined defect trapping with metal-ion filaments in amorphous or crystalline structures. 4H-SiC supplied higher mobility and stable elevated-temperature operation.

Ga₂O₃ used oxygen-vacancy filaments or ionic redistribution for digital or analog switching. Its lower thermal conductivity remains a concern for dense integration.

Diamond used defect and phase-change behavior, including reversible sp² to sp³ bonding changes and charge accumulation at doped interfaces.

The largest performance spread occurs in resistance ratio, endurance, and energy because the test conditions differ.

PlatformTest conditionReported measurement
Vertical Ag/β-Ga₂O₃/Pt memristor0.13 V programming, 0.11 V reverse reset, 10⁻¹⁰ A programming currentOn/off ratio above 10⁸ and retention above 4 × 10⁴ s
Ni/h-BN/Ni memristorOperation below 3 VOn/off ratio above 10⁷
Au/Ti/h-BN/Au memristorElectrical switching testSwitching speed below 5 ns
SiC/Si bilayerResistance states programmed by varying pulse conditionsEndurance above 10⁹ cycles and on/off ratio of about 10²

The Ga₂O₃ on/off ratio far exceeds the review’s 10² reference for practical memory. Its 10⁻¹⁰ A programming current belongs to that vertical device, not to Ga₂O₃ devices generally.

Most Ga₂O₃ memristors in the review showed 10² to 10³ cycles and 10³ to 10⁴ s retention. The SiC/Si bilayer exceeded 10⁹ cycles, but its on/off ratio only met the practical threshold.

Many listed retention windows remained below the 10-year goal. Analog updates also need stable linearity and symmetry, which reviewed reports do not consistently provide.

Ti-doped CVD diamond produced short- and long-term plasticity through dehydrogenation and hydrogenation linked to sp² and sp³ bonding. It retained its state for more than 10⁵ s, endured more than 150 cycles, and reached an on/off ratio of 10² to 10³.

The CuₓO/BDD junction responded at 3 mV and consumed 9.5 × 10⁻¹¹ J per event. Its high-resistance and low-resistance states were assigned to oxygen-vacancy clusters. That 95 pJ event energy exceeds the 1 to 10 fJ reference for three-terminal synaptic transistors, so the architectures are not directly comparable.

Si-doped diamond thin films made by heteroepitaxy and silicon doping produced broadband response, nonvolatile photocurrent, and wavelength sensitivity. The associated retina improved UV image recognition from 67.5% to 81.7%. It also simulated pupil constriction, image memory, color recognition, and dynamic light-spot trajectory tracking.

Nitrogen-doped ultrananocrystalline diamond on silicon responded to 595 nm and 808 nm pulsed light. Interface charge in physiological medium induced ionic displacement currents. Photoresponsivity reached 7.5 mA/W below ocular safety limits, and the system stimulated degenerated rat retinas.

Diamond selection should follow the target mechanism. Ti-doped diamond suits phase-change switching, BDD suits vacancy-based conductance, nitrogen-doped ultrananocrystalline diamond suits interface-mediated photocurrent, and Si-doped diamond suits broadband nonvolatile photoresponse.

Diamond/graphene heterostructures modulate interfacial bonding and add photoinduced conductivity. Their behavior depends on controlling interface bonding and defect states.

Device measurements do not guarantee useful array behavior. The review separates device operation from system demonstrations.

GaN HEMTs on 6-inch GaN-on-Si wafers reached 95% yield. Dendritic dislocations acted as three-dimensional charge stores for neuron-like processing and synaptic transmission.

BEOL-compatible SiC memristors used CVD-fabricated Cu/SiC/W stacks with compliance-free, self-rectifying switching. A physical reservoir accepted pulse streams and a trained readout. It classified digits 0 to 9 with 100% accuracy after training and tolerated noisy pixels.

AlN/Si waveguides combined microring feature extraction with Mach-Zehnder network computation. Gesture and gait accuracy exceeded 96%, with latency below 10 ns and energy below 0.34 pJ.

At 650 K, a 152-device ε-Ga₂O₃ array compressed 4 × 4 images into 4 × 1 vectors and reached 100% accuracy after 32 epochs. A 4H-SiC p-n array retained optical function in ambient air at 350°C, while its artificial neural network simulation reached 95% MNIST accuracy.

Defects serve as active memory media, but their stochastic behavior causes device-to-device and cycle-to-cycle variation. Vacancies, dislocations, and interface traps must therefore be controlled rather than accepted as uncontrolled noise.

Reliability under repeated programming remains uncertain. The authors identify conductance drift, retention loss, endurance loss, high operating voltages, and limited testing under elevated temperatures and high electric fields.

Integration maturity differs by material. GaN and SiC benefit from wafer-scale growth and CMOS compatibility. Ga₂O₃ and diamond still face thermal management, substrate availability, process standardization, and heterogeneous integration.

Diamond has two stated optical limits. Its intrinsic photoresponse is mainly in the deep ultraviolet, and its photocurrents are volatile. These limits hinder optical storage and nonvolatile neuromorphic functions.

Benchmarking lacks a common protocol for linearity, symmetry, dynamic range, event energy, temperature stability, and noise tolerance. Several recognition results were neural-network simulations rather than end-to-end hardware measurements.

The review argues that spiking and reservoir architectures should be matched to switching volatility, nonlinearity, and optoelectronic coupling. This co-design would use the material physics rather than force conventional neural-network models onto each device.

The paper’s neuromorphic memristor and optoelectronic synapse applications require diamond active material. Boron-doped diamond is specified for the CuₓO/BDD heterojunction. The diamond materials catalog lists MPCVD-grown single-crystal and polycrystalline plates and wafers, boron-doped diamond, and metallization with Au, Pt, Pd, Ti, W, or Cu, covering the material and electrode choices identified for that route. 6CCVD’s engineering support can assess those options against the intended device geometry and test conditions.

View Original Abstract

ABSTRACT Neuromorphic computing has emerged as a promising paradigm to overcome the energy inefficiency and data‐transfer bottlenecks of conventional von Neumann architectures by emulating the parallel and adaptive information processing of biological neural systems. To date, most neuromorphic hardware has relied on silicon‐compatible or narrow‐bandgap materials, which often face intrinsic trade‐offs among operating voltage, thermal stability, endurance, and multifunctionality. Wide‐bandgap semiconductors (WBGSs)—including Group III nitrides, gallium oxide, silicon carbide, and diamond—provide an alternative material platform enabled by their large bandgaps, strong polarization effects, diverse defect states, and compatibility with electronic and optoelectronic device architectures. This review surveys recent progress in WBGS‐based neuromorphic computing, with an emphasis on material‐enabled device physics rather than isolated demonstrations. Typical device concepts, including memristors, synaptic transistors, and neuronal devices, are systematically discussed together with their underlying resistive switching, charge trapping, polarization modulation, and optoelectronic mechanisms. Strategies for device integration and performance benchmarking are also addressed. Finally, remaining challenges and future research directions toward scalable and energy‐efficient neuromorphic systems based on WBGSs are outlined.