2D Materials Powering Neuromorphic Intelligence

As the demand for energy-efficient and adaptive computing architectures continues to surge, conventional von Neumann systems face critical limitations in power consumption and scalability for AI and machine learning workloads. Now, a collaborative team led by Professor Jamal Kazmi (Shanghai University), Professor Peijian Wang (Wenzhou University / National University of Singapore), Professor Mohd Ambri Mohamed (Universiti Kebangsaan Malaysia), Professor Federico Rosei (University of Trieste), Professor Zhiming M. Wang (University of Electronic Science and Technology of China), and Professor Hongwei Song (Jilin University) has presented a comprehensive review on how two-dimensional materials are powering next-generation neuromorphic intelligence.

Why 2D Materials Matter

Traditional digital processors suffer from the von Neumann bottleneck—separating memory and computation creates massive latency and energy costs. Biological systems, by contrast, operate at roughly 20 watts while consuming only 1–100 femtojoules per synaptic event. The introduction of atomically thin 2D materials—including transition metal dichalcogenides (TMDs), hexagonal boron nitride (h-BN), black phosphorus (BP), and emerging tellurene—enables neuromorphic devices with unprecedented control over electronic and optoelectronic properties, bridging the gap between biological efficiency and machine intelligence.

Innovative Material Platforms and Mechanisms

The review systematically maps how distinct 2D material properties enable specific neuromorphic functionalities:

• TMDs (MoS₂, WSe₂, WS₂): Their tunable bandgaps, strong optical response, and defect-mediated switching enable both volatile and non-volatile synaptic behaviours. WSe₂-based photoelectric synapses achieve ~0.1 fJ per operation—100× lower than a human synapse. Controlled chalcogen vacancy migration and 2H–1T′ phase transitions provide analogue conductance modulation for long-term potentiation (LTP) and depression (LTD).
• h-BN: With its large bandgap (~6 eV), ultra-low leakage current, and high breakdown strength, h-BN serves as an outstanding dielectric and active switching medium. Wafer-scale h-BN memristor arrays demonstrate 98% yield with endurance exceeding 10⁷ cycles, while atomically thin devices achieve switching speeds as fast as 120 ps.
• Black Phosphorus: Its superior carrier mobility and in-plane anisotropy enable direction-sensitive neuromorphic devices and polarization-resolved optoelectronic synapses, though environmental stability remains a key challenge requiring advanced encapsulation.
• Emerging Xenes (Tellurene, Silicene, Trigonal Selenium): These mono-elemental 2D materials offer layer-dependent bandgap tuning, multifunctional piezoelectric/thermoelectric properties, and high mechanical flexibility—ideal for wearable and bio-integrated neuromorphic systems.

Outstanding Device Performance

The review highlights record-breaking metrics across multiple architectures:

• Memristors: MoS₂/graphene heterostructure devices exhibit near-linear weight updates with nonlinearity factor of 0.276, 10⁴ s retention, and ~100 distinct conductance states. Graphene/MoS_2-xO_x/graphene (GMG) memristors achieve exceptional endurance of 10⁷ cycles at 340°C.
• Memtransistors: Polycrystalline MoS₂-based three-terminal devices demonstrate gate-tunable resistance states spanning four orders of magnitude, cycling endurance of 475 times, and projected one-year retention.
• Vertical Heterostructures: WS₂/MoS₂ band-modulation memristors and PdSeOₓ/PdSe₂ heterostructures enable convolutional image processing in crossbar arrays, exploiting "lift-gate"-like carrier control without directly damaging the switching layer.

Machine Learning Integration

When paired with neural network algorithms, these devices achieve remarkable results:

• Supervised Learning: MoS₂ memristor-based hardware neural networks reach 55% recognition accuracy on MNIST, closely matching software performance (99.41%).
• Unsupervised Learning: Graphene memristive synapses with STDP and lateral inhibition achieve ~80% accuracy on MNIST clustering tasks.
• Reinforcement Learning: 2D ferroelectric α-In₂Se₃ devices implementing Markov Decision Process algorithms improve success rates from 68% to 82% in maze-navigation tasks through cooperative learning modes.
• In-Memory Computing: A 32×32 monolayer MoS₂ floating-gate FET array performs vector–matrix multiplication for real-time signal filtering (low-pass, high-pass, feedthrough) in a single cycle.

Applications and Future Outlook

The review establishes a clear roadmap for 2D neuromorphic systems across six frontiers:

1. Wearable & Edge Computing: Ultra-flexible MoS₂/LiSiOₓ synapse arrays on polyimide substrates achieve 94.5% MNIST accuracy with sub-5 mm bending radius, enabling real-time localized processing for IoT and medical monitoring.
2. Brain–Machine Interfaces: 2D devices with biocompatible, ultra-low-power signal transduction and high spatial resolution pave the way for precision neuroprosthetics and closed-loop cognitive enhancement.
3. Quantum Neuromorphic Systems: Hybrid quantum–classical architectures leveraging tunnelling and spin phenomena in TMDs and CrI₃ enable noise-resilient spiking neural networks and secure AI.
4. Material-Level Solutions: Wafer-scale CVD, ALD passivation, and defect engineering address synthesis challenges.
5. Device-Level Innovations: vdW heterostructures and strain engineering tackle performance variability and contact resistance.
6. System-Level Integration: High-density 2D architectures with energy-efficient vdW systems and CMOS compatibility drive scalable deployment.

This work establishes a new paradigm for neuromorphic computing, explicitly linking material properties → device architectures → machine learning modes → target applications in a unified framework. By combining atomic-scale thickness with exceptional electronic, optoelectronic, and quantum mechanical properties, 2D materials offer a transformative pathway toward sustainable, adaptive, and intelligent technologies that rival biological neural networks.

Stay tuned for more groundbreaking research from this international collaborative team spanning Shanghai University, Southern University of Science and Technology, Liverpool John Moores University, University of Glasgow, King Fahd University of Petroleum and Minerals, Wenzhou University, National University of Singapore, Universiti Kebangsaan Malaysia, University of Trieste, University of Electronic Science and Technology of China, and Jilin University!