Two-dimensional materials for artificial sensory devices: advancing neuromorphic sensing technology

Abstract

Two-dimensional materials like graphene and transition metal dichalcogenides offer unique properties ideal for neuromorphic and artificial sensory devices, enabling precise emulation of neural functions with enhanced energy efficiency and flexibility. Despite their potential, challenges such as scalability, uniformity, and stability hinder widespread adoption. This review explores advancements, applications in robotics and AI, and future challenges in scaling 2D-based neuromorphic devices for real-world use.

Introduction

Neuromorphic devices, inspired by the architecture and functionality of biological neural networks, represent a significant advancement in modern computing^1,2. By integrating memory and computation within a single framework, these devices overcome the inefficiencies of traditional von Neumann computing systems, where the separation of processing and storage creates substantial bottlenecks in speed and energy efficiency^3,4. This integration enables parallel processing and local data storage^5,6, making neuromorphic devices particularly suitable for applications requiring high efficiency, such as artificial intelligence (AI), robotics, and edge computing^7,8. Existing commercial technologies, such as CMOS-based processors and memory devices, have long dominated the computing landscape. While these technologies excel in tasks requiring high computational power, they are limited in their ability to emulate the dynamic and learning-driven processes found in biological systems. Despite advances in AI and machine learning, the underlying hardware still faces significant challenges in adapting to real-time environmental stimuli and learning from experience in an energy-efficient manner. Furthermore, the ability of these devices to emulate human sensory functions, such as vision, touch, and hearing, opens new possibilities for artificial sensory systems that can process and respond to environmental stimuli in real time⁸. These advancements hold the potential to enable technologies capable of seamlessly adapting to and interacting with complex and dynamic environments^9,10,11.

Despite their potential, several challenges hinder the widespread adoption of neuromorphic devices^12,13. Device fabrication remains a significant hurdle, particularly in achieving uniform and scalable device arrays with consistent performance^14,15. Device reliability is another major concern, as the gradual degradation of materials and switching characteristics over extended operation can hinder the stable performance necessary for long-term applications¹⁶. Achieving energy-efficient operations without sacrificing computational speed is a persistent challenge, especially in high-performance computing regions¹⁷. Compatibility with existing semiconductor technologies poses integration difficulties, requiring innovative approaches to material synthesis and device engineering¹⁸. Finally, developing artificial sensory systems demands hardware capable of mimicking the complex dynamics of biological sensors with high precision resulting in current neuromorphic devices requiring significant improvement.

Two-dimensional (2D) materials, such as graphene^19,20,21,22, transition metal dichalcogenides (TMDs)^23,24, and hexagonal boron nitride (h-BN)²⁵, offer unique advantages for addressing these challenges²⁶. Their atomic-scale thickness and exceptional electrical, mechanical, and optical properties enable the miniaturization and enhancement of neuromorphic devices. In artificial sensory systems, the versatility of 2D materials supports the development of susceptible and adaptive sensors. For example, their tunable electronic properties allow precise signal processing control, while their mechanical flexibility facilitates integration into wearable and flexible electronics²⁷. Additionally, the high surface-to-volume ratio of 2D materials enhances interaction with external stimuli^28,29, making them ideal for replicating biological sensory responses such as detecting light, pressure, or chemical signals³⁰. These attributes position 2D materials as crucial components in achieving the next generation of neuromorphic and sensory systems^31,32,33.

In this review, we first explore the application of 2D materials in neuromorphic devices, highlighting their opportunities to address existing technological challenges and advancement in the field. We then present recent studies highlighting various operating mechanisms in neuromorphic devices enabled by novel 2D materials, illustrating the diversity of approaches and their impact on device performance. Following this, we discuss the application of 2D materials in artificial sensory neuromorphic devices, emphasizing their unique potential to replicate biological sensory functions with high precision. The integration of these devices into larger neuromorphic systems will be discussed, focusing on how 2D materials contribute to achieving scalable, efficient, and multifunctional architectures. The outlook and perspectives for achieving the next-generation 2D material-based artificial sensory devices will be provided.

2D-materials in neuromorphic devices

2D-materials have recently emerged as a groundbreaking component in developing neuromorphic devices. Defined by their atomically thin structure, 2D materials like graphene, TMDs, and h-BN exhibit unique electronic, optical, and mechanical properties that differentiate them from bulk materials^26,34,35,36. These materials possess high carrier mobility, tunable bandgaps, and mechanical flexibility, which, combined with their reduced dimensionality, make them ideal for miniaturized, energy-efficient neuromorphic applications. The atomic-scale thinness of 2D materials enables rapid ion and electron movement, reducing the switching time in synaptic devices^37,38,39,40. Their structures allow for strong in-plane covalent bonding and van der Waals (vdW) interactions in the out-of-plane direction, offering robust mechanical stability and flexibility. Furthermore, 2D materials benefit from minimal structural defects due to the absence of dangling bonds, which enhances device reliability and longevity. These properties collectively allow for improved control over charge dynamics, enabling the construction of devices with better synaptic performance and reduced power consumption^41,42,43,44.

A key advantage of 2D materials in neuromorphic devices lies in their diverse electronic properties, enabling various components within a single system. Graphene, for instance, is an excellent conductive channel due to its semi-metallic nature and high electron mobility⁴⁵. In contrast, TMDs, which range from semiconductors (e.g., MoS₂, WSe₂) to insulators (e.g., h-BN), could be tailored for different functional layers in memory and switching devices⁴⁶. These figure of merits of neuromorphic devices based on the 2D materials are summarized in Table 1. These versatile roles in neuromorphic systems allow 2D materials to mimic synaptic behaviors, such as plasticity, essential for simulating learning processes in artificial neural networks (ANNs). Beyond their inherent material properties, 2D materials can be stacked into vdW heterostructures, enhancing their versatility. By combining different 2D layers, researchers can design complex devices that integrate diverse functionalities, such as memory and sensing, in a single, scalable structure^47,48,49,50. These heterostructures can form interfaces with unique electronic properties, which are unattainable with traditional 3D materials. This reconfigurability opens various routes to innovative device architectures, such as vertical memristors that utilize ion migration and phase changes for reliable and multi-level resistance states, a crucial feature for neuromorphic applications.

Table 1 2D materials figure of merit comparison for neuromorphic devices

The fabrication of 2D material-based neuromorphic devices involves several important steps to ensure high performance and reliable operation. Making 2D materials, such as graphene⁵¹, TMDs, and h-BN, is done using different methods, each with its own benefits for neuromorphic applications. Chemical vapor deposition (CVD) is commonly used for large-scale production, creating even and high-quality thin films with controlled thickness. This method uses chemical reactions of gases at high temperatures on a surface, making it ideal for producing single-layer or few-layer 2D materials. Mechanical exfoliation, on the other hand, peels layers from bulk crystals using adhesive tape or other simple tools. While this method produces very pure and defect-free samples, it is usually limited to small areas. Another method, molecular beam epitaxy (MBE), allows for precise growth of layered materials in ultra-high vacuum conditions, giving excellent control over material composition and thickness. Each of these methods helps provide the necessary electronic and mechanical properties for neuromorphic devices⁷. That methods to prepare 2D thin film is depicted in Fig. 1.

Fig. 1: Schematic of various preparation methods for 2-dimensional thin films^{76,240,241,242,243,244}.

Reproduced from ref. ⁷⁶ with permission from IEEE, Copyright 2017. Reproduced from ref. ²⁴⁰ with permission from Springer Nature, Copyright 2023. Reproduced from ref. ²⁴¹ with permission from MDPI, Copyright 2021. Reproduced from ref. ²⁴² with permission from American Chemical Society, Copyright 2013. Reproduced from ref. ²⁴³ with permission from Elsevier, Copyright 2024. Reproduced from ref. ²⁴⁴ with permission from John Wiley and Sons, Copyright 2025.

The transfer of these materials onto functional surfaces is a key step in making devices because it affects how well the materials work together and how reliable the device is. For example, wet transfer is good for large-area production and involves coating the 2D material with a polymer layer (e.g., PMMA)^52,53, removing the original surface, and then placing the material on a new surface. However, this method can leave behind polymer residues and wrinkles. Dry transfer, using tools like PDMS stamps, reduces contamination and allows for accurate placement of 2D materials on the new surface. This accuracy is important for ensuring reliable synaptic-like behavior, as even small misalignments or defects can lower the device's performance. Transfer methods also make it possible to combine 2D materials with pre-made structures and build complex layered systems, which are essential for advanced neuromorphic devices.

Heterostructure fabrication, often using van der Waals stacking, creates interfaces that mimic biological synaptic interactions. Additionally, direct CVD growth of 2D material heterostructures provides a scalable approach for creating complex, multi-functional systems⁵⁴. Encapsulation, using h-BN, stabilizes the materials by protecting them from environmental degradation, ensuring stable performance for long-term learning and adaptation process^55,56. The integration of these materials with pre-fabricated substrates is achieved through lithography for precise patterning, metallization for electrode deposition, and annealing to optimize contact resistance and enhance device reliability.

However, several challenges remain in utilizing 2D materials effectively in large-scale neuromorphic devices. One significant obstacle is achieving uniform, large-area synthesis of monolayers, especially for complex heterostructures^10,57,58. Although chemical vapor deposition (CVD) and physical vapor deposition (PVD) have advanced, maintaining consistent quality across larger areas is challenging and costly. Additionally, synthesis methods often introduce grain boundaries or defects that impact electronic performance, complicating the integration of 2D materials in industrial-scale production^59,60,61,62. Further research also emphasizes the need to move beyond simple synaptic devices toward more complex networks that emulate biological neural circuits. Recent progress has focused on individual synapse-like memristors, while full neural circuits remain underdeveloped. To achieve full neural circuits, innovative material designs and scalable fabrication techniques are highly demanded to create 2D-based device arrays capable of complex information processing, like human brain functions^63,64,65,66.

This review will explore the foundational working principles of 2D materials-based neuromorphic devices, highlighting the mechanisms by which these materials mimic biological synaptic functions as described in Fig. 2. We will examine how various 2D materials enable the formation of artificial sensory systems, designed to emulate human sensory experiences like hearing, smell, taste, and touch. These developments are critical in advancing artificial neural systems, where the high sensitivity, selectivity, and low energy demands of 2D materials present unique opportunities. Additionally, we will address the integration of these devices within larger sensory arrays and neural networks^67,68,69,70, providing insights into their practical applications for real-world artificial sensory technologies. This overview will include recent progress in device structuring and the engineering challenges of integrating artificial sensory elements into cohesive, scalable neuromorphic systems^{48,71,72,73,74}.

Fig. 2: Operation mechanisms of neuromorphic devices based on 2D materials^{90,140,151,157,167,245,246,247,248}.

Reproduced from ref. ²⁴⁵ with permission from American Chemical Society, Copyright 2018. Reproduced from ref. ²⁴⁶ with permission from American Chemical Society, Copyright 2016. Reproduced from ref. ²⁴⁷ with permission from Springer Nature, Copyright 2021. Reproduced from ref. ²⁴⁸ with permission from Springer Nature, Copyright 2021. Reproduced from ref. ⁹⁰ with permission from American Chemical Society, Copyright 2015. Reproduced from ref. ¹⁴⁰ with permission from American Chemical Society, Copyright 2021. Reproduced from ref. ¹⁵¹ with permission from Elsevier, Copyright 2022. Reproduced from ref. ¹⁵⁷ with permission from Springer Nature, Copyright 2023. Reproduced from ref. ¹⁶⁷ with permission from Springer Nature, Copyright 2022.

Neuromorphic devices and working principles

In this chapter, the applications of 2D layered materials in neuromorphic devices, particularly the basic working principles and mechanisms, will be discussed. These materials are gaining attention for their unique electrical and physical properties, enabling advancements in non-volatile memory storage and logic functions. In memristive systems, materials like graphene and TMDs can effectively manipulate resistance states, leading to energy-efficient memory solutions. For neuromorphic computing, 2D materials facilitate the creation of artificial synapses and neurons that mimic biological processes, enhancing information processing and pattern recognition. However, challenges such as reproducibility and integration with existing fabrication technologies should be addressed. Despite these issues, ongoing research aims to unlock the full potential of 2D layered materials in transforming the future of computing technologies.

Memristor-based neuromorphic devices

In recent research by Weng and Zhao et al., memristor technology based on 2D layered nickel phosphorus trisulfide (NiPS₃) has demonstrated significant potential in neuromorphic computing, particularly for emulating synaptic plasticity and improving computational efficiency⁷⁵. The Ti/NiPS₃/Au memristor device, shown in Fig. 3a, features a robust architecture with titanium (Ti) as the top electrode, NiPS₃ as the active layer, and gold (Au) as the bottom electrode. As shown in Fig. 3b, their I-V characteristics exhibit a pronounced hysteresis loop during voltage sweeps from −2 V to 2 V, indicating stable resistive switching between high-resistance states (HRS) and low-resistance states (LRS), which is crucial for non-volatile memory applications. The compliance current (I_CC), adjustable from 100 μA to 1 mA (Fig. 3c), enables precise control over the set and reset processes, ensuring reliable operation for neuromorphic systems. Device stability and retention were examined through a 10,000-second retention test (Fig. 3d), where the memristor exhibited consistent current levels with minimal degradation, even under varying I_CC conditions. This stability is essential for applications requiring persistent memory states. Neuromorphic computing relies on mimicking biological synaptic plasticity, which the device achieves by simulating long-term potentiation (LTP) and long-term depression (LTD), as shown in Fig. 3e. At a pulse voltage of 1 V, the device demonstrates an increase in current up to 46.7 μA after repeated pulses (LTP), while LTD is observed at −1 V, with the current decreasing to 14.8 μA. This capability to adjust synaptic weights underpins learning and adaptability in neuromorphic systems. As shown in Fig. 3f, the plasticity of the device was investigated by varying pulse durations during LTP and LTD. For instance, at a pulse width of 0.1 ms and 0.8 V, the current increases to 128.7 μA after 50 pulses, compared to smaller increases at shorter pulse widths. This demonstrates the device's ability to regulate the rate of synaptic modifications, an essential feature for neuromorphic networks. Figure 3g highlights the repeatability of these synaptic behaviors, showing stable LTP and LTD cycles across multiple repetitions, critical for continuous learning without significant degradation. The memristor also allows fine-tuning synaptic weights by adjusting voltage amplitudes as shown in Fig. 3h. LTP voltages ranging from 0.75 V to 1.25 V and LTD voltages from −1.1 V to −1.22 V enable precise current modulation, vital for optimizing neural network performance. The application of this technology in neuromorphic image processing is illustrated in Fig. 3i, j. The memristor effectively performs edge detection on an input image (e.g., the "#NUS" logo, as shown in Fig. 3i) using horizontal and vertical Prewitt filters to generate a comprehensive feature map and the extended functionality to the classification of MNIST digits within a neuromorphic network of 400 input neurons and 100 hidden neurons depicted in Fig. 3j. The memristor facilitates analog computation and efficient signal processing, enabling low-power and high-performance pattern recognition tasks⁷⁶. By integrating these features, memristor-based systems bring neuromorphic computing closer to mimicking biological neural networks, marking a significant advancement in both computational efficiency and artificial sensory processing.

Fig. 3: 2D layered nickel phosphorus trisulfide-based memristic properties for neuromorphic computing.

a Schematic of the Ti/NiPS₃/Au memristor device. b I–V curves over 100 cycles with a compliance current (I_CC) of 500 μA. c I–V curves for different positive and negative voltage sweeps. d Retention of multiple conductance states. e LTP process modulated by varying pulse amplitudes. f LTP and LTD processes controlled by pulse width. g Identical pulse amplitude scheme for LTP/LTD. h Nonidentical pulse amplitude scheme. i Prewitt kernels (horizontal and vertical) programmed for edge detection. j Schematic of a three-layer ANN for MNIST pattern recognition. Reproduced from ref. ⁷⁵ with permission from John Wiley and Sons, Copyright 2024.

Dong and Hu et al. recently introduced a memristor device based on a Bi₂O₂Se nanosheet, showcasing significant promise for neuromorphic computing applications⁷⁷. The device is constructed on a Si/SiO₂ substrate, with the Bi₂O₂Se nanosheet sandwiched between Au bottom and Ag top electrodes, as illustrated in Fig. 4a. The cross-sectional TEM image confirms the precise fabrication of the device, revealing a uniform 50 nm Bi₂O₂Se layer and clean interfaces, essential for consistent switching behavior as depicted in Figure 4b^78,79. The memristor exhibits bipolar resistive switching, as seen in the I-V curves. It transitions between HRS and LRS at a threshold voltage of ±1.2 V, demonstrating non-volatile memory capabilities as presented in Fig. 4c. The device endurance is proven over 3500 cycles (Fig. 4d), with stable resistance states of approximately 10⁵ Ω (for HRS) and 10³ Ω (for LRS), and a switching speed in the nanosecond range, ideal for high-speed applications. Notably, it achieves ultrafast switching times (< 5 ns) and low switching energy (< 3.02 pJ), making it one of the most energy-efficient memristors reported to date as depicted in Fig. 4e. In neuromorphic computing, the Bi₂O₂Se memristor emulates synaptic behaviors such as synaptic plasticity and PPF⁸⁰. Figure 4f illustrates the memristive mechanism's analogy to biological synapses, with conductance changes representing synaptic weight modulation^81,82. Gradual conductance increases with repeated voltage pulses (Fig. 4g) mimic LTP, a feature critical for hardware-based learning algorithms. The PPF behavior (Fig. 4h) mirrors biological synapse responses, with an exponential decay in response strength as the pulse interval increases. The memristor further supports LTP and LTD, as shown in Fig. 4i, enabling adaptive learning in neuromorphic systems. This feature aligns with Hebbian learning principles, essential for implementing hardware neural networks capable of dynamic adaptation⁷⁹. Finally, the practical applicability of the Bi₂O₂Se memristor is demonstrated through its integration into a neural network trained on the MNIST dataset as depicted in Fig. 4j. The network achieved 91% accuracy after 100 training epochs, with hardware performance closely matching ideal software simulation results as shown in Figure 4k⁸³. These findings underline the Bi₂O₂Se memristor's potential for scalable and efficient AI hardware.

Fig. 4: Neuromorphic computing devices based on memrisitive 2D Bi₂O₂Se nanosheet.

a Schematic illustration of the Bi₂O₂Se device structure. b Cross-sectional TEM image showing conductive filaments (white dotted line). c–e Electrical characteristics of the Bi₂O₂Se memristor; (c) typical I − V hysteresis curve, (d) retention test at 0.01 V, (e) comparison of switching speed and energy across different memristors. f−k Emulation of synaptic behavior and ANN simulation; (f) schematic of a biological synapse, (g) conductance modulation by potentiation and depression pulses, (h) PPF as a function of pulse interval, (i) normalized conductance states with fitting curves, (j) schematic of an ANN for MNIST recognition, and (k) image recognition accuracy of the Bi₂O₂Se memristor versus the ideal case. Reproduced from ref. ⁷⁷ with permission from American Chemical Society, Copyright 2023.

Electrochemical redox reaction-based neuromorphic devices

Electrochemical neuromorphic devices, particularly electrolyte-gated transistors (EGTs), utilize the reversible motion of ions, such as Li⁺ or H⁺, under an applied electric field to modulate channel conductance, closely mimicking biological synaptic weight updates⁸⁴. This mechanism provides excellent linearity, multistate capabilities, and low power consumption, making EGTs highly suitable for neuromorphic hardware applications^85,86,87. Park and Kim et al. introduced a three-terminal 2D MoS₂ artificial synapse device with a lithium silicate (LSO) solid-state electrolyte thin film, designed for high-precision analog computing⁸⁸. The schematic in Fig. 5a illustrates the measurement setup, showing how Li-ion motion within the LSO electrolyte adjusts MoS₂ channel conductance to emulate synaptic functions. Charge transport mechanisms, depicted in Fig. 5b, demonstrate distinct behaviors across frequencies. Bulk electrolyte charging dominates at high frequencies due to the absence of an electric double layer (EDL). At lower frequencies, EDL formation at the MoS₂/electrolyte interface significantly affects capacitance. Pseudo-capacitance charging occurs at the weakest frequencies due to Li-ion intercalation into the MoS₂ channel, driving conductance changes via redox reactions⁸⁹. Gate bias experiments reveal how Li⁺ intercalation and extraction dynamically modulate channel conductance⁹⁰. Figures 5c–e illustrate the synaptic behavior of the device during LTP and LTD processes induced by gate voltage pulses. The device demonstrates high linearity, critical for neuromorphic computing, with low nonlinearity factors (v_p = −0.35, v_d = −2.35) ensuring precise and predictable synaptic weight modulation⁹¹. PPF is shown in Fig. 5f, highlighting short-term synaptic plasticity through decreasing facilitation indices as the interval between spikes increases. Decay times (τ₁ = 0.37 s, τ₂ = 3.52 s) align with biological timescales, further validating the device's ability to mimic real synaptic dynamics⁹¹. In neuromorphic computing simulations, the LSO-based EGT achieved recognition accuracy of 96.77% on the MNIST dataset using a multilayer network with 784 input-, 300 hidden-, and 10 output-neurons as shown in Fig. 5g, h. This result approaches the software baseline of 98%, illustrating excellent device efficiency for pattern recognition. Figure 5i shows consistent performance across seven devices, with a mean accuracy of 95.37% and a standard deviation of 0.68, demonstrating reliability and scalability for high-density systems^92,93. These findings highlight the potential of electrochemical EGTs with LSO electrolytes for next-generation neuromorphic hardware, offering low power consumption, high precision, and CMOS compatibility.

Fig. 5: Electrochemical characteristics of the lithium silicate electrolyte-gated MoS₂ synaptic transistor for neuromorphic computing.

a Schematic of the electrical measurement setup with MoS₂ and lithium silicate electrolyte. b Modulation of Li-ion states: pristine, EDL formation, Li intercalation, and extraction. c Measurement setup for synaptic characteristics. d LTP and LTD measured by sequential positive and negative pulses. e Linearity plot (v_p = −0.35, v_d = −2.35) with asymmetric ratio 0.32. f PPF index versus spike interval. g Neural network schematic for MNIST training. h Recognition accuracy over training epochs. i MNIST accuracy across seven devices with mean 95.37% and standard deviation 0.68. Reproduced from ref. ⁸⁸ with permission from American Chemical Society, Copyright 2022.

Ferroelectric synaptic devices

Ferroelectric semiconductor field-effect transistors (FeSFETs) based on α-phase In₂Se₃ have emerged as promising neuromorphic device candidates for next-generation functional electronics, including digital memory and brain-inspired computing circuits. Wang et al. demonstrated that the combination of ferroelectricity and semiconducting properties in α-In₂Se₃ enables multiple conductance states controlled by out-of-plane polarization, making these devices ideal for emulating synaptic behavior⁹⁴. Compared to traditional charge-trapping synaptic devices, FeSFETs offer enhanced controllability, faster learning, and seamless integration with gate dielectrics, positioning them as valuable components for neuromorphic computing applications^95,96,97. As illustrated in Fig. 6a, the device structure features a degenerate p-type Si substrate with a 50 nm Al₂O₃ dielectric layer and 15 nm Ti/65 nm Au electrodes. A 15 nm Al₂O₃ passivation layer enhances performance and stability, ensuring robust polarization switching critical for synaptic operation. The transfer characteristics at V_DS = 0.01 and 0.1 V display a significant hysteresis loop as V_GS sweeps from -10 to 10 V, as shown in Fig. 6b, attributed to reversible polarization switching in the α-In₂Se₃ channel. The resulting ferroelectric behavior provides an on/off current ratio exceeding 10⁵, with the hysteresis window expanding under higher gate potentials, enabling multiple programmable states⁹⁸. Figure 6c shows the conductance stability across 250 continuous sweep cycles, with HRS and LRS states demonstrating excellent endurance without degradation. Figures 6d–f highlight the synaptic functionality of the device, including excitatory and inhibitory responses. A negative gate pulse (6 V, 200 ms) induces a stable excitatory postsynaptic current (EPSC) lasting over 150 s, in contrast to a positive gate pulse generates an inhibitory postsynaptic current (IPSC), showcasing nonvolatile and versatile synaptic behavior⁹⁹. LTP and LTD are illustrated in Fig. 6g, applying 50 negative pulses (−5 V, 100 µs) and 50 positive pulses (2.2 V, 100 µs). These sequential pulses produce repeatable and stable potentiation and depression, mimicking synaptic plasticity. The LTP and LTD are characterized by a gradual increase and decrease of PSC, emulating the strengthening and weakening of synaptic connections¹⁰⁰. Synaptic weight modulation during LTP and LTD is further explored in Fig. 6h, with higher pulse amplitudes (−4 to −10 V for LTP; 3.3 V for LTD) inducing greater weight changes, enabling fine-tuned conductance adjustment. Figure 6i demonstrates spike-timing-dependent plasticity (STDP), a critical feature for learning processes. Following the Hebbian learning rules, the weight changes depend on the interval (Δt) between pre- and post-synaptic spikes. Positive Δt results in potentiation, while negative Δt causes depression, with time constants of 2.1 ms and 1.2 ms for negative and positive Δt, respectively, reflecting biological synaptic behavior. These findings underscore the potential of α-In₂Se₃ FeSFETs for neuromorphic computing, combining precise conductance modulation, long-term stability, and compatibility with complex learning algorithms.

Fig. 6: Ferroelectric neuromorphic device characteristics and performance of α-In₂Se₃-based FeSFETs.

a Diagram of the FeSFET structure featuring an α-In₂Se₃ channel within a ferroelectric setup. b Schematic of the upward and downward polarized state of α-In₂Se₃ channel, and energy band diagram. c Hysteresis in transfer characteristics attributed to ferroelectric polarization. d Stability of conductance in high and low resistance states across 250 cycles. e Non-volatile behavior observed in the excitatory postsynaptic current. f Stable modulation of inhibitory postsynaptic current (IPSC). g Synaptic responses of long-term potentiation (LTP) and long-term depression (LTD) to varying gate pulses. h Modulation of synaptic weight under different amplitudes during LTP and LTD processes. i Illustration of spike-timing-dependent plasticity (STDP), demonstrating changes in synaptic weight based on the spike timing interval. Reproduced from ref. ⁹⁴ with permission from John Wiley and Sons, Copyright 2020.

Kwon et al. introduced ferroelectric neuromorphic devices based on ultrathin tin monosulfide (SnS) films, showcasing their potential for analog synaptic emulation¹⁰¹. These devices utilize robust in-plane ferroelectric switching and exhibit excellent linearity in LTP and LTD, making them ideal for energy-efficient and high-density neuromorphic computing. The spontaneous polarization switching in SnS enables precise and reversible control of synaptic weights under external electric fields, mimicking the biological synaptic learning and memory functions¹⁰². The lateral ferroelectric device structure was illustrated in Fig. 7a using integrated Pt-interdigitated electrodes for ferroelectric switching measurements. The I-V characteristics as shown in Fig. 7b reveal distinct ferroelectric switching peaks around ±3 V, indicating effective polarization reversal. This multistate conductance capability underpins the device's ability to replicate synaptic functions like potentiation and depression. Additionally, endurance test (Fig. 7c) shows stable resistance states across 10⁵ cycles, demonstrating the reliability of the ferroelectric device for long-term operation without significant degradation¹⁰³. Figure 7d, e highlight the dynamic adjustment of synaptic weights in response to presynaptic spikes, with the device exhibiting EPSC behavior like neuronal action potentials. This response underscores the nonvolatile nature of the SnS synaptic device, which retains memory states even after stimuli are removed, ensuring energy-efficient performance. The device further mimics short-term synaptic behavior, as shown in Fig. 7f, where PPF demonstrates dynamic synaptic strength adjustments. Figures 7g and 6h illustrate consistent postsynaptic current (PSC) responses under varying stimulation conditions, with conductance proportional to the number of presynaptic spikes. This adaptability reflects the complex temporal and spatial learning patterns crucial for neuromorphic systems¹⁰⁴. The potential of SnS-based ferroelectric devices is validated through ANN simulations for pattern recognition tasks using the MNIST dataset, where they achieve high accuracy, demonstrating their practicality for neuromorphic computing applications. Ferroelectric neuromorphic devices stand out for their precise and tunable synaptic weight adjustments, which closely emulate neural processes. They exhibit excellent endurance, stability, and dynamic learning capabilities while maintaining low power consumption. Continued research on optimizing ferroelectric properties and integration techniques is expected to address scalability challenges, advancing brain-inspired computing systems capable of addressing the limits of conventional AI hardware.

Fig. 7: Ferroelectric neuromorphic computing device based on SnS thin films for synaptic emulation.

a Schematic of the fabricated SnS device with Pt-interdigitated electrodes used to measure electrical properties like I-V and P-E curves. b schematic of polarization switching between Sn and S. c I-V characteristics showing ferroelectric switching peaks near ±3 V, indicating typical ferroelectric behavior. d Resistance stability over multiple cycles during program/erase testing, demonstrating the device’s endurance. e Comparison of biological synapses with artificial synapses, highlighting similar operational principles. f Excitatory postsynaptic current (EPSC) response to a single presynaptic spike, indicating synaptic-like behavior. g Paired-pulse facilitation (PPF) index as a function of time interval between consecutive spikes, showing short-term synaptic enhancement. h, i Postsynaptic current response with varying numbers of presynaptic spikes (N = 10, 30, and 50), demonstrating spike-dependent plasticity. Reproduced from ref. ¹⁰¹ with permission from American Chemical Society, Copyright 2020.

Floating gate-based neuromorphic devices

Floating-gate neuromorphic devices are emerging as efficient candidates for neuromorphic computing due to their tunable synaptic weight control, high linearity, and reliability. Tang and Zhang et al. reported a two-terminal floating-gate memory (2TFGM) device composed entirely of 2D materials, including monolayer MoS₂ as the channel, few-layer graphene (FLG) as the floating gate, and h-BN as the tunneling layer¹⁰⁵. This structure minimizes power consumption while enabling precise conductance control, crucial for high-speed neuromorphic applications^64,106,107. The device achieves an ultra-high on/off ratio of ~10⁸, with symmetric switching hysteresis loops observed for h-BN thicknesses of 7, 10, and 15 nm. By varying programming voltages, up to 131 distinguishable conductance states are achieved, demonstrating excellent retention over 10⁴ seconds and endurance across 10⁴ cycles even at elevated temperatures up to 400 K (Fig. 8b–d). These multilevel states enable gradual synaptic weight adjustments, essential for accurate neuromorphic processing^108,109. LTP and LTD are replicated using programming and erasing pulses, producing 3000 quasi-continuous conductance states with transient responses that mimic EPSC and IPSC. Further tuning of pulse amplitudes optimizes the balance between linearity and conductance density, reducing energy consumption while maintaining high computational accuracy as shown in Fig. 8f–h. In MNIST pattern recognition tasks, the device achieves a recognition accuracy of 97.7%, closely matching the 98% software baseline, showcasing its high-resolution neural computation potential, as shown in Figure 8i¹¹⁰. These results demonstrate the potential of all-2D 2TFGM devices in neuromorphic systems, combining precision, scalability, and low power consumption, offering significant advancements for ANNs.

Fig. 8: Floating-gate neuromorphic device based on an all-2D 2TFGM structure.

a Schematic of the device with monolayer MoS₂, h-BN tunneling layer, and FLG floating gate. b I-V hysteresis loops showing symmetric switching with h-BN thicknesses (7, 10, 15 nm). c Band diagrams of programming and erasing processes. d Multilevel conductance switching under various programming voltages is essential for synaptic behavior. e Stable conductance retention over 10⁴ seconds with minimal drift. f Tuning of synaptic weights with up to 400 conductance states. g Potentiation and depression curves demonstrating high linearity. h MNIST recognition accuracy reached 97.7%. i Error rate decreases with increased conductance states, supporting precise neuromorphic computing. Reproduced from ref. ¹⁰⁵ with permission from John Wiley and Sons, Copyright 2021.

The synaptic function can be emulated by purely electronic charge trapping/de-trapping effects without microstructure change^111,112. It makes charge trapping/de-trapping-based synaptic devices promising in terms of reliability, while other synaptic devices based on diffusive dynamics show device performance variation due to inevitable microstructure change¹¹³. Therefore, there have been many reports on synaptic devices based on various materials using charge trapping/de-trapping effects^{114,115,116,117}. For 2D materials, various 2D vdW structures have been investigated to emulate synaptic functions, which harness the trapping or de-trapping electrons in a charge-trapping interfacial layer as the synaptic weight control layer. Seo et al. reported a vdW synaptic device with h-BN/WSe₂ heterostructures, in which a weight control layer was formed on h-BN with an O₂ plasma treatment¹¹⁴. The conductance of the h-BN/WSe₂ heterostructures was controlled to modulate the synaptic weight by the charges trapped in the weight control layer under electrical and optical stimuli. Owing to the charge trapping/de-trapping effects, the h-BN/WSe₂ structures exhibit various synaptic dynamics including LTP/LTD and STDP. The tunable charge trapping behavior can be achieved from the native oxidation of 2D materials¹¹⁸. For example, a native InO_x layer can form on InSe under ambient conditions due to the instability of InSe. The native oxide layer acts as an efficient charge-trapping layer, facilitating dynamic control of charge transport. Under a positive (negative) gate bias, electrons are trapped (released), resulting in a reduction (enhancement) of the conductance of the InSe channel, thereby mimicking an EPSC response. The charge trapping/de-trapping dynamics contribute to nonvolatile memory characteristics and enable the demonstration of key synaptic behaviors, such as short-term plasticity and long-term plasticity.

With charge trapping/de-trapping effects, it is possible to integrate all human memory types (sensory memory, short-term plasticity, and long-term plasticity) in one system. Zhou et al. reported a single vdW heterostructure of MoS₂/h-BN/graphene/h-BN for complex time-tailoring requirements of human memory by stacking volatile and nonvolatile function layers as shown in Fig. 9a, b¹¹⁷. When a negative voltage is applied to the top gate electrode, electrons are injected from the electrode into the graphene through the h-BN tunneling layer illustrating in Fig. 9c. As the h-BN tunneling limits the retention of these injected carriers, the device shows a volatile behavior of anticlockwise hysteresis. On the other hand, a positive (negative) bias at the back gate drives electron (hole) tunneling into the Al₂O₃ charge trapping layer, and the accumulation of electrons in the Al₂O₃ layer, causing a positive shift of the threshold voltage. It makes the device show non-volatile characteristics with a clockwise memory window, as the trapped electrons remain stable over a long period as shown in Fig. 9d. Due to the volatile and nonvolatile characteristics based on charge trapping/de-trapping effects, the MoS₂/h-BN/graphene/h-BN vdW heterostructures exhibit flexible memory transition and time-tailoring capabilities, enabling them to emulate various stages of human memory. When the frequency of top gate bias increases from 2 Hz to 10 Hz, it was observed to transition from sensory memory to short-term memory with stronger intensity and longer memory (Fig. 9e). Notably, sensory memory cannot switch to long-term memory under top-gate bias, originating from volatile characteristics. Figure 9f shows the measured current triggered by the stimulus from the back gate. It was found that 3 successive input pulses from the back gate trigger the transition from short-term memory to long-term memory with extended retention time. This approach effectively emulates a human memory system, such as forgetting, learning, and memorization using the unique properties of 2D materials and charge trapping/de-trapping effects.

Fig. 9: Charge trapping/de-trapping based synaptic transistor.

a, b Schematic and a false-color SEM image of the device composed of the MoS₂/h-BN/graphene/h-BN vdW heterostructure. Schematic illustration for device operating mechanism of (c) volatile characteristics and (d) nonvolatile characteristics. e The transition from sensory memory to short-term memory with increasing the top gate bias pulse frequency. f The short-term memory and long-term memory of the device depend on the number of the back gate pulse. Reproduced from ref. ¹¹⁷ with permission from John Wiley and Sons, Copyright 2019.

2D heterostructured neuromorphic devices

Recent studies have highlighted the advantages of 2D material heterostructures in enhancing the performance of neuromorphic devices. MoS₂/h-BN heterostructures, for instance, leverage the complementary properties of a semiconducting MoS₂ and an insulating h-BN layers, enabling superior synaptic plasticity and low-power consumption. Similarly, graphene/h-BN heterostructures consists of the highly conductive graphene and insulating and chemically stable h-BN demonstrating the faster switching speeds and enhanced device stability. These heterostructures offer distinct advantages, such as tailored electronic properties and multi-functional integration, that are crucial for the development of high-performance 2D material-based neuromorphic devices. Mahata et al. introduced a heterostructure-based neuromorphic device using a layered palladium diselenide (PdSe₂) and aluminum oxide (Al₂O₃) redox layer as shown in Figure 10a¹¹⁹. The sandwiched Al₂O₃/PdSe₂/Al₂O₃ device structure is proposed for the effective oxygen ion migration based memristive operation. As presented in Fig. 10b, the cross-sectional HR-TEM image shows the fabricated layered structure of the memristor using well-oriented PdSe₂ with interlayer spacing of 0.439 nm. This device exploits oxygen ion migration at the Al₂O₃/PdSe₂ interface to enable precise synaptic weight modulation via conductive filament formation and rupture under applied electric fields. The dynamic resistive switching behavior, as shown in Fig. 10c, demonstrates stable transitions between HRS and LRS, mimicking synaptic plasticity essential for neuromorphic applications. The device exhibits excellent endurance, maintaining stable performance over 800 cycles (Fig. 10d), highlighting its potential for long-term operation^120,121. By controlling RESET voltages, nearly stable multistate memory is achieved (Fig. 10e), allowing fine-tuned synaptic weight adjustments. Figure 10f illustrates the device capability to retain distinct memory states as a function of reset voltage, critical for complex neuromorphic tasks. During potentiation, positive voltage pulses incrementally increase synaptic weight, closely emulating the strengthening of biological synaptic connections (Fig. 10g). The learning and forgetting processes are depicted in Fig. 10h, where conductance gradually decreases after the removal of potentiation pulses but can be partially recovered by applying depression pulses of varying amplitudes¹²². Final conductance recovery levels, reaching up to 99.4%, depend on the amplitude of forgetting pulses (Fig. 10i)¹²³. Multilevel conductance states, as shown in Fig. 10k, highlight the device ability to emulate complex neural processes with fine-grained conductance control¹²⁴. These properties support both short-term and long-term plasticity, enabling the device to replicate adaptive memory behavior fundamental to neuromorphic computing. The integration of Al₂O₃ enhances device reliability, scalability, and compatibility with high-density hardware applications, making this PdSe₂-based memristor a promising platform for future neuromorphic systems. Table 2 summarizes the features and figure-of-merits of 2D materials based artificial synapses reported in recent studies, including their structure, dimension, mechanism, synaptic characteristics, and energy consumptions.

Fig. 10: Heterostructured neuromorphic device featuring a PdSe₂ memristor layered with Al₂O₃.

a Diagram illustrating the ITO/Al₂O₃/PdSe₂/Al₂O₃/TaN synaptic structure with applied bias settings. b Cross-sectional HRTEM image displaying the layered structure of the memristor, with a detailed inset showing the interlayer spacing within the PdSe₂ layers. c Bipolar resistive switching behavior of the ITO/Al₂O₃/PdSe₂/Al₂O₃/TaN memristor, measured at compliance currents of 100 and 500 µA during SET operations. d Durability test results for the PdSe₂-based memristor, monitored at a read voltage of +0.2 V. e Switching dynamics as influenced by varying V_RESET from 0.7 V to 1.6 V. f Achievement of nearly stable ten high-resistance states (HRS) by modulating V_RESET at a read voltage of 0.2 V. g Potentiation characteristics using 20 presynaptic pulses at +3.5 V/1 ms and synaptic weight depression from 0 to 2.0 V/1 ms with pulse intervals of 5 ms. h Detailed view of the initial potentiation and depression cycles under varying conditions. i PPF index as a function of the interval (Δt) between two consecutive pulses. j Characterization of potentiation and depression in the artificial synapse using 100 consecutive pulse trains at three sets of pulse amplitudes: +3.0 V/1.4 V, +3.2 V/1.6 V, and +3.4 V/1.8 V. k Curve fitting of the gradual normalized conductance changes observed during potentiation and depression phases. Reproduced from ref. ¹¹⁹ with permission from Elsevier, Copyright 2024.

Table 2 Structure, dimension, mechanism, synaptic characteristics, and energy consumptions for various artificial synapses based on 2D materials that have been reported in recent studies

Artificial sensory

2D materials have emerged as promising candidates for developing artificial sensory systems owing to their unique atomic structures and novel electronic properties^{37,41,125,126,127}. The atomically thin nature of 2D materials offers advantages such as mechanical flexibility, low operating voltage, and high energy efficiency^128,129,130. In addition, 2D materials show superior sensitivity and selectivity to target stimuli due to their large active sensing area and tunable electrical properties, making them ideal platforms for developing various artificial sensory synaptic devices^{37,131,132,133}. The following sections mainly introduce recent progress of various sensory synaptic devices based on 2D materials.

2D materials-based artificial sensory synapse

2D materials-based artificial visual synapses

Artificial visual synapses, inspired by the biological synapses in the human visual system, which collect over 80% of information from the environment¹³⁴, have attracted significant interest due to their ultrafast computing speed, low crosstalk, high bandwidth, and excellent energy efficiency^135,136,137. To build an artificial vision system, optical detection, signal processing, and memory are fundamental requirements. Artificial visual synapses can perform visual perception, signal processing, and data storage within a single device. Among the available materials, 2D materials have been considered an ideal platform for artificial visual synapses due to their novel physical and chemical properties, particularly their tunable bandgap, strong light-matter interactions and atomically thin, dangling-bond-free surfaces^138,139. With these intriguing properties of 2D materials, many researchers have proposed various approach to develop artificial visual synapses and its applications for neuromorphic computing^{57,114,140,141,142,143,144}.

optoelectronic transistor with optical tunable ferroelectric switching is one approach to emulate visual synaptic function. Ge et al. reported an optoelectronic synapse consisting of MoS₂/BaTiO₃ heterostructure, where MoS₂ is the light-sensitive channel and BaTiO₃ is the ferroelectric gate (Fig. 11a)¹⁴². This configuration enables the MoS₂/BaTiO₃ optoelectronic transistor to exhibit light-tunable synaptic behavior, allowing it to dynamically adjust its conductance depending on the light dosage and light history. As shown in Fig. 11b, The MoS₂/BaTiO₃ optoelectronic transistor shows multi-level conductance state and more conductive states under a longer pulse duration time. Under the illumination of 100 pulses of light with various wavelengths (405, 532, and 650 nm), the device shows four orders of magnitude in the channel conductance, indicating the high sensitivity of the MoS₂/BaTiO₃ optoelectronic transistor to blue light (Fig. 11c). By applying an electrical pulse to the BaTiO₃ back gate, the MoS₂/BaTiO₃ optoelectronic transistor decreases channel conductance with the increasing electrical pulse duration (Fig. 11d). These results indicate that the MoS₂/BaTiO₃ optoelectronic transistor shows LTP and LTD behavior under optical and electrical pulse, respectively, which is capable of being applied to neuromorphic computing. With programmed optical pulse and electrical pulses, the MoS₂/BaTiO₃ optoelectronic transistor successfully exhibits multi-level conductance with long retention time by optical potentiation and electrical depression (Fig. 11e).

Fig. 11: Optoelectronic synaptic devices based on 2D materials.

a Schematic of the MoS₂/BaTiO₃ artificial visual synapse. b Optically controlled conductance of the device under the illumination with various durations of optical pulses at 450 nm. c Wavelength-dependent conductance variation of the device. d Electrically modulated conductance changes of the device. e The non-volatile multi-level conductance switching under programmed optical and electrical pulses. Reproduced with permission from¹⁴². Copyright 2021 Elsevier. f Schematic illustration of the photoelectroactive artificial synapse structure and bidirectional synaptic weight updating process of the MoS₂/h-BN heterostructure. Energy band diagrams for (g) LTP and (h) LTD process. i Negative V_th shift and (j) positive V_th shift characteristics controlled by optical stimuli under negative and positive gate bias, respectively. k LTP and LTD characteristics of the photoelectroactive synaptic device. Reproduced from ref. ¹⁴³ with permission from Springer Nature, Copyright 2021.

For most artificial visual synapses, optical stimuli are limited to use as an excitatory spike pulse for the potentiation process rather than an inhibitory spike for the depression process, constraining artificial visual synapses in practical applications. For optically tunable potentiation and depression, Park and co-workers suggested a photoelectroactive synapse based on the MoS₂/h-BN heterojunction, where the potentiation and depression of synaptic weight were controlled by optical pulses under negative and positive gate bias, respectively (Fig. 11f)¹⁴³. The MoS₂ channel, with few surface defects is the active layer for conducting electrical signals. The h-BN layer, containing inherent photoelectroactive defects, is crucial for storing and updating synaptic weights. Figure 11g, h illustrate the synaptic weight-updating mechanisms of the LTP and LTD process in the MoS₂/h-BN heterojunction. When negative gate bias is applied along with optical spikes, the photo-excited electrons from inherent defects of h-BN transfer to the MoS₂. These excited electrons transfer to the MoS₂ channel, creating positively charged defects in the h-BN, resulting in an increase in the surface potential as well as the channel conductance of MoS₂, a process known as LTP. For the LTD process, the excited electrons fill the inherent defects in the h-BN under a positive gate bias along with optical spikes, neutralizing the positively charged defects and simultaneously reducing the surface potential of MoS₂. It leads to a reduction in the conductance of the MoS₂ channel. As shown in Fig. 11i, j, the MoS₂/h-BN heterojunction shows the shift of threshold voltage (V_th) from −2.1 to −23.9 V and from −23.9 to −1.5 V in the reverse sweep (+30 V → −30 V) and the forward sweep (−30 V → +30 V) of the gate bias under the optical spike pulse (405 nm, 65 mW cm^-2). Notably, it was found that higher photon energy enhanced the neutralization process, resulting in a pronounced V_th shift in the MoS₂/h-BN heterojunction. Additionally, Fig. 11k presents the dynamic LTP and depression characteristics, where the synaptic conductance steadily increases when optical spikes were consecutively applied at a gate bias of −16 V, and gradually decreases to the initial conductance level when the optical spikes were applied at a gate bias of 30 V, corresponding to LTP and LTD characteristics of the photoelectroactive synapse, respectively#. These results confirm the multistate memory capabilities and artificial visual synaptic function of the MoS₂/h-BN heterojunction under electrical and optical stimulation due to the ionization and neutralization of photoelectroactive defects.

Fully optical-modulated synaptic devices, which use only optical pulses rather than electrical pulses to modulate conductance, provides an opportunity to build a neuromorphic computing for visual information processing, recognition, and memory in a flexible and energy-efficient way. Zhang et al. reported large-scale and flexible two-terminal optical synaptic devices based on graphdiyne/graphene/PbS quantum dot (Pyr-GDY/graphene/PbS-QD) heterostructure (Fig. 12a)¹⁴⁰. The Pyr-GDY/graphene/PbS-QD heterostructure can emulate both excitatory and inhibitory synaptic behaviors through optical stimulation without the need for electrical stimulation to control its conductance. In the Pyr-GDY/graphene/PbS-QD heterostructure, Pyr-GDY acts as a photoresponsive hole trapping layers, graphene serves as the conductive channel, and PbS QDs acts as a photoresponseive layer, trapping electrons under illumination. Figure 12b illustrates the mechanism of fully optical modulation synaptic behavior of the Pyr-GDY/graphene/PbS-QD heterostructure. Under the illumination at 980 nm, the PbS quantum dots absorb the light and generate electron-hole pairs, with electrons being trapped in the PbS QDs and holes moving into the graphene layer. This increases the conductance of the graphene, resulting in a positive photoresponse analogous to EPSC. In contrast, when the device is exposed to an incident light with wavelength of 405 nm, Pyr-GDY mainly generates electron-hole pairs. In this case, while holes are trapped in Pyr-GDY layer, electrons move into the graphene, decreasing its conductance and creating and emulating IPSC. Therefore, the Pyr-GDY/graphene/PbS-QD heterostructure exhibits LTP and LTD behavior with excellent linearity and symmetricity of the weight update trajectory under repetitive optical pulses of 980 nm light and 405 nm light, respectively, as shown in Fig. 12c. Based on the fully optical modulation synaptic behavior, visible information sensing-memory-processing system based on a 7 × 6 optical synapse array constructed from the Pyr-GDY/graphene/PbS-QD heterostructure (Fig. 12d). Initially, all the synapses are in a low-conductance state. After applying 150 successive optical stimuli (980 nm), the system detects and memorizes a visual input, represented by the letter "G." The synaptic conductance differences between irradiated and non-irradiated areas remain stable for over 1000 s, demonstrating in situ image memorization. The system can also distinguish between different images by comparing the input with a reference image and calculating a distinction output, allowing it to differentiate between the letter "O" (reference) and "S" (unknown). This integrated functionality eliminates the need for separate memory and processing units, reducing energy consumption and latency in artificial visual systems.

Fig. 12: Fully light-modulated synapse based on the Pyr-GDY/Gr/PbS-QD heterostructure.

a Schematic of the artificial visual synapse based on the Pyr-GDY/graphene/PbS-QD heterostructure. b Energy band diagram of the Pyr-GDY/graphene/PbS-QD heterostructure under 450 nm and 980 nm illumination. c LTP and LTD characteristics of the Pyr-GDY/graphene/PbS-QD heterostructure under 150 consecutive 980 nm optical pulses for potentiation and 150 consecutive 450 nm optical pulses for depression. d Conductance map of a 7 × 6 artificial visual synapses array at the (i) initial state, (ii) after 150 optical training (980 nm, 150 mW cm^-2, 20 ms) and (iii) after 1000 s at the end of training. Reproduced from ref. ¹⁴⁰ with permission from American Chemical Society, Copyright 2021.

2D materials-based artificial tactile synapses

Tactile perception allows us to detect texture, pressure, vibration, temperature, and pain, playing an essential role in our daily life. Inspired by this natural system, artificial tactile synapses have been developed to emulate the neural functions of touch in neuromorphic devices^145,146,147. Artificial tactile perception systems offer opportunities for the development of robotics, human-machine interfaces, and haptic feedback systems. To realize artificial tactile system, various strategies have been proposed to mimic mechannoreceptors in skins, capacitive¹⁴⁸, piezoelectric¹⁴⁹, and triboelectric methods¹⁵⁰. Among these, capacitive methods utilize deformation-induced capacitance changes to achieve high sensitivity and rapid response, suitable for applications such as flexible electronic skins and human-machine interfaces¹⁴⁸. Piezoelectric approaches employ strain-induced polarization charges in non-centrosymmetric materials, enabling functionalities like spike-timing-dependent plasticity and dynamic temporal filtering¹⁴⁹. Triboelectric systems harness contact electrification effects to convert mechanical stimuli into electrical signals, providing energy-efficient, self-powered tactile sensing capabilities¹⁵⁰.

To integrate tactile sensing and processing into each single devices, many researchers have employed 2D materials based synaptic transistors and triboelectric nanogenerators (TENGs) to emulate the biological synapses^151,152,153. Li et al. reported a touch-modulated vdW heterostructure based InSe floating-gate device with self-writing power for tactile synaptic simulation¹⁵¹. As shown in Fig. 13a, the device consists of a floating-gate synaptic transistor with an InSe/h-BN/graphene vdW heterojunction integrated with a TENG (Al/Kapton/Silicone rubber/Al). This structure allows for charge storage and release upon touch stimulation, enabling the device to emulate tactile synaptic plastic behaviors. Figure 13b illustrates how positive and negative separation distances (D₊ and D₋) between the TENG layers influence the carrier density and band structure of the device. The electrostatic potentials from TENG cause charge injection or release in the graphene floating gate, shifting the V_th of the InSe transistor. As a result, the device exhibits reliable and complex memory and tactile synaptic behaviors driven by the touch-modulated charge injection/release. The device exhibits EPSC and IPSC response when negative and positive displacement was applied, respectively. The duration of the touch also plays a critical role, as longer touch pulses lead to greater synaptic weight changes. Figures 13c, d present how the device emulates synaptic plasticity through repeated touch stimulation from the TENG. As the number of input spikes increases, the EPSC and IPSC progressively transition from short-term plasticity to long-term plasticity, simulating LTP and LTD, respectively. Under consecutive touch stimulations, the device exhibits the sustained LTP and LTD states with gradual strengthening or weakening of synaptic weights (Fig. 13e).

Fig. 13: Touch-modulated van der Waals heterostructure based InSe floating-gate transistor integrated with a TENG for synaptic simulation.

a Schematic of the artificial tactile synapse based on the InSe/h-BN/graphene heterostructure integrated with a TENG (b) Schematic illustration of the operating mechanism of the InSe/h-BN/graphene heterostructure integrated with a TENG under touch stimulation. c EPSC and (d) IPSC of the triggered by InSe/h-BN/graphene heterostructure integrated with a TENG under a sequence of various displacement pulses. e LTP and LTD of the InSe/h-BN/graphene heterostructure integrated with a TENG under 100 consecutive displacements. Reproduced from ref. ¹⁵¹ with permission from Elsevier, Copyright 2022.

Wang et al. also reported artificial tactile synaptic devices, integrating a floating-gate MoS₂ synaptic transistor with a TENG (Fig. 14a)¹⁵². The MoS₂ transistor uses Au nanoparticles as a floating-gate layer, trapping and de-trapping electrons across the HfO₂ tunneling layer. The MoS₂ synaptic transistor with a TENG exhibits mechanoplasticity, resulting from charge injection and release controlled by triboelectric potentials under displacement pulses. Due to the mechanoplasticity, the device emulates biological synaptic functions such as short-term memory and long-term memory, as well as complex neural processes like facilitation and depression, all controlled through displacement pulses. Based on the excellent mechanoplasticity of the MoS₂ synaptic transistor with a TENG, a neural network capable of logic switching was built, where three artificial synapse transistor, synapse A (S_A), synapse B (S_B), and mechanoplastic synapse M (S_M) are connected in parallel (Fig. 14b). The mechanoplastic synapse, driven by the TENG, modulates the synaptic weights and allows for the realization of logic gates like AND and OR, without the need for complex CMOS circuits. By varying the mechanical input to the TENG, the synaptic transistor network can perform basic logic operations. The sum of the postsynaptic currents from different logic combinations (e.g., "0,0," "1,1") can be used to generate AND and OR gate functions (Fig. 14c, d). These results demonstrate distinct logic truth tables based on mechanical stimuli, providing proof of concept for neuromorphic computing applications.

Fig. 14: Mechanoplastic tribotronic floating-gated neuromorphic MoS₂ transistor.

a Schematic of the artificial tactile synapse based on the mechanoplastic triboelectric foliating-gated neuromorphic MoS₂ synaptic transistor. b Circuit diagram of a neural network composed of three artificial synapse transistors for logic switching. The resultant postsynaptic current corresponds to the eight combination ways of the logic input signals, demonstrating the (c) “AND” logic and (d) “OR” logic. Reproduced from ref. ¹⁵² with permission from John Wiley and Sons, Copyright 2020.

2D materials-based artificial olfactory synapses

Artificial olfactory synaptic devices mimic the brain’s neural processing of smells, enabling AI of things with human-like perception, personalized healthcare to diagnose diseases, and environmental monitoring to identify potential hazards. 2D materials have shown great potential for gas sensing applications due to their large specific surface areas, superior surface activities^154,155,156. Using these advantages, there have been reports of artificial co-sensory systems, including olfactory perception. Although many efforts have been made to develop artificial olfactory synapses response to chemical pulse using 2D materials^157,158,159, they remain in the early stages of development and there have been issues to overcome due to their complexity.

Jiang et al. reported that artificial visual-respiratory synapse with monolayer oxidized Ti₃C₂T_x MXene (VRSOM). The hydroxyl concentration and airflow rate influence synaptic weight of VRSOM, providing insights for designing artificial olfactory synapses¹⁵⁸. Figure 15a, shows the forgetting process in the human brain, influenced by respiratory dynamics. Similarly, the VRSOM shows the decrease of PSC under hydroxy pulse and stable PSC after the hydroxy pulse. This result indicates the potential of VRSOM as an artificial olfactory synapse to simulate biological forgetting and memory processes. The VRSOM shows the faster decay of PSC with increasing hydroxyl concentration and flow rate, showing transition from long-term memory to short-term memory (Fig. 15b). It is attributed that more hydroxyl groups bind to oxidized vacancies in the Ti₃C₂T_x MXene, which opens pathways for photocarrier dissipation. As shown in Fig. 15c, under successive light and hydroxyl pulse stimulation, the VRSOM exhibits stable and reproducible LTP and LTD, respectively. These results demonstrate cross-modal updating of synaptic weights utilizing a combination of chemical atmospheric and photonic plasticity. This approach, multisensory integration of environmental signals, can serve as a proof-of-concept for developing advanced artificial olfactory systems.

Fig. 15: Perception of hydroxyl signals of artificial visual-respiratory synapse with monolayer oxidized Ti₃C₂T_x Mxene.

a Schematic diagram of respiration-induced excitability in the human brain and the attenuation of postsynaptic current of artificial olfactory synapse with monolayer oxidized Ti₃C₂T_x MXene (VRSOM) under hydroxyl pulses. b The depression of the device’s synaptic weight depends on the hydroxyl concentration in flowing air. c Repetitive cyclic LTP and LTD characteristics of the VRSOM for memory and forgetting process Reproduced from ref. ¹⁵⁸ with permission from John Wiley and Sons, Copyright 2020.

2D materials-based multi-mode artificial sensory synapses

The human brain is capable of processing individual sensory inputs while also combining multisensory information through specialized neurons located in the primary sensory regions, enabling more efficient and rapid responses within the nervous system¹⁵⁹. Likewise, multi-sensing and multi-mode in a single synaptic device improve the ability of AI and robotics to perceive and recognize the surrounding environment accurately, enhancing information collection and process efficiency. Therefore, there have been remarkable advancements in multi-mode sensing and perception of visual, auditory, and tactile information using artificial sensory systems, such as visual-tactile artificial synapse^160,161,visual-olfactory artificial synapse^162,163, and multisensory neural network with cross-modal integration and recognition^162,163.

2D materials, owing to their unique optoelectronic and mechanical properties as well as versatility, have emerged as an ideal platform for multi-mode artificial sensory synapse. For example, Gong et al. reported multimodal mechano-photonic synaptic device based on 2D asymmetric ferroelectric heterostructure¹⁶⁴. The mechano-photonic synaptic device consists of an asymmetric MoS₂/CuInP₂S₆ (CIPS) ferroelectric heterostructure field-effect transistor integrated with a TENG to supply triboelectric potential for gating, programming, and plasticizing (Fig. 16a). As depicted in the equivalent circuit diagram in Fig. 16b, the light pulse illuminated to the MoS₂ channel and mechanical displacement applied to the TENG can act as presynaptic inputs, inducing the postsynaptic current and updating the synaptic weight. The device demonstrates optoelectronic functionality, as illustrated in Fig. 16c, where light pulses (405 nm) induce EPSC, effectively emulating the memory retention and forgetting behaviors of biological systems. Furthermore, the EPSC increases as both displacement and light intensity increase, demonstrating precise tunability of the synaptic weight through TENG displacement and optical illumination (Fig. 16d). The multimodal capabilities are further exemplified in Fig. 5h, which demonstrates light-induced potentiation and triboelectric depression, enabling dynamic modulation of synaptic weights. These results collectively underline the potential of 2D materials-based devices for neuromorphic applications, particularly in AI and bionic electronics, offering low-energy, versatile solutions for replicating complex human sensory and memory functions. The multimodal capabilities are further explored in Fig. 16e. It was observed that 100 consecutive light pulses induce excitatory behaviors, while 100 consecutive mechanical displacements result in inhibitory behaviors. These results demonstrate the dynamic modulation of synaptic weights through light-induced potentiation and triboelectric depression, effectively emulating the multimodal synapse properties.

Fig. 16: 2D materials-based multi-mode artificial sensory synapse.

a Schematic of the mechano-photonic synaptic device based on an asymmetric MoS₂/CIPS ferroelectric heterostructure field-effect transistor integrated with a TENG. b The equivalent circuit diagram of a mechano-photonic synaptic device. c EPSC of the MoS₂/CIPS heterostructure field-effect transistor integrated with a TENG under a 405 nm light pulse. d The synergistic modulation of EPSC by mechanical displacement (from the TENG) and optical power density. e light-induced potentiation and triboelectric depression behavior of the MoS₂/CIPS synaptic device. Reproduced from ref. ¹⁶⁴ with permission from John Wiley and Sons, Copyright 2024. EPSC modulation of the MXene/VP vdW heterojunction by 10 consecutive light pulses under (f) various gas atmosphere and (g) various relative humidity concentrations. h The recognition and (i) forgetting process of the letter ‘X’ pattern in various relative humidity concentration. Reproduced from ref. ¹⁶⁵ with permission from Springer Nature, Copyright 2024.

Ma and coworkers reported MXene/violet phosphorus (VP) vdW heterojunctions for visual-olfactory multimodal perception¹⁶⁵. The hierarchical design, spanning from nanoscale heterojunctions to hybrid films with a large specific surface area, enhances photoelectric response and gas sensitivity, enabling visual-olfactory multi-modal perception. The vdW heterojunctions generate a built-in electric field that promotes the separation and transport of photogenerated carriers, leading to a persistent photoconductivity effect that emulates neuromorphic functionalities such as EPSC, STP, and LTP. In addition, the large specific surface area of the 2D materials allows for the efficient adsorption of gas molecules, which modulate the optoelectronic response of the device through charge transfer with the nanosheets and changes in the heterojunction barrier. Figure 16f–i shows the ability of MXene/VP vdW heterojunction to emulate visual-olfactory cross modal perception, where environmental factors such as gas type and relative humidity significantly influence synaptic behaviors. The increased EPSC observed in higher humidity environments, coupled with faster memory decay, mimics the retrieval-strength trade-offs characteristic of human cognition. These results indicate the potential of 2D materials for advancing neuromorphic sensory applications in robotics, artificial vision, and sensory augmentation systems.

Advanced integrated artificial sensory

Bio-plausible sensory information processing

Biological sensory systems exhibit remarkable capabilities in perceiving and processing environmental stimuli. These systems have evolved sophisticated mechanisms for adapting to diverse conditions, efficiently encoding information, and performing complex computations. In recent years, the field of neuromorphic devices has developed arrays or matrices using pixel unit devices to build highly efficient sensory information processing systems that emulate such abilities of biological sensory systems. In this section, we will explore diverse studies that have suggested novel advanced sensory information processing systems, inspired by individual phases or integrated steps of the biological sensory systems.

Receptor to cortex

Afferent neurons, the primary conduits of sensory information, carry signals from receptors to the central nervous system. As these signals propagate, they undergo significant preprocessing, including lateral inhibition and adaptation, which enhance contrast and adjust sensitivity. This preprocessing extends to preliminary feature detection, where neurons selectively respond to specific stimulus characteristics before reaching the primary sensory cortex¹⁶⁶. Upon arrival at the cortex, local feedback circuits further refine and interpret the sensory data. These cortical networks implement computations that reduce redundancy and enhance neural efficiency, integrating top-down influences from higher brain areas to incorporate experience, attention, and expectations with current sensory input. By mimicking the efficient coding strategies of biological sensory systems, from initial transduction to higher-level cortical computations, these technologies aim to create more responsive and adaptable artificial sensing systems.

In addition to fundamental synaptic functions introduced in the previous chapters, a key feature of biological sensory systems is their remarkable capacity for environmental adaptation, which is the procedure of adjusting sensitivity in response to varying intensities of external stimuli. This adaptive mechanism, is ubiquitous across diverse sensory modalities, with notable examples including light adaptation in visual systems¹⁶⁷, nociceptor inactivation in pain perception¹⁶⁸, and olfactory adaptation in smell detection¹⁶⁹. Visual adaptation, for instance, modulates light sensitivity based on average light intensity, alternating between rod and cone cells for dim and bright light conditions, respectively. This adaptive response significantly reduces the subsequent processing burden on the brain through automatic response normalization. A MoS₂ phototransistor matrix was developed to emulate such human visual adaptation and provides a wider light intensity dynamic range than state-of-the-art image sensors (Fig. 17a)¹⁶⁷. Analogous to the role of photoreceptors in the human visual adaptation, the MoS₂ phototransistor with charge trapping and de-trapping mechanism achieved a similarly broad light detection range. Table 3 summarizes the features and synaptic functions of 2D materials-based artificial sensory synapses reported in recent studies, including the utilized 2D materials, device structure, synaptic functions, performances, and energy consumption^{141,142,153,170,171,172,173,174,175,176,177}. The gate voltage-dependent response emulates the role of horizontal cells in switching between rod and cone cells (Fig. 17b); for example, under a positive gate bias, the device behaves like cone cells, while under a negative bias, it functions similarly to rod cells. Moreover, the MoS₂ phototransistor matrix demonstrated visual adaptation, including both scotopic and photopic responses, by adjusting its photosensitivity and illumination time under different gate biases. Tested with MNIST images in both dim and bright light conditions, it achieved a high recognition rate of over 96% through its adaptive process (Fig. 17c). Building upon this concept, chromatic adaptation has been realized through the implementation of a bilayer PtSe₂ transistor-based retinomorphic device with wavelength-dependent nonvolatile photoconductivity¹⁷⁸. This device leverages the distinct energy levels of different wavelengths of light and the physisorption and desorption of O₂ molecules under low and high energy conditions, respectively. As a result, PtSe₂ forms nonvolatile negative photocurrent under blue and green illumination and positive photocurrent under red illumination, enabling red-blue and red-green antagonistic adaptation. The contrast enhancement capabilities of this chromatic adaptation mechanism were experimentally verified, demonstrating its potential for improving visual processing in artificial systems. Pseudocolor images with the wavelength-dependent response of the device were compared to three other types of images with the magnitude of red, green, and blue, respectively. Pseudocolor images fed into a LeNet-5 structured convolutional neural network (CNN) produced classification accuracies of over 90%, which is a significant improvement above other image under 75%.

Fig. 17: Bio-plausible sensory information processing mimicking receptor to cortex.

a Schematic of visual adaptive MoS₂ phototransistor and optical image of single phototransistor. Scale bar, 100 µm; S source, D drain, BG bottom gate. b Gate voltage V_G resulting in same photo sensitivity S_ph by each illumination intensity P_in. c Top: illustration of phototransistor array and an ANN for image recognition. Bottom left: the simulated scotopic adaptation result of MNIST image ‘8’. Bottom right: the simulated photopic adaptation result of MNIST image ‘8’. Reproduced with permission from ref. ¹⁶⁷. Copyright 2022, Springer Nature. d Illustration of neuromorphic visual system for static image and motion trace pre-processing. e Demonstration of self-denoising ability by each wavelength of image ‘U’. First column: noisy input at σ = 0.3. Second column: scanning procedure. Third column: output image immediately after scanning. Fourth column: output image 60 seconds after scanning. f Illustration of in-sensor trace extraction based on distinct current intensity by time. g Sample images of trajectory image by each step. Reproduced ref. ¹⁹⁵ with permission from Springer Nature, Copyright 2023.

Table 3 Summary of structure, synaptic functions, and energy consumptions for various artificial sensory synapses based on 2D materials that have been reported in recent studies

In visual receptive fields, composed of combinations of ON and OFF cells in the retina, contribute to various levels of signal processing, from low-level processing such as contrast enhancement to high-level processing like object detection. Over 50 different types of visual neurons contribute to diverse levels of signal processing¹⁷⁹, removing redundant information and enhancing the efficiency of rapid visual processing. Table 2 summarizes the features and synaptic functions of 2D materials based artificial sensory synapses reported in recent studies, including their structure, synaptic functions, performances, and energy consumptions^{68,114,117,180,181,182,183,184,185,186,187,188,189,190}. Object detection relies on identifying object segments through edge detection, a process enhanced by edge enhancement and contrast correction techniques. These methods typically employ diverse convolutional operations, with the differential-of-Gaussian (DoG) kernel, resembling a Mexican hat, being a popular choice for image filtering. While conventional approaches apply these operations post-image acquisition, necessitating additional data storage and transfer, biological visual systems integrate this processing at an earlier stage. The mammalian retina utilizes ON and OFF bipolar cells, responsive to bright and dark light respectively, to form a natural DoG kernel. This biological structure, where one cell type constitutes the center of the receptive field and the other forms an antagonistic surround, facilitates innate edge detection and image processing. Inspired by this biological model, a retinomorphic vision sensor using a WSe₂/h-BN/Al₂O₃ heterostructure phototransistor enabled near-data processing, where the visual information can be both sensed and processed simultaneously⁵⁷. The device operates as ON bipolar cells at a gate voltage bias of 0 V and as OFF bipolar cells when the gate voltage is negative (Vg < 0 V). By implementing artificial receptive field matrices, where photoresponsivity of multiple pixels was controlled individually with independent gate voltage biases, and various image processing techniques such as edge detection of moving objects, image stylization, edge enhancement, and contrast correction have been demonstrated.

While conventional noise reduction methods such as Gaussian smoothing filters, median filters, wavelet denoising, and deep learning-based approaches require considerable computation^{191,192,193,194}, the retina accomplishes this task with remarkable efficiency. Temporal coding preserves and enhances temporal information, reinforcing past information while mitigating noise (Fig. 17d). It reinforces information that is similar to the past information and cancels out noisy information. Non-volatile memory device which can cancel out the noise and maintaining informative pixels was emulated with a NbS₂/MoS₂ phototransistor array¹⁹⁵. The device memorizes patterns through LTP based on the gate-tunable transfer of photoinduced charge carriers at the NbS₂/MoS₂ heterointerface, enabling differentiation between persistent patterns and transient noise. Pixels corresponding to memorized patterns undergo 'relearning', while noise pixels only achieve STP due to insufficient stimulation. The significant responsivity difference between LTP and STP states facilitates effective pattern recognition with contrast enhancement, under stronger light intensity, shorter intervals between successive light pulses, and a greater number of pulses maintaining signals longer. This image noise reduction device improved pattern recognition rates from 78% (unenhanced) to 90% (enhanced). Building upon this work, a multi-wavelength self-denoising device, incorporating UV-to-visible sensitive bilayer MoS₂ channel and IR sensitive PtTe₂/Si gate electrode, was developed¹⁹⁶. This device added wavelength-dependent responsivity for color discrimination, successfully denoising input patterns with multi-colored inputs of different wavelengths. The array maintained the denoised pattern memory for at least 1200 s for single-colored input and 60 s for multi-colored inputs (Fig. 17e), which may be the enough timescale for image preprocessing before moving on to the next process, demonstrating the potential of biomimetic approaches in advancing sensory processing technologies.

Motion detection is a more complex process that involves processing vast amounts of three-dimensional video data in real-time, a feat remarkably accomplished by the human visual system. This process requires both spatial and temporal information for effective object tracking. Temporal coding plays a crucial role in this process, retaining memory of the location and existence of past stimulations, thereby enabling the brain to follow an object's trajectory^197,198. Recent advancements in biomimetic sensor technologies have made significant strides in replicating this capability and demonstrate the potential to improve data processing efficiency and simplify system designs compared to traditional CMOS image sensors, which rely on successive data transmission. A NbS₂/MoS₂ hybrid film-based phototransistor, capable of processing spatiotemporal data, successfully traced the trajectory of a moving spotlight over time¹⁹⁵. The time-dependent decay of the photocurrent enabled the sensor to directly process and generate spatiotemporal information, with each pixel able to store the data in real-time and in parallel (Fig. 17f). This system achieved over 92% accuracy in trajectory registration across diverse motion datasets (Fig. 17g), effectively mimicking the temporal coding mechanism of biological systems. In a parallel development, MoS₂ phototransistors demonstrated potential for object detection across a wide velocity range¹⁹⁹. The MoS₂ phototransistors exhibit a volatile, exponentially decaying response to light stimulation due to the shallow charge trapping in MoS₂ and produce graded current responses like non-spiking graded neurons of insect vision systems. Such non-spiking behavior without a refractory period in spiking neurons enhances data transmission bandwidth and enables agile motion perception in the tiny vision system of insects. This characteristic enables efficient temporal information encoding, functionally like the temporal coding observed in biological systems despite the different neural architecture. In addition, depending on the speed of motions, a temporal resolution of the device was modulated from 10¹ to 10⁶ ms by applying gate voltage bias. A 20-by-20 matrix configuration of these photosensors demonstrated remarkable object tracking capabilities across velocities ranging from 0.005 km/h to 50 km/h, with the ability to detect objects moving in various directions (up, down, left, right, approaching, and leaving).

Network level

The sensory cortex plays a pivotal role in processing and interpreting information from the physical world, receiving electrochemical signals via afferent neurons and performing complex cognitive functions that underpin perception and recognition²⁰⁰. This sophisticated capability stems from two key features: a hierarchical structure and the capacity for learning and memory. The hierarchical organization of the sensory cortex facilitates a progressive, multi-stage approach to signal processing, primarily through weighted sum operations across various levels. Concurrently, its learning and memory capabilities enable the analysis of environmental stimuli and, crucially, allow for the dynamic adjustment of synaptic connectivity based on accumulated experiences. This adaptive process, fundamentally defining intelligence, has inspired the scientific and engineering communities, leading to the conceptualization and development of ANNs²⁰¹. These biomimetic systems aim to emulate the brain's ability to process information, learn from experiences, and adapt to new stimuli, marking a significant step towards machine intelligence. By modeling the hierarchical structure and plasticity of the sensory cortex, ANNs represent an ambitious attempt to replicate aspects of biological intelligence in artificial systems, potentially revolutionizing our approach to data processing, pattern recognition, and decision-making in various technological domains. Traditionally, ANNs have been predominantly applied in software, which, while effective, often incurs substantial computational costs in terms of signal transfer and data storage. To address these limitations, researchers have begun exploring innovative hardware-based approaches that integrate memory and sensing units directly with synaptic components capable of exhibiting plasticity akin to the weights in conventional ANNs²⁰². In this sub-section, we will delve into the diverse landscape of ANNs that leverage 2D material-based devices, examining their architectures, functionalities, and the advantages they offer over traditional software-based neural networks.

The multi-layered perceptron (MLP) represents an early model designed to emulate the hierarchical structure and learning capabilities of biological neural systems. At its core, the MLP consists of perceptrons, fundamental units comprising operational and activation subunits. The operational subunit performs linear transformations on multiple inputs, mirroring the weighted signal accumulation at the axon hillock, while the activation subunit normalizes or binarizes the signal, analogous to the threshold-dependent activation of postsynaptic neurons. These perceptrons, arranged in layers, form the MLP network, which can perform intelligent tasks such as classification and regression through a learning process that involves weight tuning. Recent advancements have seen the integration of MLPs with 2Dmaterials and devices, exemplified by the application of an h-BN/WSe₂ heterostructure as a color-discriminative optical sensor and tunable synaptic device for pattern recognition¹¹⁴. The near-ideal nonlinearity of 1.5 for both LTP and LTD of the device were employed to construct a double-layered perceptron for colored and color-mixed pattern recognition and the network achieved accuracy rates exceeding 90% after the 50^th epoch of training for an RGB-colored MNIST dataset. Subsequently, WSe₂ optoelectronic devices have been applied for physically implementing these neural network structures, demonstrating their potential for both supervised and unsupervised learning tasks²⁰³. The system has N pixels with M subpixels as illustrated in Fig. 18a, and subpixels with the same color are connected to the same node in the network to generate M outputs. The physical realization involves a 3 × 3 matrix of pixels, each subdivided into three sub-pixels and trained through bias voltage adjustment. In classifier implementation, sensor responses were applied to the input layer, weighted sums of the input data were realized as the device response and off-chip activation functions were applied via software programs (Fig. 18b). The weights were trained through backpropagation. The classifier achieved 100% accuracy across various noise levels (σ = 0.2, 0.3, 0.4) for the alphabet 'n', 'v', and 'z' classification task. In addition, in autoencoder implementation, the device responses were applied to the off-chip activation function and decoder network (Fig. 18c). The network not only correctly reconstructed noisy images (σ = 0.15, 0.3) but also reduced the noise of the original image. Moreover, the integration of sensor and processor limited the operation time to photocurrent generation, enabling real-time classification within ~50 ns. The autoencoder effectively demonstrated the significant potential of these advanced MLP implementations in improving AI system efficiency and capability.

Fig. 18: Hierarchical neural network-inspired sensory information processing.

a Schematic of WSe₂ optoelectronic device array for ANN implementation. ‘n’ indicates the pixel number and ‘M’ indicates the number of subpixels. b Illustration of ANN for classification. R_MN indicates the weight at m^th subpixel of nth pixel. Double-network perceptron (dashed box) is the device implementation, and the rest of the network performs off-chip activation function. c Top: illustration of ANN for autoencoder. Device implementation is identical to that of the (b), and the rest of the network performs off-chip activation and reconstruction functions. Reproduced from ref. ²⁰³ with permission Springer Nature, Copyright 2020. d Optical image of WSe₂/h-BN/Al₂O₃ heterostructure device. e Illustration of device array, analogous to the receptive field. Total output current I_Output is the sum of drain-source current change ΔI_ds. f Illustration of CNN implementation classifying alphabet ‘n’ (blue), ‘j’ (yellow), and ‘u’ (green). k indicates training number and (i, j) denotes the location in the device array. Reproduced from ref. ⁵⁷ with permission from AAAS, Copyright 2020. g Schematic of multi-terminal floating gate memristor. V_FG is floating gate voltage, and V₁ to V₅ indicate each electrode. h Illustration of LIF model. Membrane potential V_FG is compared to V_th by comparator circuit. i Optical image of device array. j Illustration of device array circuit for unsupervised learning of MNIST dataset. Reproduced from ref. ²⁰⁵ with permission Springer Nature, Copyright 2023.

The field of AI, particularly in the domain of computer vision, has long sought inspiration from biological systems to enhance machine learning capabilities. The human visual system, with its high efficiency in processing and interpreting complex visual information, serves as a prime model for these efforts. Image recognition in the biological visual system begins with the processing of optical stimuli by localized receptive fields in the retina, followed by hierarchical analysis in the visual cortex to extract meaning from the perceived images. This natural process has inspired the development of CNNs, which emulate the biological approach by employing 2-dimensional convolution operations in the early stages of their architecture. An advancement towards bioinspired ANNs was demonstrated through the simulation of MXene-based memristive synapses emulating CNN functionality²⁰⁴. The MXene synapses exhibited efficient weight update capabilities through their LTP and LTD properties, with low nonlinearity ratios of 2.3 and 1.9 for LTP and LTD, respectively. By applying those parameters of the MXene devices to each kernel and weight in an eight-layered CNN architecture, the system achieved an image recognition rate of 87.5%, which approaches the 91.1% recognition rate obtained by an ideal CNN model for the CIFAR-10 dataset, a benchmark in image classification tasks. Further progress was made in bridging the gap between artificial and biological visual systems through the experimental implementation of a WSe₂/h-BN/Al₂O₃ heterostructure retinomorphic image sensor (Fig. 18d)⁵⁷. As introduced in the earlier section (receptor to cortex) the sensor's ability to change its activation mode between ON and OFF light responsive states based on the applied gate voltage is analogous to the behavior of ON and OFF bipolar cells in the retina, key components in early visual processing. As illustrated in Fig. 18e, the matrix of devices with distinct states generates a kernel that corresponds to the receptive field in retina. To demonstrate the capabilities of this bioinspired system, a CNN with a 3 × 3 input image was implemented, where the retinomorphic image sensor directly converted light illumination into electrical current responses. These responses were then utilized to classify images through the network, with the weights of the sensors determined by the applied gate voltage (Fig. 18f). The system achieved 100% recognition accuracy in fewer than 10 training epochs for the three alphabet letters ('n', 'j', and 'u'). This rapid and accurate learning demonstrates the potential of bioinspired hardware to significantly enhance the efficiency and performance of ANNs.

Despite exceptional perception abilities that surpass human perception, their significant energy consumption limits their practical application. The remarkable energy efficiency of biological neural networks, primarily attributed to their sparse and binary coding mechanisms, has inspired the development of spiking neural networks (SNNs) as a promising approach for low-power, low-latency computing in fields such as robotics and edge computing. SNNs aim to emulate the event-driven nature of biological neural systems, with the leaky integrate-and-fire (LIF) model serving as their fundamental unit and spike timing dependent plasticity (STDP) as the primary learning rule. The LIF model, a simplified mathematical representation of neural spiking behavior, operates by accumulating input signals until a threshold is reached, at which point it generates a spike output. Hardware implementation of this model using multi-terminal memristors and memtransistors was demonstrated (Fig. 18g)²⁰⁵. The core structure consists of a monolayer graphene floating gate, an h-BN tunneling insulator, and a MoS₂ semiconductor channel. In this configuration, inputs from multiple electrodes on the MoS₂ layer, analogous to presynaptic signals from multiple neurons, collectively determine the potential of graphene floating gate and the current response of MoS₂ channel, like a biological membrane potential. The device with efficient blocking of spontaneous release of trapped charges at the floating gate demonstrated excellent modulation of synaptic weight in terms of linearity, symmetry, long-term retention, cyclic endurance and energy efficiency. In addition, when the thickness of h-BN tunneling insulator is thinner enough to tunnel out the trapped charges, the device emulated the leaky nature of the neuronal membrane. The LIF neuron model is realized by linking the floating gate voltage to a comparator biased at the threshold voltage (Fig. 18h). The system, incorporating artificial LIF neurons and synapses (Fig. 18i), demonstrated STDP learning and successfully simulated unsupervised learning on unlabeled MNIST datasets, achieving a noteworthy maximum accuracy of 83.08% (Fig. 18j). This work not only demonstrates the feasibility of hardware-based SNNs but also highlights their potential for efficient, brain-inspired computing in various applications requiring low energy consumption and rapid response times.

Artificial synapses for brain-like cognitive functions

Recent advancements in neuromorphic engineering have drawn inspiration from biological systems to develop innovative array-type devices that enhance the efficiency of ANNs and process information in a way similar to the human brain, aligning with principles from neuroscience^{17,206,207,208}. This emerging field explores specialized circuits with neurobiological architecture, which mimic the remarkable capabilities of nature, such as sound localization in barn owls and multisensory integration across various species. By emulating these biological processes, researchers have developed artificial neurons capable of integrating multiple sensory inputs, including visual, chemical, tactile, and even subjective cues. These innovations not only contribute to the enhancement of machine learning performance but also pave the way for more sophisticated AI that can better navigate complex, real-world environments. From audiomorphic computing inspired by barn owls to visuochemical neurons mimicking Heliconius butterflies²⁰⁹, these bio-inspired approaches are pushing the boundaries of what's possible in artificial sensory processing and decision-making. In this chapter, we will explore studies that suggest specialized circuits and that can mitigate current limitations of ANNs.

Sound localization is one of the achievements of evolution, under resource-constrained environments. Using neurons which can fire once in a few milliseconds, the auditory system must process the signal within hundreds of microseconds. Thus, auditory system undergoes such a challenging computational task. Remarkably, barn owls mitigated such issues using specialized neural circuits (Fig. 19a). An audiomorphic computing device inspired by the neural map that exists in barn owl was suggested for sound localization²⁰⁹. According to the Jeffress model (Fig. 19b), barn owls have time delay neurons and coincidence neurons in their neural maps. Time delay neurons fire in order starting from each cochlear nucleus when the sound is given. Coincidence neurons, which are connected to corresponding time delay neurons, fire only when concurrent spikes arrive at matching delay neurons. Through this neural map, barn owls can localize a sound source with a precision of 1-2 degrees. Inspired by this ability, neuromorphic devices capable of identifying the location of a sound source was proposed. A MoS₂ FET composed of multiple split-gates with different nanoscale separation distances functioned as a coincidence neuron (Fig. 19c). The split gate FET operates only when both gates in a pair are activated mimicking the behavior of a biological coincidence neuron (Fig. 19d). Time constant tunable RC circuits, emulating the time delay neurons, are individually connected to each split-gates with designed resistance values (channel resistance of MoS₂ FETs) according to spatial arrangement of the gates. Using simulation and implementing a specialized learning mechanism for their computing device, the authors demonstrated its performance as a sound localization device.

Fig. 19: Artificial synapses for brain-like cognitive functions.

a Illustration of auditory cortex of the barn owl. b Illustration of the Jeffress model. c Optical image of audiomorphic MoS₂ split-gated FET. d Activation of coincidence audiomorphic split-gated FET. Each color corresponds to the color of the split gate in (c). ITD, interaural time difference. Reproduced from ref. ²⁰⁹ with permission from Springer Nature, Copyright 2019. e Illustration of the integration between tactile and optical sensory information. f Schematic of mechano-photonic graphene/MoS₂ optoelectronic transistor with TENG; (i) Optical image of the optoelectronic transistor. Scale bar, 5 µm. (ii) Illustration of the charge transfer of optoelectronic transistor. (iii) Illustration of image recognition using mechano-photonic device. Reproduced from ref. ¹⁶¹ with permission AAAS, Copyright 2021. g Butterfly-inspired visuo-chemical platform. Signals from MoS₂ memtransistor visual neuron and graphene chemo receptor neuron are sent through MoS₂ memtransistor mating circuit to be processed with AND or OR gate. h Schematic of AND gate with four different chemical stimuli (C1 – C4) and four different color stimuli (Blue, Green, Orange, and Red). i Activation of visuo-chemical neuromorphic AND gate. j Schematic of OR gate with same stimuli. k Activation of visuo-chemical neuromorphic OR gate. Reproduced from ref. ²¹⁷ with permission from John Wiley and Sons, Copyright 2023.

Multisensory integration is a fundamental process in biological systems that enhances perception, particularly to complement weak sensory stimuli or enhance or speed up the decision making by including additional information. This phenomenon is observed across a wide range of species, from insects to humans, playing a crucial role in survival and environmental interaction. In insects, such as fireflies²¹⁰, bees²¹¹, moths²¹², and butterflies^213,214, multisensory integration facilitates mate selection and navigation. Birds also utilize multisensory integration for various behaviors, including foraging and predator avoidance²¹⁵. Recent studies have sought to emulate these biological processes in artificial systems. Building upon these insights, researchers have developed more sophisticated artificial neurons that more closely mimic biological multisensory integration by incorporating visual, chemical, tactile, and even subjective sensory inputs.

In humans, multisensory integration is particularly evident in the interplay between visual and tactile perception (Fig. 19e). In low-light conditions, tactile information supplements visual perception, enhancing overall spatial awareness. Conversely, in well-lit environments, visual cues inform tactile responses, such as determining the appropriate grip strength for object manipulation. A mechano-photonic neuron that combines a TENG with a graphene/MoS₂ heterostructure was demonstrated to achieve synergistic mechanical and optical plasticity (Fig. 19f)¹⁶¹.The triboelectric potential generated by mechanical displacement in the TENG made a gating effect that controlled charge transfer within the heterostructure, modulating optoelectronic synaptic properties of the graphene/MoS₂ phototransistor. Programming and erasing were achieved through distinct stimuli, i.e., optical inputs for programming and mechanical inputs for erasing. When implemented in an ANN, this approach improved image recognition accuracy up to 92%, highlighting the potential of multisensory integration in enhancing machine learning performance. Also, a visuotactile spiking neuron was implemented with a MoS₂ phototransistor, a triboelectric tactile sensor, and a MoS₂ memtransistor-based spike encoding circuit²¹⁶. This artificial neuron exhibited three key features of biological multisensory integration: super-additive response, inverse effectiveness, and temporal congruency. In this way, the device successfully emulated multisensory integration occurring in the human brain.

Heliconius butterflies integrate visual and chemical information to identify suitable mates, demonstrating the synergistic effect of multiple sensory inputs. A combination of a monolayer MoS₂ phototransistor-based visual sensor and electrochemical graphene transistor-based chemo-sensor, both connected to a non-volatile floating-gated MoS₂ memtransistor-based comparator neuron, is implemented to emulate visuochemical neuron of Heliconius butterfly (Fig. 19g)^209,217. The comparator neuron converted signals from visual and chemical sensors, triggered by different visual and chemical stimuli, into digital output signals. In addition, the decision-making circuit performed more complex operations based on the combination of specific visual and chemical stimuli using AND and OR gate configurations (Fig. 19h-k). With the similar approach, a gustatory circuit in feeding behavior that integrates physiological (hunger) and psychological (appetite) cues, emulating complex human decision-making processes have been developed²¹⁸. The circuits were integrated with graphene-based chemitransistors (gustatory taste receptors) to detect the five basic tastes (sweet, bitter, sour, salty, and umami) and MoS₂ memtransistor-based logic circuits to process this information. This system incorporates both objective (physiological) and subjective (psychological) factors in decision-making, exemplified by modeling the feeding circuit (F) as a function of hunger (a physiological factor, H) and appetite (a psychological factor, A) (${\rm{F}}={\rm{H}}+{\rm{A}}\bar{{\rm{H}}}$). The feeding circuit is activated when an individual is either hungry or has an appetite, even if not physiologically hungry ($\bar{{\rm{H}}}$). This approach represents a significant step towards artificial systems that can make more nuanced, human-like decisions by considering both objective data and subjective factors. The integration of these multisensory artificial neurons into larger networks holds promise for developing more sophisticated and adaptable AI systems that can better navigate complex, real-world environments.

Outlook and perspective

2D materials-based neuromorphic devices have emerged as the next-generation information processing units based on their various electronic characteristics, such as excellent charge carrier mobility and various switching functionalities. In this regard, we reviewed neuromorphic devices, particularly for artificial synapse systems from single units to more integrated circuit-level applications, which utilize various 2D materials including graphene, MoS₂, WSe₂, h-BN, and so on^33,62,68,219. Neuromorphic devices with different working mechanisms such as memristor, electrochemical, and ferroelectric have been developed and combined to become a fundamental building block of advanced neuromorphic artificial sensory systems that process optical, mechanical, and chemical information in a bio-plausible way These devices have expanded beyond individual components to encompass comprehensive system-level implementations replicating biological sensory pathways^16,50,67, from initial signal detection through hierarchical neural networks and specialized circuits. Such architectures present compelling alternatives to traditional von Neumann computing paradigms, delivering enhanced processing speed while significantly reducing power requirements.

The development of artificial sensory neuromorphic devices using 2D materials faces critical challenges, including limited material stability under harsh operational conditions, scalability constraints in large-area device integration, and variability in synaptic performance due to fabrication inconsistencies. To address these issues, a multidisciplinary approach through material innovation is required, for example, the synthesis of defect-engineered or alloyed 2D materials for achieving enhanced stability and tailored electronic properties, as well as advancements in device architecture, such as hybrid integration with robust encapsulation techniques to mitigate environmental degradation^{23,25,28,44,68}. The foundation of efficient data transmission in these devices relies on van der Waals heterostructures, enabling precise vertical integration of 2D materials for energy-efficient, compact sensory networks²²⁰, enabling precise vertical integration of 2D materials for energy-efficient, compact sensory networks with minimal contact resistance and efficient charge transfer. Spike-based signal conversion transforms analog outputs into temporal binary patterns, providing noise immunity and reduced bandwidth through temporal encoding while enabling asynchronous communication²²¹. Essential intermediate adaptation circuits include voltage/current amplifiers, impedance matching networks, and signal conditioning units to bridge component gaps and optimize sensor outputs^222,223,224. The integration of hybrid 2D-CMOS microchips extends functionality beyond traditional van der Waals heterostructures, advancing device performance and applications^225,226.

Recent advances in 2D material-based artificial sensory systems have shown significant progress, from simple single-layer perceptrons to SNNs^{57,114,203,204,205}. However, these devices face practical limitations due to their basic architecture, especially when compared to biological sensory systems that process information through complex, multi-layered neural networks. The primary challenge is implementing deep neural networks (DNN) that can effectively transmit and process data across multiple layers using nonlinear activation functions, similar to biological systems. Researchers in optics^{227,228,229,230} and electronics^231,232 have addressed this by using intrinsically nonlinear devices. To address these challenges, researchers across diverse fields, including optics^{227,228,229,230} and electronics^231,232 have approached this challenge by utilizing intrinsically nonlinear devices, marking a significant step forward. For example, researchers have been demonstrated external circuitry involving transimpedance amplifiers and analog-to-digital converters incorporated with neuromorphic devices, enabling the hardware implementation of multilayer neural networks built with memristors^{220,221,222,223,224,225}. Also, the integration of nanoelectric devices with nonlinear functionalities providing excellent scalability, remarkably low power consumption, and the potential for high-density integration. In addition, a recent ReLU device developed with Cu/p(V3D3-co-VI)/Al₂O₃/Al memristors proved successful physical implementation of nonlinear functions using neuromorphic devices, which not only validates the theoretical frameworks but also demonstrates the practical viability of 2D material-based neuromorphic computing systems^{220,221,222,223,224,225,226}.

The interdisciplinary effort to bridge the gap between artificial sensory devices and the sophisticated architecture of biological systems represents a crucial advancement toward developing more capable and practical sensory technologies that could revolutionize fields from robotics to medical diagnostics. However, several common technological bottlenecks emerge, suggesting opportunities for cross-domain solutions. The primary challenges include operational speed limitations for multi-modal and large dataset processing, hardware integration complexities involving 1T1R structures, power regulators, and signal processing circuits (including filters and amplifiers) which are essential for stable system operation, and device uniformity issues in large-scale fabrication. These shared challenges can be addressed through cross-domain solutions: the three-dimensional integration of functional units via device stacking as well as wafer-scale fabrication and transfer methods offer promising approaches to enhance operational speed, latency, device uniformity and yield in mass production^{233,234,235,236}. For user-specific applications (such as customized chatbots in AI and robotics, and individual medical devices), localized data pre-processing and lightweight network architectures capable of real-time adaptation are crucial, particularly given the limitations of current data-driven learning approaches that require extensive computational resources and large training datasets. To this end, specialized neuromorphic device matching algorithms and corresponding hardware configurations are essential. The optimization of these systems requires interdisciplinary collaboration across computer science, neuroscience, materials science, and electrical engineering, where the integration of computer science and neuroscience perspectives with material and electrical engineering technologies can significantly accelerate practical implementation across all application domains. These research directions collectively address the fundamental challenges in the field while pushing the boundaries of what is possible in neuromorphic applications.’