Two-dimensional material-based devices for in-sensor computing

Abstract

In-sensor computing (ISC) integrates sensing, memory, and processing at the point of data acquisition, enabling real-time, low-power operation. Two-dimensional (2D) materials offer unique advantages for ISC due to their atomic thickness and multifunctional properties. This review highlights 2D material-based ISC devices, covering mechanisms, performance, and architectures, and discusses challenges and solutions toward scalable fabrication and practical deployment in emerging technologies like Internet of Things (IoT), analog computing, and motion detection.

Background and motivation for 2D in-sensor computing

Due to the explosive growth of the IoT^1,2 and artificial intelligence (AI)^3,4, novel computing schemes capable of efficiently handling massive data are crucially needed^5,6. By 2025, a total of 80 zettabytes of data will be collected and processed by global IoT devices⁷. Traditional computing schemes relying on von Neumann architectures, which separate memory, sensing and computing hardware, result in frequent data transfer with undesirable latency and power consumption^8,9. ISC is an emerging solution to the von Neumann bottleneck, incorporating both processing and memory within sensors to facilitate real-time decision-making with low power consumption^5,10,11,12 (Fig. 1).

Fig. 1: Conceptual illustration of in-sensor computing.

a In conventional computing, sensors, memory, and processing are physically separated. Signals from sensors undergo analog-to-digital conversion before being processed. b In in-memory computing, data processing occurs directly within memory arrays, but sensors remain separate. c In in-sensor computing, sensors, memory, and computing are physically integrated into a single platform. The bottom-left panel illustrates multimodal inputs, including visual scenes (e.g., cactus, javelina, tortoise, and roadrunner) and physical stimuli (e.g., light, temperature, strain, and magnetic fields). These inputs are encoded and processed in a distributed resistive network, leading to applications such as edge hardware, space computing, health monitoring, and robot control.

2D material-based ISC devices distinguish themselves from other ISC technologies by offering unparalleled properties such as high sensing sensitivity¹³, high-density memory states¹⁴, and good compatibility with materials in various dimensions^15,16,17. 2D materials enable efficient sensing of light, temperature, magnetic fields and mechanical stress with strong optoelectronic responses¹⁸, rapid temperature-dependent phase transitions¹⁹, intrinsic spin alignment²⁰, and strong piezoelectric effects²¹. 2D material-based memory devices exhibit up to 1280 memory states¹⁴, leading to high-resolution information storage. Mixed-dimensional integrating 2D materials onto complementary metal-oxide-semiconductor (CMOS) circuits¹⁵ or flexible substrates^22,23 lead to the discovery of 2D devices with unprecedent functionality and enhanced wearability. 2D ISC devices offer strong potential in biomedical fields where real-time, on-device processing of multiple input signals is critical²⁴, including brain-machine interfaces^25,26, seizure detection systems^27,28, and health monitoring platforms^29,30. With more 2D material properties being unveiled, the ISC capabilities of 2D material-based devices continue to evolve rapidly³¹.

Many 2D ISC applications such as adaptive behavior^32,33, associative learning^33,34,35, pattern recognition^36,37, and edge detection^38,39 have been demonstrated. These applications are realized through three distinct ways: using single multifunctional device³⁵, deploying arrays of ISC devices^38,40, and heterogeneously integrating sensing and memory devices⁴¹. A multifunctional single device, which integrates sensing and computation within a single 2D platform, offers a compact and cost-effective approach to ISC. Scaling individual devices into array-based ISC configuration enhances parallel processing speed and sensing resolution. It requires complex fabrication techniques to minimize device-to-device variation¹⁹. Alternatively, hybrid integration systems, combining top-tier sensors and memory devices of cross-species in a modular manner⁴², merge their complementary strengths for enhanced functionality. It requires optimized coupling and interconnection strategies to maximize efficiency⁴¹.

This review begins by exploring state-of-the-art 2D material-based ISC devices. In Sections “Ferroelectric devices” and “Memristor and Memristive devices”, we discuss ferroelectric (FE) and memristive devices, analyzing their operational mechanisms, performance metrics, and ISC applications. We then explore the potential of utilizing the unique spatially modulated electronic phases in 2D materials for next-generation ISC devices. Section “Charge density wave devices” delves into charge-density-wave-based devices, while Section “Spintronic devices” examines spin-based devices. These devices hold promise for ISC due to their ultrafast switching dynamics and high sensitivity to optical and magnetic stimuli. Key challenges and future directions in 2D devices-based ISC research are elaborated in Section “Summary and Outlook”, where we discuss the strategies for advancing real-time multimodal ISC applications, the development of emerging 2D topological insulators with unique sensing capabilities, and approaches toward wafer-scale integrations.

Ferroelectric devices

Ferroelectricity originates from non-centrosymmetric crystal structures, where polarization emerges via subtle ionic displacements^43,44. This polarization can be reversibly switched by electric fields, allowing for non-volatile data storage, erasure, and reprogramming⁴⁵. Depending on the origin of ferroelectricity, 2D FE devices are categorized into three types: intrinsic FE devices, interfacial FE devices, and interlayer FE devices. Intrinsic FE devices utilize 2D materials with inherent FE properties, such as In₂Se₃⁴⁶, SnS⁴⁷, SnSe⁴⁸ and CuInP₂S₆ (CIPS)⁴⁹, as the active channels or gate dielectrics (Fig. 2a). Interfacial FE devices employ non-ferroelectric 2D materials (e.g., MoS₂, WSe₂) as channels, paired with traditional 3D FE materials (e.g., barium titanate (BTO)⁵⁰ and poly(vinylidenefluoride-trifluoroethylene) P(VDF-TrFE)³⁸) as gate-dielectric to modulate charge transport (Fig. 2b). Interlayer FE devices exploit the non-centrosymmetric polarization induced by sliding⁵¹ or rotating⁵² between 2D materials (Fig. 2c, d).

Fig. 2: 2D FE devices for in-sensor computing.

a Schematic of intrinsic ferroelectricity in In₂Se₃, where the displacement of the central Se atom induces interlocked in-plane and out-of-plane polarizations. b Schematic of interfacial ferroelectricity, where 3D ferroelectric materials modulate charge carrier concentration in 2D materials, leading to resistance changes. c OOP sliding ferroelectricity in bilayer MoS₂ due to charge center displacement. d Schematic of excitonic ferroelectricity arising from asymmetric moiré structure. e The three-terminal In₂Se₃ FE FETs mimicking synaptic behavior. The gate can be excited with optical stimuli. f Paired-pulse facilitation index of α- In₂Se₃/SnSe based p–n junctions as a function of optical pulse interval time (∆t). The fitted curve shows that PPF decreases exponentially with the increase of ∆t. The inset displays postsynaptic current triggered by two consecutive optical pulses. g Schematic of reconfigurable graphene/MoTe₂/P(VDF-TrFE) homojunctions. The P(VDF-TrFE) layer independently controls ferroelectric domains, resulting in reversible p-n and n-p transitions. h Schematic of 3R MoS2 sliding-ferroelectric FETs, where shear-transformation in 3R MoS₂ epilayers induces polarization switching. i Schematic of bilayer graphene/h-BN based moiré synapse transistors, which exploit electronic ratcheting states generated by moiré potential to enable non-volatile conductance modulation. j Retention characteristics of In₂Se₃ FE FETs, showing high- and low-resistance states (HRS/LRS) under ±10 V write/erase pulses for 30 s, with readout at V_gs = 0 V, V_ds = 1 V. k Schematic of a lane-keeping task, where a laser sensor observes obstacle distance and inputs data into a reservoir computing network. The network output adjusts the vehicle’s steering angle. Benchmark comparison of the 2D FE devices for ISC applications: l Energy consumption vs. retention time, m Sensing optical wavelength vs. operating voltage. Reproduced with permission from: e, j, k ref. ⁵⁶, 2024 American Chemical Society; f ref. ³⁵, John Wiley & Sons; g ref. ³⁸, Springer Nature; h ref. ⁷⁵, Springer Nature; i ref. ⁷⁷, Springer Nature.

Two widely studied intrinsic FE materials for 2D ISC devices are α-In₂Se₃ and CIPS. The α-In₂Se₃ exhibits two distinct ground polar states driven by the motion of the middle Se atom, resulting in interlocked in-plane (IP) and OOP polarizations⁴³. This interlocking effect stabilizes ferroelectricity even at the monolayer limit⁵³. The FE polarization in α-In₂Se₃ can be further controlled by illumination in a non-destructive manner, leading to applications including photon detectors and optoelectrical memory⁵⁴. These properties positions α-In₂Se₃ as a workforce material for 2D FE ISC devices (Fig. 2e)^35,55,56,57. For example, α-In₂Se₃/SnSe based p-n junctions emulate synaptic behaviors, including short-term/long-term plasticity (STP/LTP) and excitatory/inhibitory functions, achieving a ultra-high paired-pulse facilitation (PPF) index of 457% which is crucial for promoting the development of artificial vision³⁵ (Fig. 2f). The p-n junctions implement Pavlovian associative learning: an initial conditioned stimulus (CS, +2 V electrical pulse) induces a subthreshold current (0.12 nA < I_th = 0.14 nA), failing to trigger a response. After five co-stimulation cycles with an unconditioned stimulus (US, 0.69 mW cm⁻² light), the current surges to ≈1.68 nA, establishing a strong associative reflex. Subsequent CS stimuli alone generate suprathreshold currents (0.16 nA), confirming the circuit’s ability to retain and execute the learned CS-US association.

CIPS exhibits spontaneous OOP polarization due to the displacement of Cu atoms from their lattice centers and the cation displacement in the In lattice^49,58. Its insulating nature (bandgap ~ 2.9 eV), high tunneling electroresistance⁵⁹, and dangling-bond-free surface⁶⁰ support robust non-volatile FE memory functionality. While recent studies demonstrate light-induced polarization switching in CIPS^61,62, its role in ISC remains limited to gate modulation rather than direct light-sensing channels. For instance, SnS₂/hexagonal boron nitride (h-BN)/CIPS-based ferroelectric field-effect transistors (Fe-FETs) utilize the optoelectrical properties of SnS₂ and the ferroelectricity of CIPS to emulate optoelectrical synaptic behaviors³⁴. These Fe-FETs form a fully FET-driven reservoir computing (RC) system with a reservoir layer and a fully connected layer. In the reservoir layer, Fe-FETs with optical STP process stimulus-dependent current relaxation. In the fully connected layer, Fe-FETs with electrical LTP and long-term depression serve as tunable synaptic weights for training and classification. This RC system achieves 93.62% accuracy in MNIST image recognition, showcasing a streamlined approach to ISC. CIPS can integrate with telecom-wavelength materials, as shown in CIPS/graphene/h-BN/Te FE-FETs³⁷, where the thickness-tunable bandgap of Te enables efficient 1550 nm (telecom-band) photoresponse. When integrated into an RC system, these Fe-FETs directly process optical fiber signals, achieving ~80% accuracy in digit recognition.

2D Janus MoSSe exhibits spontaneous OOP polarization due to its asymmetric structure, which breaks the OOP structural symmetry of MoS₂⁶³. Its direct bandgap (2.14 eV), high carrier mobilities (157 cm²V⁻¹ s⁻¹ for holes, 74 cm²V⁻¹ s⁻¹ for electrons), and efficient visible-light absorption underpin robust optoelectronic functionality^63,64. For example, ion-liquid-gated MoSSe FETs demonstrate optoelectronic synaptic behaviors, achieving a PPF index of 190%³². This device mimics the human visual system’s light adaptation: under mild illumination (450 nm, 1 s, 0.040 mW/cm²), its current remains sub-threshold (<1.3 nA), while stronger light (450 nm, 1 s, 0.061 mW/cm²) triggers overstimulation (1.44 nA), akin to retinal responses. Applying a −1 V pulse modulates synaptic weight by redistributing Li⁺ ions in the electrolyte, reducing the current to 1.2 nA to restore stable vision. A 10-by-10 array of these devices is used to preprocess optical inputs in three steps. The array converts light stimuli into electrical signals through MoSSe’s visible-light absorption. A thresholding mechanism (1.09 nA) filters out noise by suppressing weak signals, akin to retinal neurons discarding subcritical inputs. The retained signals undergo contrast enhancement, sharpening edges and improving feature resolution. By integrating transduction, noise filtering, and contrast amplification, the system enhances data quality, boosting digit recognition accuracy from 77.6% to 83.3%.

2D interfacial FE devices combine 2D materials as channels with 3D FE materials as gate dielectrics, integrating high optoelectronic sensitivity, robust non-volatile memory, and long-term retention (>90,000 s) for energy-efficient ISC^38,50,65. For instance, graphene/MoTe₂/P(VDF-TrFE) homojunctions employ split gates beneath the FE dielectric to independently control ferroelectric domains on either side of the devices³⁸ (Fig. 2g). By modulating these domains, the potential profile of homojunction can be reversibly tuned from p–n (negative photoresponsivity) to n-p (positive photoresponsivity), enabling multi-level synaptic weight tuning and photoresponsivity reversal. This functionality enables applications like reconfigurable convolutional kernels for edge detection. The devices demonstrate exceptional endurance (>10⁶ cycles) and scalability into functional arrays. A 3-by-3 array of these devices operates as an artificial neural network (ANN), performing energy-efficient pattern recognition (10^-13J per operation) and enabling real-time robotic control, such as directing a robotic dog to execute assigned tasks.

In addition to traditional 3D ferroelectrics such as BTO and P(VDF-TrFE), recently developed ferroelectric thin films like HfO₂ and Hf_xZr_1-xO₂ (HZO) have gained prominence due to their intrinsic scalability, CMOS compatibility^66,67 and robust ferroelectric behavior down to sub-10 nm thicknesses⁶⁸. When integrated with 2D semiconductors such as MoS₂ and WSe₂, these films enable the realization of steep-slope FE-FETs for in-sensor memory and neuromorphic operations^69,70. MoS₂/HZO-based FE-FETs show potential in ISC by demonstrating subthreshold swings below 60 mV/dec and energy-efficient synaptic behavior with sub-picojoule consumption per spike, along with fast switching (∼4.8 ns), high retention (>10 years), and exceptional endurance (>10¹³ cycles)⁷¹.

2D interlayer ferroelectricity originates from charge redistribution via the hybridization of occupied and unoccupied states or net charge transfer across van der Waals (vdW) interfaces, as observed in bilayers of h-BN⁷² and transition metal dichalcogenides (TMDCs)^73,74. 2D interlayer FE devices utilize emergent ferroelectricity phases formed by stacking 2D materials together. These new phases offer exceptional properties such as fast switching speeds, high endurance and low energy consumption at room temperature⁷⁵. For example, h-BN/MoS₂/graphene transistors use shear transformation to induce polarity switching in 3R MoS₂ epilayers, meeting the sub-3 nm node requirement for future CMOS technologies (Fig. 2h). Graphene/bilayer h-BN based Fe-FETs employ ferroelectricity arising from parallel-stacked bilayer h-BN, demonstrating nanosecond switching speed and endurance exceeding 10¹¹ cycles⁷⁶. Bilayer graphene/h-BN moiré synapse transistors⁷⁷ (Fig. 2i) utilize excitonic ferroelectricity^52,78 arising from asymmetric moiré structure, operating at low-power (20 pW) while enabling diverse neuromorphic computing functionalities such as reconfigurable synaptic responses and input-specific adaptation. The moiré synapse transistors exploit electronic ratcheting states generated by moiré potential to enable non-volatile conductance modulation. In addition to the memory functionality, Moiré structures demonstrate intelligent light sensing capabilities. The photodetectors based on 1.2° twisted double bilayer graphene demonstrate bulk photovoltaic effect (3.7 V W^–1) at mid-infrared wavelengths (5 μm and 7.5 μm), due to symmetry breaking and quantum geometry contributions⁷⁹. The bulk photovoltaic effect is electrically tunable and enables detection of light polarization, power and wavelength. By integrating memory, computing and sensing functionalities, 2D interlayer FE devices hold promise for ISC applications.

2D FE devices have demonstrated a broad range of ISC applications, including sociative learning^34,35, light adaptation³², digit recognition^36,57 and edge detection³⁹. These applications are achieved through individual ISC devices³⁵ or array-based configurations implementing computing architectures such as ANN³³, spiking neural networks⁸⁰ and RC systems³⁴. For instance, in RC systems, In₂Se₃-based Fe-FETs, which exhibit a retention time exceeding 48 h (Fig. 2j), are employed to control the motion of robotic vehicles⁵⁶. The vehicles are equipped with lidars to detect the obstacles, which serve as the input to network. The RC system processes input and dynamically adjust steer angle of the vehicles, leading to smooth lane navigation and temporal signal processing (Fig. 2k). The approach operates at 10⁴ times lower power and achieves 25% higher data throughput per second compared to conventional GPU-based systems.

We benchmarked various 2D FE devices for ISC applications, as shown in Fig. 2l, m, evaluating key metrics such as energy consumption, retention time, sensing wavelength, and operating voltage. Energy per switching event was calculated using E = V⋅I⋅t, based on reported voltage, current, and pulse duration, unless the original study provided a measured value directly. Retention, wavelength, and voltage values were extracted from published data; where precise values were unavailable, we estimated typical values from average data or figure interpretations. Variations across studies are expected due to differences in device architecture, material quality, fabrication processes, and measurement setups. These benchmarks show that 2D ferroelectric ISC devices generally exhibit low operating voltages and excellent retention times, although switching energy can be relatively high due to polarization thresholds. The data also reveal clear trade-offs: devices such as Gr/MoTe₂/P(VDF-TrFE) achieve ultralow energy consumption (~10⁻¹³ J) with solid retention but require higher voltages (>10 V), while systems like WSe₂/In₂Se₃ offer broad spectral sensitivity (~600–1900 nm) but show limit retention time (~10² s). Overall, low-power-optimized platforms tend to operate at higher voltage costs, while broadband or high-retention systems may sacrifice endurance.

Memristor and memristive devices

Memristors and memristive devices are components whose resistance depends on the history of applied voltage or current, allowing them to store information as resistance states⁸¹. 2D memristor and memristive devices⁸² enable multimodal and energy-efficient ISC by achieving non-volatile multiple memory states in response to external stimuli such as voltage, light, temperature, or mechanical stress⁸³. Based on the resistance switching (RS) mechanisms, 2D memristor and memristive ISC devices are classified into three types: conductive filament, charge trapping and phase transition⁸⁴.

2D conductive filament devices leverage 2D materials as switching layers to achieve multiple resistance states by dynamically modulating the dimension of filaments⁸⁵. The filaments are formed via metal ion migration (e.g., Ag⁺, Cu²⁺), or oxygen vacancy redistribution. Emerging 2D materials such as MXene-ZnO composites⁸⁶ and oxidized black phosphorus (BP)⁸⁷ enable multimodal ISC applications by integrating memristive switching with multimodal sensing capabilities of light, humidity, and strain. For instance, MXene-ZnO memristors (Fig. 3a) utilize UV light to tune oxygen vacancies and humidity to alter proton coupling (Fig. 3b), facilitating noise-reduced, environment-adaptive neuromorphic data preprocesses that accelerate training processes by 5 times⁸⁶. Similarly, MXene-based piezoresistive memristors detect mechanical stress with high sensitivity (23.9 kPa⁻¹) and broad range (>100 kPa), where pressure adjusts filament dimensions for multilevel switching²³. These devices enable real-time Morse code recognition by encoding “dots” and “dashes” through dynamic and static pressures, respectively. The pressure signals alter the resistance states of the memristors, which store and process the signals. Excitatory postsynaptic currents decode the changes in resistance into alphanumeric outputs, facilitating Morse code interpretation without external conversion circuits.

Fig. 3: 2D memristors and memristive devices for in-sensor computing.

a Schematic of MXene-ZnO-based flexible memristive devices with dual sensing capabilities for light and humidity. The resistance is modulated by formation of oxygen vacancy (V_ö) filament. Humidity influences resistance by introducing protons (H⁺) that interact with oxygen ions (O^2-) and V_ö⁸⁶. b I–V characteristics of MXene-ZnO memristors under varying relative humidity (RH) conditions, showing changes in resistive switching behavior. c Schematic of a multifunctional SnS-based memristor array, responsive to electrical and optical stimuli⁹⁵. d Schematic of a VO₂-based neuromorphic transistor stimulated by 375 nm UV light, where the VO₂ film serves as a channel between source (S) and drain (D) electrodes, with ionic liquid as the gating medium⁹³. e Endurance characteristics of a VO₂-based memristor, showing no signal degradation over 10¹² cycles, using 3.0 V/1 μs write pulses and 0.5 V/1 μs erase pulses. f Schematic of MoS₂ phototransistor arrays for spatiotemporal vision sensing, visualizing motion through a sequence of temporal frames. Pixels from specific columns form a temporal vision sequence s(t), which is processed into temporal compressive states x(t), mimicking bioinspired vision sensors. Benchmark comparison of the 2D memristors and memristive devices for ISC applications: g Energy consumption vs. retention time, h Sensing optical wavelength vs. operating voltage. Reproduced with permission from: b ref. ⁸⁶, John Wiley & Sons; e ref. ¹⁹, Springer Nature; f ref. ⁹⁶, Springer Nature.

2D charge trapping devices achieve multilevel non-volatile resistance states by modulating charge trapping at defect sites, 2D-2D/dielectric interfaces, or gate dielectrics, while leveraging the optoelectronic properties of 2D materials for high-sensitivity, high-endurance ISC applications^88,89,90,91. For example, MoS₂-based memtransistors detect light at ultralow intensities (0.001 mW/cm²) and endure >5 × 10⁸ cycles, enabling secure data encryption via wafer-scale arrays of MoS₂ memtransistor⁹² (Fig. 3c). These systems encode light signals into encrypted data through photosensitive analog programming, resisting eavesdropping and brute-force attacks while operating at ultra-low energy (~100 pJ/operation). An integrated 2D SnS-based memristor circuit demonstrates optoelectronic RC, where spatiotemporal electrical and optical inputs generate high-dimensional reservoir states. The optoelectronic RC maps complex temporal inputs into high-dimensional reservoir states, achieving 91% accuracy in classifying practical sentences with minor natural errors. The 2D h-BN/WSe₂ heterostructure offers enhanced properties for optical synapses⁴¹. The integrated optical sensor detects light in the 405–655 nm range, while the charge-trapping memristor controls synaptic weight. Light reduces the resistance of sensors, increasing carrier density in WSe₂ and enhancing charge trapping in the weight control layer to tune synaptic dynamics. The synapse devices form an ANN architecture which is capable of >90% accuracy in colored and color-mixed pattern recognition.

2D phase transition devices utilize resistance variations between distinct material phases to achieve multilevel non-volatile memory states, offering gigahertz response, multimodal sensing, and ultrahigh endurance for ISC¹⁹. For instance, VO₂-based memtransistors sensor both UV and visible light (Fig. 3d), inducing RS with long retention (>4000 s)⁹³. Their distinct responses to UV and visible light enable RGB noise reduction in digit image preprocessing. Initially, the recognition accuracy was 24% due to noise from random Gaussian interference, which masked key features. After preprocessing with VO₂ memtransistors, which emphasize UV-specific information, the system filtered out the noise and extracted relevant features more effectively, boosting the accuracy to 93%. Additionally, atomically thin VO₂ films—which undergo light- and temperature-driven metal-to-insulator transitions—enable cross-modal spiking sensory neurons¹⁹ with high endurance over 10¹² cycles (Fig. 3e). These devices integrate VO₂ memristors with pressure sensors to encode pressure and temperature signals, enabling robotic hands to dynamically grasp or release objects (e.g., balls, hot water cups) with low latency (<30 ns).

2D memristors and memristive devices enable diverse ISC applications—such as associative learning⁸⁹, pattern recognition⁹⁴, data encryption⁹², language learning⁹⁵, and human-machine interaction¹⁹—by leveraging multimodal sensing and ultralow energy consumption. A notable example is MoS₂ memtransistor-based optoelectronic graded neurons deployed in a two-stream neural network for motion detection and action recognition⁹⁶ (Fig. 3f). The spatial stream processes static frames for image recognition, while the temporal stream analyzes motion information to perceive object direction and visual saliency. This dual architecture achieves 99.2% recognition accuracy with a temporal resolution spanning 10¹–10⁶ milliseconds, enabling real-time tracking of dynamic environments.

In comparison with conventional memristive technologies, 2D-material-based devices exhibit unique functional advantages but also face maturity-related challenges. Metal-oxide memristors, such as those based on TiO₂ and HfO₂, are well-established and offer high endurance (>10⁹ cycles), fast switching (<10 ns), and robust integration with CMOS platforms, making them commercially viable for resistive memory and neuromorphic arrays ^97,98. Yet, these systems typically operate as isolated memory elements and are limited to electrical input⁹⁹. Organic memristors, in contrast, feature mechanical flexibility, solution-processability, and low switching energy (<1 pJ), which are attractive for wearable computing and bioelectronics¹⁰⁰. Yet, their limited environmental stability and short retention time restrict long-term deployment ¹⁰¹. Compared to both, 2D memristors strike a compelling balance: they combine atomic-scale thickness and tunability via external stimuli making them promise for ISC. Nevertheless, their current limitations, including device-to-device variability, scalability, and endurance, must be addressed before they reach technological maturity.

We benchmarked various 2D memristor and memristive devices for ISC applications, as shown in Fig. 3g, h, assessing the same critical performance metrics such as energy consumption, retention time, sensing wavelength, and operating voltage as in Section “Ferroelectric devices”. Variations reflect differences in switching mechanisms, material quality, and device structure. The results show clear trade-offs. Ti₃C₂T_x MXene achieves the lowest energy consumption (~10⁻¹⁵ J), with moderate retention (~10² s), suitable for low-power sensing. MXene-ZnO shows much higher energy (~10^-3 J) but long retention (>10⁴ s), favoring memory stability. MoS₂-based heterostructures offer balanced performance. In Fig. 3h, most devices sense in the 300–900 nm range. TiN_xO_2-x/MoS₂/ITO and SnS provide broad spectral coverage but operate at higher voltages (>1 V), while Black Phosphorus and Ti₃C₂T_x MXene achieve similar response under <1 V. Overall, devices with broader wavelength sensitivity tend to require higher voltages, while low-voltage systems offer narrower spectral windows.

Charge density wave devices

CDWs are collective electronic states characterized by periodic modulations in electron density, often coupled with lattice distortions, forming phases distinct from conventional metals or insulators^102,103. These phenomena are observed in 2D TMDCs such as 1T-TaS₂¹⁰⁴, 1T-TaSe₂¹⁰⁵, 2H-NbSe₂¹⁰⁶, and 1T-VSe₂¹⁰⁷. As temperature increases, CDWs transition between phases—commensurate (C-CDW), nearly commensurate (NC-CDW), and incommensurate (I-CDW)—each with distinct electronic ordering¹⁰⁸ (Fig. 4a). CDW phase transitions induce RS, enabling their use in ultrafast memory devices with picosecond-scale switching speeds and ultralow energy consumption (few femtojoules per operation)¹⁰⁹.

Fig. 4: 2D charge density wave devices for in-sensor computing.

a Phase transitions in 1T-TaS₂ at different temperatures, showing the commensurate CDW (CCDW, <180 K), nearly commensurate CDW (NCCDW, 180K–340 K), incommensurate CDW (ICCDW, 340K–540 K), and metallic (>540 K) phases. b Artificial neuronal devices based on 1T-TaS₂. Optical micrograph of a 3-synapse-1-neuron network, where three memristors are connected in parallel to a 1T-TaS₂ device, mimicking biological neurons. Scale bar: 100 μm. c I−V characteristics of the 1T-TaS₂ film, showing oscillator circuit behavior. Oscillations occur when V_dc is within 2.21 V (green line) to 2.56 V (violet line). d Firing rate vs. V_dc for the 1T-TaS₂-based neuronal oscillator, with a 3 kΩ series resistor, showing nonlinear behavior. e Oscillation waveforms corresponding to different V_dc values in (c). At V_dc = 2.21 V (green) and 2.56 V (violet), oscillations occur irregularly. At V_dc = 2.40 V (cyan), continuous oscillations are observed. f Reflectance spectra of 1T-TaS₂ under 0.4 mW/cm² and 4 mW/cm² incoherent white light illumination at 250 K. Inset: Optical image of a 1T-TaS₂ film (~100 μm). g Temperature-dependent reflectance of 1T-TaS₂ at 600 nm. Maximum optical tunability occurs around 250 K. h Schematic of the experimental setup for photo-manipulation of the polar electronic state in EuTe₄, using an 800 nm pump-probe system and a photomultiplier tube (PMT) for detection. i Temperature dependence of second-harmonic generation (SHG) intensity in pristine EuTe₄. A thermal hysteresis is observed, matching the hysteresis in the resistance curve (inset), indicating phase transition behavior. j Non-volatile switching of SHG intensity and resistance induced by high-intensity laser pulses. Strong excitation above the threshold fluence drives EuTe₄ into a new non-volatile phase, resulting in sharp resistance switching. Reproduced with permission from: a ref. ²¹⁶, 2024 IOP Publishing; b–e ref. ¹¹², 2021 American Chemical Society; f, g ref. ¹¹³, 2025 AIP Publishing LLC; h–j ref. ¹²², Springer Nature.

1T-TaS₂, a prototypical 2D CDW material, is widely studied for neuromorphic computing. 1T-TaS2 exhibits phase transitions from NC-CDW to C-CDW at room temperature through thermal or electric field stimuli^110,111. The phase transition leads to 1T-TaS₂-based stochastic artificial neurons¹¹² (Fig. 4b), where a Pearson-Anson oscillator circuit is used to generate voltage-dependent oscillations ranging from 500 Hz to 5000 Hz. In the regular oscillation regime (Fig. 4c, blue region), increasing DC voltage (V_dc) enhances RC dynamics, elevating firing rates. Conversely, in stochastic regimes (green/purple), firing rates exhibit abrupt sensitivity to V_dc shifts (ΔV < 0.1 V) (Fig. 4d). Oscillation waveforms are tunable: low (2.21 V) and high (2.56 V) V_dc stabilize NC-CDW or IC-CDW phases, while intermediate voltages (2.40 V, cyan) induce stable oscillations (Fig. 4e), showcasing 1T-TaS₂’s potential for reconfigurable stochastic neuron circuits. Additionally, optically tunable CDW domains in 1T-TaS₂ enable light-responsive (Fig. 4f) and temperature-responsive sensors^113,114 (Fig. 4g), bridging CDW physics with adaptive optoelectronic properties for ISC.

Recent studies have expanded 1T-TaS₂’s functionality across multiple ISC-relevant domains. Optical excitation has been shown to induce a metastable heterochiral CDW state with coexisting α and β domains. These form a moiré superstructure with 43.7 Å periodicity and Kagome-like symmetry, resulting in emergent metallicity and flat bands near EF—features promising for ultrafast, light-reconfigurable memory¹¹⁵. In heterostructures, proximity-induced CDWs have been observed in graphene atop 1T-TaS₂, accompanied by a ~31% reduction in the Mott gap and a ~0.3 eV Dirac point shift, suggesting tunable hybrid electronic states¹¹⁶. At the system level, coupled oscillator arrays based on CDW quantum oscillators (CDW-QOs) have demonstrated second-harmonic injection locking and Ising spin encoding. These networks solve Max-Cut optimization problems in under 10 µs at room temperature, with low-voltage operation (~0.01 V) and frequency tunability from 195 to 537 kHz¹¹⁷.

EuTe₄, an emerging quasi-2D CDW material¹¹⁸, exhibits improper polarization enabled by its unique crystal structure¹¹⁹: planar Te-sheets separated by insulating EuTe slabs. CDW formation breaks in-plane inversion symmetry within the Te-sheets, inducing a polar order. The material hosts a ~200 meV CDW gap with incommensurate wave vectors along multiple directions, and a primary modulation (q ≈ 0.643 b*) stable up to 400 K¹²⁰. Scattering and STM measurements reveal competing trimerized domains with opposite polarizations, contributing to a wide thermal hysteresis between 50 K and 400 K and enabling domain-based memory retention¹²¹.

Recent studies demonstrate that light pulses (800 nm) can non-volatilely manipulate the polar state and resistance of EuTe₄, as shown in Fig. 4h¹²². Pump-probe second harmonic generation (SHG) measurements (400 nm) track structural changes, while electrical resistance (measured via the four-electrode method) correlates with temperature-dependent SHG intensity, both displaying thermal hysteresis (Fig. 4i). Under strong excitation (~7.5 mJ/cm²), EuTe₄ transitions to a high-resistance phase, whereas moderate pulses (~4.5 mJ/cm²) fine-tune SHG intensity and restore resistance. Thermal annealing fully reverses the transition, resetting the system to its original state (Fig. 4j). This non-volatile, reversible optical control positions EuTe₄ as a promising candidate for adaptive optoelectronic memory. Additionally, EuTe₄ exhibits large negative magnetoresistance (~86%) at 2 K under magnetic fields above 4 T, attributed to spin canting of Eu²⁺ ions¹²⁰. This behavior suggests that magnetic fields can modulate the CDW state, reinforcing EuTe₄’s potential as a reconfigurable platform for in-sensor memory systems driven by electric, optical, and magnetic inputs¹²³.

2D CDW systems have successfully demonstrated key in-sensor computing functionalities, including light- and temperature-responsive sensing^110,111,122, non-volatile memory through polar state switching¹²², neuromorphic behavior¹¹² and coupled oscillator arrays¹¹⁷. These independent achievements establish a strong foundation for future ISC development. However, several challenges currently limit their practical deployment. Many CDW transitions occur below room temperature, with only a few materials (e.g., 1T-TaS₂, EuTe₄) exhibiting switching near ambient conditions¹²⁴. Domain control is often difficult, as CDW phase switching involves metastable or hidden states that are highly sensitive to local structure and hard to program deterministically^125,126. Ongoing efforts are being made to address these issues. For example, vdW heterostructures combining 1T-TaS₂ with graphene have been investigated to enable vertical transport and interlayer functionality¹²⁷, while h-BN-capped three-terminal CDW devices provide early demonstrations of gate tunability¹²⁸. On the materials growth side, wafer-scale synthesis of monolayer 2H-TaSe₂ and TaS₂ films via APCVD has been reported^129,130, demonstrating a potential route to scalable integration.

Spintronic devices

2D materials exhibit different types of magnetism such as ferromagnetism, antiferromagnetism, ferrimagnetism and altermagnetism. Ferromagnetism originates from parallel alignment of atomic magnetic moments, driven by positive exchange interactions (Fig. 5a). 2D ferromagnetic (FM) materials like CrI₃¹³¹, p-SnSe¹³², VSe₂¹³³, and Fe₃GaTe₂¹³⁴, retain magnetization even without external fields. This inherent magnetization introduces challenges for high-density integration, as stray field coupling between adjacent magnetic domains can cause mutual interference, limiting scalability in ultra-compact spintronic devices¹³⁵. Antiferromagnetism results from antiparallel spin alignment, canceling net magnetization via negative exchange interactions¹³⁶. 2D antiferromagnetic (AFM) materials like FePS₃¹³⁷, CuCrP₂S₆¹³⁸, NiI₂¹³⁹ and MnSe₂¹⁴⁰ exhibit robust spin ordering with zero macroscopic magnetization(Fig. 5b). Ferrimagnetism arises from unequal antiparallel magnetic moments, resulting in a net magnetization. 2D ferrimagnets like Cr₂S₃¹⁴¹ and supramolecular Kondo lattices¹⁴² provide tunable spin properties with reduced stray field effects for spintronic manipulation¹⁴³. Altermagnetism, a recently discovered class, uniquely combines compensated magnetic moments in real space with alternating spin splitting in momentum space^144,145 (Fig. 5c). This dual character enables spin-polarized currents without net magnetization, unlocking an extra spin degree of freedom for ISC paradigms. Emerging 2D altermagnetic (ALM) candidates include Mn₅Si₃ thin film¹⁴⁶, Co_1/4NbSe₂¹⁴⁷ and Fe_1/4NbS₂¹⁴⁸, host unprecedent spin alignment for spintronic applications.

Fig. 5: 2D spintronic devices for in-sensor computing.

a–e Schematic of spin alignments in 2D magnetic systems: a Ferromagnetism: All spins align parallel, producing net magnetization. b Antiferromagnetism: Spins align in antiparallel sublattices, canceling net magnetization. c Altermagnetism: Antiparallel spin sublattices connected by crystallographic rotational symmetry (e.g., 2-fold, 4-fold, 6-fold), which are distinct from inversion, translation, or mirror symmetry. d Skyrmions: Topological spin textures with fixed chirality, where spins swirl continuously from “up” at the periphery to “down” at the core. e Merons: Half-skyrmion excitations where spins transition from in-plane orientations at the periphery to “down” at the core. f Schematic of a spin valve device using Fe₃GeTe₂ as the magnetic layer and h-BN as an insulating barrier. Thin Fe₃GeTe₂ layers (L1: ~7 nm, L2: ~20 nm) are separated by an atomically thin h-BN layer and encapsulated by a thicker h-BN layer. g Schematic of a Cr₂Ge₂Te₆/P(VDF-TrFE) multiferroic heterostructure, sandwiched between an ITO/Au top electrode and a SiO₂/Si substrate. h Probabilistic computing with p-bits, where bias voltage controls the probability distribution between two states. i Transfer curve of a perpendicular magnetic tunnel junction based on an AlO_x barrier, demonstrating magnetic sensing capabilities. j Voltage-dependent magnetic coercivity in a 2L-Cr₂Ge₂Te₆/P(VDF-TrFE) heterostructure at 4 K, showing coercivity modulation and memory functionality under different applied voltages. k Schematic of spin splitter torque devices, where spin polarization from altermagnetism flips the adjacent free ferromagnetic layer. l Schematic of a skyrmionic synaptic device, where bidirectional learning stimuli move skyrmions into (potentiation) or out of (depression) the postsynapse region, mimicking biological synapses. m Skyrmion-based neuromorphic computing concept, showing a Hall bar device and a magnetic skyrmion reservoir for computation. n Synaptic behavior of skyrmion-based devices, showing the number of skyrmions in the postsynapse during potentiation and depression modes under different learning stimulus densities. o Magnetic sensing using skyrmion devices, where transverse resistance (R_xy) evolves under an in-plane magnetic field (B_x) for positive and negative currents (±25 mA). Reproduced with permission from: f ref. ¹⁶⁹, 2018 American Chemical Society; g, j ref. ¹⁷⁴, Springer Nature; h ref. ¹⁷⁰, 2025 AIP Publishing LLC; i ref. ²¹⁷, 2025 IEEE; k ref. ¹⁴⁴, 2025 American Chemical Society; l, n ref. ¹⁶¹, 2017 IOP Publishing Ltd; m ref. ¹⁸³, 2025 American Association; o ref. ¹⁸², arxiv.

Topological spin textures, such as skyrmions¹⁴⁹ and merons¹⁵⁰, transcend conventional magnetic order in 2D materials by organizing spins into stable, topologically protected configurations rather than simple parallel or antiparallel arrangements¹⁵¹. Skyrmions, nanoscale vortex-like spin structures, are stabilized by the interplay of dipolar interactions, Dzyaloshinskii-Moriya interactions (DMI)¹⁵², and magnetic anisotropy¹⁴⁹(Fig. 5d). DMI is induced by broken inversion symmetry in chiral crystals or interfaces, driving the emergence of skyrmions in 2D systems. Examples include ultrathin metallic magnets (e.g., FeGe¹⁵³), Cr-based vdW compounds (CrI₃¹⁵⁴, CrGeTe₃¹⁵⁵), and heterostructures like h-BN/Co¹⁵⁶, WTe₂/CrCl₃¹⁵⁷ where interfacial DMI generates skyrmions that move and interact as particle-like entities with distinct dynamical modes. Merons, resembling half-skyrmions, exhibit partial spin winding and arise in materials such as MnBr₂¹⁵⁸ or strained twisted 2D magnets¹⁵⁹(Fig. 5e). Meron-based devices are promising for high storage density due to their nanoscale size¹⁶⁰. These textures enable ultra-dense, low-energy spintronic devices, as their topological stability permits robust motion with minimal current¹⁶¹.

2D spintronic devices harness the spin degree of freedom of electrons, leveraging magnetization dynamics, robust spin ordering, and topological stability in magnetic materials to manipulate electronic states and achieve tunable resistance¹⁶². The magnetoresistance effect inherent to these devices enables unique capabilities in magnetic field sensing for ISC. Traditional 2D spintronic devices include magnetic tunnel junctions (MTJs)¹⁶³ and multiferroic systems¹⁶⁴. Emerging ALM order and topological spin textures enhance the performance of these systems, offering pathways to novel ISC devices with high-density memory and low-power operation¹⁶⁵.

MTJs are spintronic devices composed of two ferromagnetic layers separated by an ultrathin insulating tunneling barrier. These devices exploit tunneling magnetoresistance (TMR), where the relative magnetization alignment of the ferromagnetic layers governs the tunneling resistance, enabling non-volatile resistance switching. Two-dimensional materials such as h-BN and TMDCs (e.g., MoS₂ and WS₂) serve as atomically thin, smooth tunneling barriers, enhancing TMR ratios and scalability^166,167 (Fig. 5f). For the ferromagnetic layers, 2D vdW magnets like Fe₃GeTe₂, Cr₂Ge₂Te₆, and CrI₃ provide tunable magnetism, room-temperature ferromagnetic order, and robust spin filtering^168,169. Beyond memory, MTJs enable novel computing paradigms such as probabilistic computing¹⁷⁰, where stochastic MTJs paired with 2D-MoS₂ FETs¹⁷¹ realize “p-bits”— fluctuating units that harness intrinsic stochasticity— for tasks like random number generation and spin logic (Fig. 5h). MTJs exhibit magnetic field-sensing and strain-sensing capabilities (Fig. 5i), as their magnetization direction and free energy in the ferromagnetic layer are sensitive to applied stress, allowing tunable TMR for detecting strain amplitudes and direction^172,173. The non-volatile memory and multimodal sensing capabilities make MTJs promise toward ISC applications.

2D multiferroic devices integrate sensing, memory, and processing functionalities into a single platform by exploiting tunable magnetoelectric coupling. These systems often employ heterostructures combining 2D magnetic materials (e.g., Cr₂Ge₂Te₆¹⁷⁴) and ferroelectric layers (e.g., In₂Se₃), enabling reconfigurable spin-dependent optoelectronic responses¹⁷⁵ (Fig. 5g). For memory applications, the ferroelectric layer controls the magnetic state of the heterostructure, allowing nonvolatile data storage through reversible transitions between ferromagnetic (spin-polarized photocurrent, “1”) and AFM (unpolarized photocurrent, “0”) states¹⁷⁶ (Fig. 5j). Simultaneously, the spin-constrained photoelectric effect underpins their sensing capability: variations in light exposure modulate charge transfer and photocurrent characteristics, enabling optical readout of stored information without altering its magnetic state¹⁷⁶. This dual functionality positions 2D multiferroics as promising candidates for ISC applications.

Altermagnetism enables spin current generation without spin–orbit coupling. In altermagnets, a charge current applied perpendicular to the Néel vector generates a spin current parallel to it. This phenomenon, termed spin-splitter torque, provides a basis for new spintronic devices. Altermagnets provide a new mechanism for the generation of spin currents that do not rely on spin−orbit coupling, offering a mechanism termed spin splitter torque devices (Fig. 5k). Integrating ferroelectricity and ferromagnetism into a single material remains a significant challenge. Altermagnets provide a promising avenue for achieving such coexistence, as their unique spin-symmetry properties enable simultaneous ferroelectric and spin polarization. Recent experimental progress supports this potential: studies of (Ge, Mn)Te alloys, which span the phase diagram between FE GeTe and ALM MnTe, demonstrate tunable magnetoelectric coupling¹⁷⁷. In 3D systems, large TMR has been predicted in MTJs with altermagnets theoretically^178,179, and also realized experimentally¹⁸⁰ at room temperature. This is due to the momentum dependent spin polarization even though the total density of states for spin-up and spin-up down electrons at the Fermi energy are the same in these unique materials. This intriguing effect has just emerged as a new frontier in the research of MTJs with 2D materials.

2D magnetic skyrmions, with their nanoscale size, defect tolerance, and ultralow depinning current density, are promising information carriers for ultra-dense, high-speed, and energy-efficient spintronic applications such as racetrack memories¹⁸¹, logic gates, neuromorphic computing¹⁶¹, and magnetic sensing¹⁸² (Fig. 5l). For neuromorphic architectures, skyrmion-based devices like Pt/Co/Ir heterostructures leverage nonlinear magnetic field-driven dynamics in Hall bars to emulate reservoir computing¹⁸³ (Fig. 5m). Skyrmions act as reconfigurable processing units, enabling high-dimensional mapping and short-term memory effects (Fig. 5n). The devices demonstrate a skyrmion density-dependent accuracy (highest 94.7%) in handwritten digit recognition. Skyrmions show magnetic sensing capabilities: a [W/CoFeB/MgO]₁₀ multilayer Hall bar sensor exploits spin-orbit torque-induced transformations between skyrmions, stripe domains, and type-II bubbles¹⁸². By monitoring anomalous Hall effect signals via a differential readout scheme, it detects in-plane (±17 mT) and out-of-plane (±30 mT) fields with linear response and higher sensitivity over conventional anomalous Hall effect sensors (Fig. 5o).

CMOS compatibility remains a challenge for 2D spintronic ISC devices due to three key limitations: low thermal and chemical stability, poor interface quality, and limited synthesis scalability. Many 2D magnets, such as CrI₃ and Cr₂Ge₂Te₆, are air-sensitive and exhibit sub-room-temperature Curie points, requiring encapsulation and cryogenic operation incompatible with CMOS processes^131,184. Approaches like strain engineering, chemical doping, and heterostructure design have shown promise in improving coercivity and thermal stability^185,186. Notably, MBE-grown Fe₃GeTe₂ has demonstrated higher Curie temperatures and improved film uniformity¹⁸⁷. Interface quality also remains a concern, as metal contacts (e.g., Pt, Ta) often introduce contamination and spin scattering^188,189. Encapsulation with h-BN and other passivation methods help preserve interface integrity but remain difficult to scale¹⁸⁶. Meanwhile, wafer-scale synthesis of 2D magnets remains limited. Techniques such as CVD and mechanical exfoliation often yield defects and poor uniformity¹⁹⁰, which hinder reliable integration, especially in multilayer heterostructures where interface control and doping precision are critical¹⁹¹.

Spintronic devices offer promising ISC capabilities, with MTJs enabling memory-in-sensor architectures^59,172, multiferroic devices supporting low-power spin logic¹⁷⁶, and skyrmion-based devices demonstrating magnetic sensing¹⁸², and neuromorphic functions¹⁸³. However, spintronic devices face several key challenges for ISC applications. Many 2D magnets suffer from low Curie temperatures and require cryogenic conditions, limiting practical deployment¹⁶². Interface quality in magnetic heterostructures is critical, as oxidation and defects degrade spin injections and suppress magnetoresistance signals¹⁶². Spintronic readout signals, such as TMR, are often weak and demand amplification or improved sensing architectures¹⁹². Scalability remains an issue: wafer-scale synthesis of uniform 2D magnets is difficult, and conventional spintronic devices face thermal stability and interface control challenges as dimensions shrink²⁰. Recent efforts have begun addressing these issues. For example, vdW magnetic electrode transfer has enabled high-performance 2D spin valves with improved interface quality¹⁹³, while integrated multiferroic tunnel junctions using Mn₂Se₃, TiTe₂, and In₂S₃ have demonstrated in-memory logic and multilevel storage via magnetic and electric-field control¹⁹⁴.

Summary and outlook

Two-dimensional material-based ISC devices achieve their performance through fundamentally distinct physical mechanisms, each shaping how information is sensed, stored, and processed at the material level. FE devices utilize switchable polarization domains arising from broken crystal symmetry. These bistable dipole configurations can be flipped with low energy and minimal leakage, enabling fast, non-volatile operation and long retention³⁴. Memristive devices operate via ion migration, filament formation, or phase transitions. Such localized and stimulus-responsive processes allow high-speed switching and multimodal sensing¹⁹, but their variability and diffusion-driven mechanisms limit endurance and retention⁸⁶. CDW systems exploit collective electron behavior, where the entire electronic structure reorganizes coherently¹¹². This allows for femtojoule, picosecond transitions, ideal for neuromorphic oscillators and memory, though the metastable nature of CDW phases often complicates control and reproducibility. Spintronic devices rely on spin polarization¹⁷⁰, magnetic anisotropy¹⁸², and topological textures such as skyrmions¹⁹⁵. These enable non-volatile magnetoresistive states and vector-sensitive sensing¹⁸², yet their performance is constrained by interface quality, temperature stability, and efficient spin injection. Each mechanism offers specific advantages while also introducing performance trade-offs.

Scaling up high-performance 2D ISC devices is essential for bridging the gap between laboratory innovations and practical real-world applications. Critical challenges span wafer-scale synthesis, defect-minimized transfer, and heterogeneous integration^17,196. To address synthesis, techniques like chemical vapor deposition (CVD) and metal-organic CVD enable growth of uniform 2D transition metal dichalcogenides (TMDCs, e.g., MoS₂, WS₂) on 300-mm silicon wafers, compatible with back-end-of-line (BEOL) semiconductor processes¹⁹⁷. Substrate engineering, such as step-controlled templates, further ensures single-crystal film uniformity. In addition, wafer-scale growth of single-domain 2D monolayer arrays via geometric confinement¹⁹⁸ and stacked multilayer vdW superconductors through a high-to-low temperature growth strategy¹⁹⁹ have been demonstrated. However, transferring these atomically thin layers to target substrates risks introducing defects or contamination, necessitating advanced transfer protocols (e.g., polymer-free methods) to preserve material integrity²⁰⁰. To address these issues, recent efforts have focused on (i) polymer-assisted wet transfer methods, (ii) polymer-free transfer methods, such as adhesive tape-based solvent-free transfer²⁰¹, and (iii) deterministic dry transfer methods²⁰⁰. High-throughput layer-by-layer exfoliation techniques have also been proposed for generating multiple compound semiconductor membranes from a single wafer²⁰². For system-level integration, monolithic 3D architectures, where 2D material layers are sequentially stacked onto prefabricated silicon circuitry, offer a scalable path toward compact ISC systems. This approach enables vertical interconnects with minimized footprint while preserving the performance of bottom-tier CMOS logic, especially when combined with low-temperature transfer techniques and interlayer isolation strategies¹⁶. Monolithic 3D integration of 2D material-based AI-processing hardware has recently been demonstrated, highlighting integrability and multifunctionality¹⁶. However, challenges such as interlayer misalignment, interface contamination, and poor via connectivity need to be addressed to ensure layer-to-layer reliability^16,203.

Current 2D material-based ISC applications, such as digit recognition and edge detection, are constrained to either single-modality optical sensing or lack real-time data processing capabilities. Advancing toward complex applications like autonomous robotics²⁰⁴ or wearable health monitoring²⁰⁵ requires ISC systems to achieve multimodal integration and spatiotemporal analysis. Recent progress highlights a pathway⁹⁶: motion-detection ISC systems based on MoS₂ memtransistors employ two-stream neural networks to decouple static spatial data (e.g., object shape) from dynamic temporal data (e.g., movement), leading to real-time action recognition.

Compared to conventional ISC platforms such as Si⁴², ZnO²⁰⁶, ITO²⁰⁷,and organic semiconductors²⁰⁸, 2D materials offer significant advantages in terms of energy efficiency, spectral coverage, and multimodal sensing versatility (Table 1), owing to their atomic thickness and high surface-to-volume ratio. While conventional systems benefit from mature fabrication processes, excellent endurance (>10⁶ cycles), and reliable device integration, they typically suffer from higher energy consumption (10⁻⁶–10⁻⁸ J) and limited sensing functionality, often restricted to visible light. In contrast, 2D ISC devices achieve ultralow switching energies (down to 10⁻¹³ J), broader spectral sensitivity (280–1310 nm), and support diverse inputs such as light, humidity, and pressure.

Table 1 Comparison of in-sensor computing devices based on conventional and 2D materials.

Emerging 2D topological insulators (TIs) enable ISC with unprecedented functionalities through their unique quantum properties: insulating bulk states and spin-polarized, topologically protected surface/edge states²⁰⁹. These states support phenomena like dynamic magnetoelectric effects, chiral edge transport, and the giant anomalous Hall effect²¹⁰, leading to adaptive sensing, low-power in-memory computing, and stable readout for probabilistic/neuromorphic architectures²¹¹. For example, magnetic TIs achieve high-accuracy pattern recognition at cryogenic temperatures²¹¹. Practical implementation is limited by low Curie temperatures, air sensitivity, and phase instability in candidate materials like MnBi₂Te₄^212,213. To overcome these issues, recent work has explored capping strategies using Al₂O₃ or graphene to stabilize magnetic and topological phases at room temperature²¹⁴. Topologically protected boundary states in 2D insulators enable high-mobility, backscattering-immune charge transport, which can enhance thermoelectric effects such as the Seebeck response. This supports thermoelectric-driven ISC applications, particularly in scenarios where efficient thermal management is critical²¹⁵.

2D ISC systems have the potential to significantly enhance biomedical applications that demand real-time, on-device processing of multimodal signals, such as brain-machine interfaces, seizure detection systems, and health monitoring platforms. Current 2D material-based systems have already demonstrated promising performance in these fields. For instance, in brain-machine interfaces, 2D systems enable real-time neural signal decoding with sub-millisecond latency and ultralow power consumption by processing electroencephalogram (EEG) and local field potentials directly at the sensor layer^25,26. In seizure detection, on-chip 2D devices based on charge-trap memory and dynamic memristor arrays support real-time EEG analysis and closed-loop neuromodulation, achieving sub-second latency, nanowatt-level power consumption, and classification accuracies exceeding 96%^27,28. In wearable health monitoring, 2D systems allow continuous, on-body tracking of signals such as glucose, temperature, and motion, with real-time response within milliseconds, low detection limits (e.g., <10 µM for glucose), and sub-microwatt power consumption^29,30. ISC architectures that co-integrate sensing, memory, and computing at the sensor level could further advance these systems by minimizing data transfer, lowering energy consumption, and shrinking device footprint.

The field of 2D ISC devices is rapidly evolving, driven by breakthroughs in new materials, device architectures, and integration strategies. Advances in ferroelectric, memristive, and spintronic 2D materials are pushing the boundaries of energy-efficient, multifunctional ISC, enabling real-time data processing and adaptive sensing. As research progresses, intelligent materials and sensors will seamlessly integrate with emerging technologies, accelerating the development of next-generation IoT and AI systems.