2D computational photodetectors enabling multidimensional optical information perception

Abstract

Multidimensional optical information, encompassing spatiotemporal intensity, spectral composition, and polarization, enables a wide range of advanced applications, including environmental monitoring, biomedical imaging, and optical communications. Two-dimensional van der Waals (2D vdW) materials have emerged as promising platforms for detecting such diverse optical signals. This Review highlights recent progress in 2D-based computational photodetectors, with a focus on static and dynamic intensity sensing, spectral reconstruction, and Stokes parameter measurement. Representative device concepts, including neuromorphic vision sensors, computational spectrometers, and miniaturized polarimeters, are discussed in terms of their operating principles and sensing capabilities. Moreover, we propose future directions for high-dimensional optical information acquisition and advanced device development, emphasizing coordinated advancements in detection performance, perception functionality, and chip-level integration.

This Review highlights recent progress in 2D-materials-based computational photodetectors, including neuromorphic vision sensors, computational spectrometers, and miniaturized polarimeters, toward multidimensional optical information perception of light intensity, spectrum, and polarization states.

Introduction

Light is a rich carrier of multidimensional information, encompassing intensity, spectral composition, polarization state, and spatial-temporal distribution. Capturing and analyzing this rich optical information facilitates deeper insights into complex scenes, significantly enhancing capabilities such as target recognition and situational awareness^1–6. This expanded approach facilitates advanced techniques, including in-sensor computing for intelligent data processing^7–9, multispectral detection for precise material identification^10–14, and polarimetric imaging to reveal hidden or subtle features^15–18. At the core of these advancements are photodetectors, which have evolved in response to the increasing demands toward miniaturized, compact, and intelligent photodetection systems capable of real-time detection and processing. These next-generation photodetectors combine detection and computational functionalities within a unified architecture, reminiscent of the human retina, which simultaneously captures and processes visual information^19–21. This integrated approach reduces latency, minimizes data transmission bottlenecks, and allows direct interaction with machine-learning algorithms, paving the way toward efficient, adaptive, and context-aware optical detection systems.

2D vdW materials have recently emerged as promising platforms for multidimensional optical perception^1,3,13,16. Characterized by atomically thin layers and dangling-bonds-free surfaces, 2D vdW materials exhibit high carrier mobility, low intrinsic noise, and strong light-matter interactions^22,23. Their tunable band structures and ability to form vdW heterostructures without the constraints of lattice matching, significantly enhance their versatility in detecting and processing intensity, spectral, and polarization information. For instance, the integration of detection and computation within a single 2D vdW materials-based neuromorphic in-sensor device offers a promising solution to the limitations of conventional von Neumann architectures, enabling direct, in-situ processing of complex optical data^8,24,25. Additionally, 2D vdW material-based computational spectrometers have driven advancements in compact, portable spectroscopy^13,26, while the anisotropic crystal structures and engineered symmetry-breaking geometry in these materials provide new avenues for on-chip polarization detection^27–30. Moreover, the rapid progression of machine learning further strengthens these capabilities by enabling multidimensional detection within a single device, fostering intelligent data processing and accelerating the development of next-generation photodetectors^31–33. These innovations represent a paradigm shift toward intelligent photodetectors that harness multiple degrees of freedom in light, significantly enhancing their applicability across diverse fields, including environmental monitoring, biomedical imaging, remote detection, and secure communications.

In this Review, we highlight recent advances on 2D vdW materials for multidimensional optical information perception, emphasizing their remarkable capability to efficiently capture and accurately process intensity, spectral, and polarization information (Fig. 1). Detailed discussions are provided on various underlying detection mechanisms and innovative device architectures that enable these capabilities. Notable examples include neuromorphic vision sensors capable of perceiving and preprocessing both static and dynamic intensity information, computational spectrometers that reconstruct spectral distributions from compressed signals, and miniaturized polarimeters that extract Stokes polarization states. We conclude by outlining key research directions driven by the growing demand for high-dimensional optical information acquisition and advanced device architectures. Through coordinated advancements in detection performance, perception function, and chip integration, 2D computational photodetectors are expected to present a promising route toward next-generation optoelectronic technologies.

Fig. 1. Overview of multidimensional optical information perception using 2D computational photodetectors.

Information perception includes static and dynamic intensity information detection and in-sensor preprocessing based on neuromorphic vision sensors, spectral information reconstruction based on computational spectrometers, and Stokes parameters detection using miniaturized polarimeters. P (λ) is the incident spectrum where light intensity varies with wavelength.

Intensity information detection by 2D neuromorphic vision sensors

Traditional complementary metal-oxide-semiconductor (CMOS) image sensors capture spatial intensity frame-by-frame and rely on external processors, which results in redundant computation, higher power consumption, and increased latency^2,7,34. In contrast, neuromorphic vision sensors based on 2D vdW materials integrate detection and processing in a single device, reducing computational overhead, lowering energy use, and enabling real-time sensing of both static and dynamic intensity information^35–40.

Work principles of 2D neuromorphic vision sensors

For artificial vision systems to approach such efficiency, sensory designs must address the distinct requirements of static and dynamic intensity information processing. As shown in Fig. 2a, a bio-inspired vision device based on 2D neuromorphic sensors mimics the hierarchical processing of biological retinas, converting light pulses into electrical signals to extract both static and dynamic visual information. These sensors operate through three primary photoresponse modes: synaptic, non-volatile bipolar, and spike responses. The synaptic response emulates biological synapse plasticity by gradually accumulating photocurrent and reinforcing the response to repeated stimuli (Fig. 2b). It can be generated by charge trapping and de-trapping processes, including defect charge traps and interface charge traps^41–45. In addition, light-induced valence state change and aqueous ion-based doping of the channel can lead to non-volatile resistance switch and light-tunable synaptic behaviors^46,47. These synaptic responses enable temporal integration that suppresses noise, enhances contrast, and encodes motion trajectories. The nonvolatile bipolar response uses a floating-gate transistor, where light pulses and gate bias induce alternating positive and negative photoconductivity (PPC and NPC) (Fig. 2c), allowing static background subtraction and motion boundary enhancement by alternating PPC and NPC across sequential frames^48,49. The spike response, typical of event-driven sensors, produces positive and negative current spikes corresponding to changes in light intensity. A typical design features a pixel with two complementary PN junctions that generate opposing photocurrents, canceling each other under constant illumination (Fig. 2d). Introducing a capacitor into one branch creates asymmetric response times, producing transient spikes²⁴. Moreover, integrating a light-sensitive diode with a capacitor can induce spike photocurrents arising from the capacitive charge/discharge dynamics^50,51. Consequently, these spike responses enable real-time detection of dynamic visual events.

Fig. 2. Response characteristics and working principles of 2D neuromorphic vision sensors.

a Workflow schematic showing how input light pulses are converted into static and dynamic information via photodetectors with distinct response modes. b Synaptic response and the corresponding defect- and interface-related charge trapping mechanisms. E_C and E_V denote conduction band and valence band, respectively. Red and blue circles represent electrons and holes, respectively. Horizontal dashed lines mark defect energy levels, while vertical arrows illustrate the photoexcitation process. c Non-volatile bipolar response featuring positive photoconductivity (PPC) and negative photoconductivity (NPC), driven by charge tunneling in a floating-gate phototransistor. V_g denotes gate voltage. d Spike response enabled by a complementary PN junction configuration for spike signal generation. I_ph1 and I_ph2 are the photocurrents of the two complementary PN junctions, and I_total is the sum of I_ph1 and I_ph2.

2D neuromorphic vision sensors for static/dynamic intensity perception

Static vision sensors

For static visual information, the interaction between input images and sensor operators can be described by a Hadamard product, where raw data is structured through element-wise multiplication of the input matrix (visual image) and the operator matrix (sensor responsivity)². When processing blurred images (Fig. 3a), neuromorphic vision sensors utilize synaptic plasticity to denoise and enhance contrast. Repeated optical stimulation modulates synaptic weights, suppressing low-intensity noise while amplifying target features^43,52,55. Additionally, reconfigurable bipolar photocurrents enable in-sensor convolution for edge feature extraction^42,53,56,57. In complex scenes with high dynamic range, visual adaptation (e.g. scotopic and photopic adaptation) is essential for capturing both shadowed and highlighted details^43,58,59. Furthermore, by introducing a second light source with opposite photocurrent polarity, the system can achieve simultaneous visual perception and precise spike encoding³⁶.

Fig. 3. Advances in 2D neuromorphic vision sensors for static and dynamic perception.

a Static processing via Hadamard operation between input image and operator matrix, enabling noise filtering, edge enhancement, intensity adaptation, and spike encoding. b Noise suppression of the letter “H” through differential decay rates between signal and noise, modulated by synaptic weights under repeated optical stimuli⁵². c Edge enhancement using a gate-tunable bipolar photoresponse in 2D PdSe₂/MoTe₂ vdW heterostructures⁵³. d Light intensity adaptation in dim and bright conditions via a 2D MoS₂ phototransistor⁴³. e Spike encoding and MIR image classification with a 2D b-AsP/MoTe₂ vdW heterostructure³⁶. f Dynamic information processing by leveraging bipolar-response differencing for motion detection, synaptic plasticity for trajectory detection and event-driven spiking for event detection. g Top: Motion detection via PPC–NPC differencing⁵⁴. Bottom: Motion detection demonstration in 2D BP/Al₂O₃/WSe₂/h-BN vdW heterostructure with gate-tunable non-volatile bipolar photoresponse⁴⁸. h Trajectory detection using graded contour from long-term plasticity in MoS₂ phototransistors⁴⁵. i Event-driven spike generation for event detection via 2D WSe₂ complementary PN photodiodes²⁴. Panels adapted with permission from: (b), ref. ⁵², Nature; (c), ref. ⁵³, Nature; (d), ref. ⁴³, Nature; (e), ref. ³⁶, Nature; (g), top ref. ⁵⁴ and bottom ref. ⁴⁸Nature; (h), ref. ⁴⁵, Nature; (i), ref. ²⁴, Nature.

The synaptic vision sensors mimic the plasticity observed in biological synapses, where output current amplification depends on the interval of successive light pulses and pulse number^46,60,61. Huang et al.⁵² demonstrated a neuro-inspired optical sensor based on 2D NbS₂/MoS₂ phototransistors, which achieved responsivity and detectivity up to 1.1 A/W and 2 × 10¹¹ Jones at 532 nm, respectively. By leveraging synaptic weight modulation via repeated optical stimuli, the synaptic vision sensors can achieve noise filtering and contrast enhancement (Fig. 3b). The accuracy of image recognition in the neural network training is obviously improved after the synaptic pre-processing. To realize both enhancement and suppressing of the synaptic weight, optical and electrical co-stimulation can be applied^62,63, for example, excitation by optical pulses and inhibition by electrical pulses⁵². Furthermore, electrical control also allows for the modulation of optical response in both magnitude and speed^43,45,48, thereby offering more versatile and intelligent processing functions by optical-electrical co-modulation in the synaptic vision sensors.

In addition to the independent synaptic weight modulation of the sensor pixels, the complex visual conditions require sensors to do convolution for extracting edge features. The key feature for convolution sensors is bipolar photoresponse, which operates as the bipolar weight in convolution kernel. Wang et al. designed a 2D vdW heterostructure (WSe₂/h-BN/Al₂O₃) with gate-tunable positive and negative photocurrents⁵⁷. Using different kernel configuration, they demonstrated image processing functions of image stylization, edge enhancement and contrast correction. A convolutional neural network is also implemented for in-senor classification of 3 × 3 binary figures⁵⁷. Pi et al.⁵³ developed a 2D PdSe₂/MoTe₂ vdW heterostructure to achieve gate tunable band alignment between type-II and type-III heterojunctions, which generate both positive and negative photocurrents in a broadband range. The bipolar photocurrent can be linearly tuned by the gate voltage, with responsivities ranging from tens to hundreds of mA/W and response speed up to 0.4 μs at the wavelengths of 980, 532 and 365 nm. By configuring different gate matrices as convolutional kernels (Fig. 3c), they demonstrated in-sensor image processing of sharpening and edge enhancement across multi-spectral bands from ultraviolet, visible to near-infrared⁵³. In parallel, Yang et al.⁶⁴ demonstrated precise extraction of dim target edge features across contrast-varying images using graphene-Ge vdW heterostructure. The bipolar photoresponse of the device depends on the bias polarity, and a special designed circuit implemented with the sensor array dynamically tunes the photoresponse of the central active device in the kernel, enabling accurate and robust tracking of dim targets compared to conventional optoelectronic convolutional processing⁶⁴.

Except dim conditions, ultrabright visual environment is also challenging for vision sensors as the natural light intensity spans a large dynamic range of 280 dB while the advanced silicon-based CMOS sensors only cover 70 dB. Liao et al. designed MoS₂ phototransistors with intentionally introduced trap states for dynamic modulation of the device photosensitivity under different lighting conditions⁴³. Emulating the functions of the horizontal cells and photoreceptors in the retina, the fabricated sensor array exhibit both scotopic and photopic adaptation (Fig. 3d), offering a broad perception range of up to 199 dB⁴³. Wang et al.³⁶ demonstrated visual adaption for mid-infrared (MIR) signal, which is more challenging than visible and near-infrared (NIR) lights. They employed a 2D b-AsP/MoTe₂ vdW heterostructure with high MIR detectivity of 9.6 × 10⁸cm Hz^0.5/W at 4.6 μm and fast NIR photoresponse rate (∼600 ns). With the assisted NIR stimuli, the device exhibits an adaptable MIR photoresponse to varying light intensities, analog to a human-eye’s visual adaptation behavior. The NIR sampling can also encode the MIR illumination into rate-based spike trains, enabling spiking neural network to achieve digit classification with an accuracy higher than 96% (Fig. 3e)³⁶.

Dynamic vision sensors

For dynamic visual information, unlike conventional image sensors that output continuous intensity frames, neuromorphic vision sensors can retain and process multi-frame input by encoding temporal memory (Fig. 3f). One strategy involves modulating long-term synaptic plasticity to encode motion trajectories as variations in output intensity across frames^45,48,65. Another approach leverages non-volatile bipolar photoconductivity^48,54,66, where a positive “background” frame is stored and subtracted from a negative “current” frame, highlighting motion edges while filtering out static background. Additionally, event-driven sensors respond only to changes in light intensity²⁴, enabling sparse, energy-efficient encoding with minimal data redundancy and ultrafast transmission of dynamic signals to downstream processors.

As shown in Fig. 3g, the bipolar frame differencing is operated on vision sensors with non-volatile and bipolar photoconductivity⁵⁴. For example, the sensor response for frame n is positive and the conductance of the device is kept constant until input light changes in next frame, corresponding to the PPC in Fig. 2d. Then the response for frame n + 1 is NPC as programmed by gate voltage, and the output of frame n + 1 equals the difference between the PPC and NPC of the two frames. Hence, only the moving parts are reflected in the differencing result. In these motion sensors, the programming speed between different states induced by light pulses defines the temporal resolution. Zhang et al.⁴⁸ developed a floating-gate BP/Al₂O₃/WSe₂/h-BN vdW heterostructure and Pang et al.⁵⁴ developed a rippled MoS₂ phototransistor to achieve such non-volatile PPC and NPC with fast temporal resolution up to 100 μs. As an example, in a simulated scene where two people are walking and running while the background is static (Fig. 3g), the output result clearly shows the moving profile of the two people, thereby achieving efficient motion detection and reducing large redundant data⁴⁸.

If the moving trajectory is required in motion detection, the synaptic sensor array with long-term plasticity response can be an alternative approach, emulating the functions of insects’ compound eyes which can detect fast moving objects^45,67–69. The synaptic response shows a gradual decay after light stimuli (as illustrated in Fig. 2b), and the different output currents of the array pixels reflect time-varied motion information. If the contour of an object shows graded increased intensities from left to right, it indicates the object is moving from left to right (Fig. 3h), and vice versa. Furthermore, the gate modulation of the plasticity in the MoS₂ phototransistor enables motion perception with different temporal resolutions from 10 to 10⁶ms⁴⁵, which is critical for motion perception with different speeds^24,45. A high temporal resolution enables clearly capturing of fast-moving objects while may cause oversampling and data redundancy in detecting slow movements. Implementing motion perception with well-tuned temporal resolution ensures both high efficiency and reliable accuracy. The temporal resolution can be dynamically tuned by modulating the response decay time via gate control, and may further be programmed through the integration of back-end circuitry and software control.

Beyond the above motion detection pathways, event-driven motion sensors that respond solely to changes in light intensity offer enhanced detection efficiency and reduced data redundancy. Instead of capturing frames at fixed exposure intervals, the event-driven motion sensors generate sparse spike signals only when input light intensity changes. Zhou et al.²⁴ designed a device with two complementary PN junctions based on 2D WSe₂, the fast and slow response of each PN junction generate a total current in spike form. The photoresponsivity of the sensor is tuned by the floating gate, enabling self-powered and fast event detection with a temporal resolution of 5 μs. Embedding a spike neural network in a device array, the motion of left-hand and right-hand waving can be directly recognized by the spike strains (Fig. 3i)²⁴.

Spectral information detection based on 2D computational spectrometers

Spectral information is essential for material identification, environmental monitoring, and optical communication, enabling precise analysis via spectroscopy. Leveraging the tunable spectral response of 2D vdW materials, single-detector computational spectrometers offer a compact alternative to bulky optical components, enabling efficient and miniaturized spectral reconstruction.

Mechanisms for spectral response modulation in 2D vdW materials

Bandgap engineering

One effective method for tuning the spectral response of 2D vdW materials is bandgap modulation (Fig. 4a), achievable through mechanisms like the Stark effect and electrostriction. The Stark effect shifts and splits electronic states under an external electric field^70–73, enabling dynamic, precise control of the optical response. In contrast, electrostriction involves field-induced lattice deformation^74,75, introducing strain-dependent bandgap variations. While the Stark effect alters electronic states directly, electrostriction modulates them via structural changes. Together, these effects offer versatile and fine-tuned control over spectral properties, enhancing the performance and adaptability of 2D-material-based computational spectrometers.

Fig. 4. Spectral response modulation mechanisms and reconstruction methods.

a Two electric-field-driven mechanisms in 2D vdW materials: (i) bandgap engineering and (ii) band alignment tuning. E_g and E_g’ denote the intrinsic and modulated bandgap, respectively. b Spectral reconstruction in linear spectrometers: device is first trained with known spectra to build a spectral response matrix, which is then used to reconstruct unknown spectra via linear regression. c Nonlinear reconstruction using deep neural networks (DNNs): the model learns the mapping between known spectra and photocurrent outputs, enabling prediction of unknown spectra from measured photocurrent. λ is wavelength, P (λ) is the incident spectrum, I_i is the photocurrent under modulation V_i, and R (V_i, λ) is the corresponding responsivity, γ(V, λ) is the nonlinearity factor.

Band alignment modulation

Beyond intrinsic bandgap tuning, spectral response in 2D vdW materials can be modulated by engineering heterostructures composed of stacked 2D layers. Interfacial interactions create unique band alignments that can be externally tuned to control carrier transport and optical absorption (Fig. 4a). Applying an electric field modulates the interfacial potential barrier^{13,26,71,74,76–80}, enabling dynamic adjustment of spectral response and interlayer excitonic transitions. These tunable excitonic features broaden the absorption spectrum and enhance spectral range. Such flexibility is essential for computational spectrometers, enabling precise spectral encoding, high adaptability, and improved resolution across a wide bandwidth.

2D vdW materials-based computational spectrometers

Linear computational spectrometers

Linear spectrometers operate on the principle that the photocurrent is linearly proportional to the incident light intensity, expressed as $\overrightarrow{{\mathit{I}}_{V,\lambda }}=\overrightarrow{{\mathit{P}}_{\lambda }}\cdot \overrightarrow{{\mathit{R}}_{V,\lambda }}$, where $\overrightarrow{{\mathit{I}}_{V,\lambda }}$ is the photocurrent vector measured at different electric fields (V); $\overrightarrow{{\mathit{P}}_{\lambda }}$ is the light intensity vector of the incident spectrum; $\overrightarrow{{\mathit{R}}_{V,\lambda }}$ is the spectral response matrix which depends primarily on the incident wavelength (λ) and applied modulation field (V). These spectrometers enable spectral reconstruction through linear regression and inverse problem-solving techniques¹². As illustrated in Fig. 4b, the process typically involves three steps: learning, sampling, and reconstruction. In the learning phase, the responsivity function is calibrated using reference spectra by measuring the photocurrent:

$$I_i={\int }_{{\lambda }_1}^{{\lambda }_m}P\left(\lambda \right)\cdot R\left(V_i,\lambda \right)\mathrm{d}\lambda$$

where P (λ) is the incident spectrum, I_i is the photocurrent under modulation V_i, and R (V_i, λ) is the responsivity. Varying V_i across n steps generates a responsivity matrix mapping wavelength to field strength. In the sampling step, the device records photocurrents for an unknown spectrum under different modulation conditions, forming a current vector $\overrightarrow{{\mathit{I}}_V}$. Finally, spectral reconstruction is achieved using least squares or regularization methods, allowing the unknown spectrum to be recovered from the learned responsivity matrix.

Nonlinear computational spectrometers

Most 2D-based photodetectors inherently exhibit nonlinear response due to complex carrier recombination, interfacial charge trapping, and multiphoton absorption. Consequently, nonlinear spectrometers show photocurrent dependence on electric field strength, incident wavelength, and light intensity, expressed as: $\overrightarrow{I_{V,\lambda }}=\overrightarrow{P_{\lambda }^{\gamma \left(V,\lambda \right)}}\cdot \overrightarrow{R_{V,\lambda }}$, where $\gamma \left(V,\lambda \right)$ represents the nonlinearity factor. Given the complexity of nonlinear responses, traditional analytical reconstruction methods are often inadequate. Instead, data-driven approaches especially deep learning, offer effective solutions. By training deep neural networks (DNNs) on large datasets of known spectra and corresponding photocurrent signals, the models learn to map nonlinear responses to spectral outputs (Fig. 4c). Once trained, DNNs decode voltage-dependent photocurrents and reconstruct unknown spectra via inverse mapping, significantly enhancing reconstruction accuracy in strongly nonlinear systems^80,81.

Advances in 2D computational spectrometers

Recent advancements in 2D-based spectrometers have attracted significant attention due to their ability to reconstruct spectral information across the visible (VIS), near-infrared (NIR), and mid-infrared (MIR) regions. These developments^{13,26,71,74,76–81}, illustrated in Fig. 5a, are critical for acquiring multidimensional spectral information. A pioneering work by Xia et al.²⁶ introduced a linear computational spectrometer utilizing a single black phosphorus (BP) photodetector with a footprint of 144 µm², capable of reconstructing NIR-MIR spectra within 2-9 µm. By harnessing the strong Stark effect and gate-tunable light-matter interactions in BP, this single-detector system demonstrated significant potential in reconstructing both monochromatic and broadband spectra, achieving an initial spectral resolution of 420 nm with 41 photocurrent sampling points. The resolution was further improved to 90 nm by narrowing the spectral range to 4–7 µm and increasing the sampling density to 81 points. However, due to inherent dark noise in BP at room temperature, the device was operated under cryogenic conditions to mitigate noise contributions²⁶.

Fig. 5. Overview of computational spectrometers based on 2D vdW materials.

a Typical advances in 2D linear and nonlinear computational spectrometers in VIS-NIR-MIR spectral range. Linear computational spectrometers include dual-gate tuning BP²⁶, gate-tuning BP/MoS₂⁷¹, gate-tuning ReS₂/Au/WSe₂⁷⁶, bias-tuning BP/MoS₂⁷⁷, gate-tuning MoS₂/WSe₂¹³, and gate-tuning ReSe₂/SnS₂⁷⁸. Nonlinear computational spectrometers include bias-tuning InSe/GeSe⁸⁰ and gate-tuning MoS₂ homojunction⁷⁴. b Plot comparing the spectral resolution, operational spectral range, footprint and spectral waveband number (λ_range/Δλ) for selected 2D computational spectrometers. Gra/InSe was demonstrated in ref. ⁷⁹. Panel (a) adapted with permission from: ref. ²⁶, Nature; ref. ⁷¹, Copyright © 2023 American Chemical Society; ref. ⁷⁶, Nature; ref. ⁷⁷, Nature; ref. ¹³, AAAS; ref. ⁷⁸, Nature; ref. ⁸⁰, AAAS; ref. ⁷⁴, Nature.

2D vdW heterostructures, characterized by various band alignments, have emerged as an effective platform for computational spectrometers. These structures offer built-in potentials that suppress dark currents while enabling tunable spectral detection at room temperature. Chen et al.⁷¹ developed a room-temperature self-powered infrared spectrometer based on a 2D BP/MoS₂ vdW heterostructure with an active area of 1500 µm². By applying a gate voltage, the band structure of BP and the band alignment of the BP/MoS₂ vdW heterostructure were modulated effectively, generating a responsivity matrix dependent on wavelength and voltage. This allowed for spectral reconstruction within the spectral range of 1.7–3.6 µm, achieving a spectral resolution of approximately 43 nm⁷¹. Further advancements in vdW-based spectrometers have significantly improved spectral resolution across different wavelength ranges. Sun et al.¹³ fabricated a compact miniaturized computational spectrometer using a MoS₂/WSe₂ vdW heterostructure with a 176 µm² footprint. By employing mathematical linear regression techniques and inverse problem-solving methods, they achieved a spectral resolution of 3 nm across the VIS-NIR range (405–845 nm). Sun et al.⁷⁷ further developed a high-performance broadband spectrometer based on a 2D BP/MoS₂ vdW tunnel diode, capable of operating between 500 and 1600 nm with a spectral resolution of 2 nm.

Interlayer optical transitions beyond the intrinsic optical bandgaps of 2D vdW materials exhibit sensitivity to external electric field modulation, making them suitable for tunable spectral detection and reconstruction. Zhang et al.⁷⁶ exploited this principle by constructing a spectrometer using a 2D ReS₂/Au/WSe₂ vdW heterostructure, which achieved a resolution of 20 nm in the NIR region of 1150–1470 nm with a device footprint of approximately 24 µm². Similarly, Wu et al. demonstrated a single-detector spectrometer employing a 2D SnS₂/ReSe₂ vdW heterostructure, attaining a spectral resolution of 5 nm in the VIS-NIR region of 400-800 nm⁷⁸. These findings highlight the growing feasibility of 2D vdW heterostructures in computational spectroscopy, offering a compact yet powerful alternative to traditional spectrometer configurations. These computational spectrometers obtain the response matrix through electric field modulation of band matching and utilize a linear regression algorithm model to reconstruct unknown spectra.

For 2D-based single-detector spectrometers exhibiting inherently nonlinear responses, spectral reconstruction increasingly relies on artificial neural networks^33,74,80,81. Naveh et al.⁸⁰ demonstrated the use of a fully connected DNN comprising four hidden layers to reconstruct the nonlinear spectral response of a 2D GeSe/InSe vdW heterostructure with a 625 µm² footprint. The dataset consisted of pairs of known illumination spectra generated by seven light-emitting diode (LED) sources, each modulated at ten distinct intensities, combined with corresponding photocurrent vectors measured across 101 bias voltages. The data were randomly partitioned into a training set (80%) and a test set (20%). This approach yielded a remarkable mean reconstruction error of 2 × 10⁻⁴ and a high spectral resolution of 0.35 nm within the 400-1100 nm range⁸⁰. Similarly, Li et al. developed a nonlinear computational spectrometer based on a positive-intrinsic-negative (PIN) WSe₂ homojunction⁸¹. By employing a nonlinear neural network to learn from a four-dimensional dataset comprising photocurrent, illumination intensity, bias voltage, and wavelength, they achieved inverse reconstruction of the input spectra from measured photocurrents. The spectrometer demonstrated a peak wavelength accuracy of 0.18 nm and a spectral resolution of 2 nm within the narrow 630-640 nm band⁸¹. Expanding on these strategies, Xiong et al.⁷⁴ introduced a dual-signal scheme utilizing both photocurrent amplitude and relaxation time to encode spectral information. Their spectrometer, based on a semi-floating 2D MoS₂ homojunction with a 500 µm² footprint, leveraged the electrostriction effect to dynamically tune both the bandgap and carrier kinetics. A DNN with five hidden layers was trained to map the dual-signal responses under varied gate voltages to the corresponding input spectra, achieving a spectral resolution of 1.2 nm in the 450–800 nm range⁷⁴. Beyond spectral reconstruction, deep learning-enabled computational detectors have demonstrated substantial potential for mix-dimensional optical information retrieval. For instance, DNNs have been successfully applied to simultaneously reconstruct wavelength and polarization states from the photocurrent outputs of 2D graphene photodetectors by learning the complex mapping between mixed-dimensional optical parameters and device responses^3,31.

As shown in Fig. 5b, a comparison of 2D-based linear and nonlinear computational spectrometers highlights key performance metrics, including operational spectral range (λ_range), resolution (Δλ), footprint, and spectral waveband number (λ_range/Δλ, defined as the ratio λ_range/Δλ)⁷⁴. The spectral waveband number refers to the total number of discrete wavelength intervals into which a spectrometer divides the incident light for analysis. A higher spectral waveband number enables finer spectral detail, supporting more accurate identification of subtle spectral features. In general, a trade-off exists between spectral waveband number and device footprint, as larger devices typically achieve higher waveband numbers by enhancing light absorption and responsivity, which improve spectral resolution. Compared with conventional benchtop spectrometers, 2D computational spectrometers show strong potential to overcome this trade-off by achieving comparable spectral waveband numbers within a significantly smaller footprint⁷⁴. Linear spectrometers using regression models typically achieve resolutions from 2 to 90 nm. This variation stems from rank deficiency in the responsivity matrix, where field-modulated spectral responses are often correlated. Responsivity tuning is also constrained by material bandgaps, defect states, and fabrication quality, which may result in information loss and limited reconstruction accuracy. In contrast, nonlinear spectrometers powered by deep neural networks can achieve sub-nanometer resolution. These data-driven methods overcome linear model limitations but require large, high-quality training datasets to ensure optimal performance.

Polarization information detection based on 2D miniaturized polarimeters

Polarization information is defined by the Stokes parameters (S₀, S₁, S₂, S₃), which fully describe the polarization state and are typically visualized on the Poincaré sphere, where S₁-S₃ represent the Cartesian coordinates (Fig. 6). Due to their anisotropic optical properties and tunable symmetry, 2D vdW materials show great promise for compact and intelligent polarization sensing^{17,29,37,82–85}. Recent advances include twisted 2D vdW materials and 2D vdW material–metasurface heterostructures capable of capturing complete polarization states. In parallel, reconstruction methods have progressed from empirical models to data-driven machine learning approaches (Fig. 6).

Fig. 6. Overview of 2D miniaturized polarimeters.

Two device configurations including twisted 2D vdW materials and 2D vdW material/metasurface heterostructures were demonstrated for full-stokes polarization detection. For the polarization information reconstruction, empirical model and machine learning techniques were mainly used. S₀, S₁, S₂ and S₃ are the Stokes parameters. F₁-F_n represent the modulate function of the polarization state (S₀, S₁, S₂, S₃) and I₁-I_n are the corresponding photocurrents.

Miniaturized polarimeters based on twisted 2D vdW materials

The twisted stacking of 2D vdW materials not only breaks the inversion symmetry inherent in untwisted layers but also generates moiré patterns that drastically alter interlayer coupling, resulting in significant modifications to the electronic band structure^86–89. This distinctive structural configuration gives rise to emergent polarization sensitivity, offering a novel platform for engineering compact, high-performance polarimetric detectors with precisely tailored response characteristics for specific polarization states.

Twisted 2D vdW heterostructures engineered through interfacial symmetry design present an effective platform for on-chip polarization detection. For instance, twist-stacking WSe₂ (threefold rotational symmetry) with BP (twofold rotational symmetry) while keeping their mirror planes parallel produces a heterointerface that lacks rotational symmetry and preserves only one mirror plane. This symmetry reduction generates an in-plane electronic polarization, enabling photodetection of linearly polarized light⁸⁹. Introducing chirality into twisted 2D MoTe₂/MoS₂ vdW heterostructures further allows selective responses to left- and right-circularly polarized light, facilitating efficient separation of mixed circular-polarization images⁹⁰. These methods offer valuable guidance for miniaturized polarimeter design, yet they typically extract limited polarization information.

Leveraging the configurational versatility of twisted heterostructures, current efforts are shifting toward devices capable of simultaneously resolving polarization, intensity, and even wavelength-encoded signals. A representative example is the fiber-integrated polarimeter reported by Xu et al.⁹¹ In which, a Bi₂Se₃-BP-BP vdW heterostructure with controlled twist angles is assembled on the tip of an optical fiber. Two BP layers rotated by 60° discriminate linear and circular polarization components, while the isotropic Bi₂Se₃ layer acts as a power-calibration reference independent of polarization. Under zero bias, the BP channel delivers a −3 dB bandwidth (f_-3dB) of ~1 MHz and a specific detectivity of 2.91 × 10⁸ Jones at 1500 nm. A dedicated reconstruction algorithm converts the photocurrent outputs into full Stokes parameters (S₀, S₁, S₂), providing complete linear-polarization information alongside intensity⁹¹. Pushing toward multidimensional detection, Wang et al.¹ developed photodetectors based on double-twisted black arsenic-phosphorus (b-AsP) homojunctions that exploit the photothermoelectric effect at zero bias (Fig. 7a). Operating in the 3.7–5.7 µm mid-infrared range under ambient conditions, these devices achieved a specific detectivity of 1.2 × 10⁹ Jones, a responsivity of 50 mA/W, and a f_-3dB of 17.5 kHz. These homojunctions exhibit phase-offset and polarization-dependent bipolar responses, enabling simultaneous detection of full linear polarization states (S₁, S₂) and power densities (S₀) through their output photocurrents. Additionally, by analyzing variations in responsivity, the device can concurrently measure the incident wavelength (3.7–5.7 µm) and full linear polarization states. A clear mathematical relationship links the incident light’s intensity, polarization state, and wavelength with the photocurrents, enabling their reconstruction through reverse computation. The reconstructed parameters (S₀, S₁, S₂) demonstrated a root mean square error (RMSE) of about 20% compared to the reference values, offering a promising technical route for the simultaneous acquisition of multidimensional optical information¹.

Fig. 7. Advances in 2D miniaturized polarimeters.

a Twisted b-AsP homojunctions demonstrated simultaneous detection of full linear polarization states and intensity information¹. b Full-stokes polarization detection based on a twisted double bilayer graphene empowered by a convolutional neural network (CNN)³¹. c 2D PdSe₂/metasurface heterostructures for full-stokes polarization detection⁹². d An optoelectronic polarization eigenvector comprising four MoS₂/metasurface subpixels for high-accuracy full-stokes polarization detection by using a machine learning algorithm³². e Plot comparing of Stokes parameters (S₀, S₁, S₂, S₃) reconstruction accuracy (RMSE values) and detection performance (D*: detectivity, R: responsivity, f_−3dB: -3dB bandwidth) for selected 2D miniaturized polarimeters^{1,31,32,91–96}. Panels adapted with permission from: (a), ref. ¹, Nature; (b), ref. ³¹, Nature; (c), ref. ⁹², Nature; (d), ref. ³², Nature.

Despite these advancements, the methods described above do not cover the entire Poincaré sphere in terms of Stokes parameters. In the work by Xia et al.³¹, a gate-tunable twisted double bilayer graphene (TDBG) moiré superlattice was employed to capture full-Stokes parameters and wavelength information (Fig. 7b). The TDBG moiré superlattice demonstrates a strong bulk photovoltaic effect (BPVE), achieving a specific detectivity of 4.2 × 10⁵ Jones, a responsivity of 0.694 mA/W, and a f_−3dB of 10 kHz at 7.7 µm and a temperature of 79 K. A key point in this work is the ability to tune the Berry curvature within the TDBG by applying external displacement fields through top and bottom gates. This tuning enables dynamic control over polarization-dependent photoresponse, encoding the incident light’s full-Stokes parameters and wavelength into complex photovoltage mappings that vary with the applied top and bottom gate voltages. The photovoltage mappings were used to construct the training and validation datasets, from which a deep-learning-based convolutional neural network (CNN) was trained to extract polarization and wavelength information. As shown in Fig. 7b, the normalized Stokes parameters produced by the trained CNN closely matched the measured values for various input polarization states. This method, which combines machine learning with advanced materials, achieved high parameter recognition accuracy, with a mean squared error (MSE) of 10% for S₀ and 0.2% for S₁, S₂, and S₃, corresponding to root mean square errors (RMSE) of 31.6% for S₀ and 4.47% for S₁, S₂, S₃, respectively. This framework opens new avenues for high-precision, intelligent detection of multidimensional information such as polarization, intensity, and wavelength³¹.

Miniaturized polarimeters based on 2D vdW material/metasurface heterostructures

Metasurfaces offer precise control of light at subwavelength scales, which has led to promising advances in miniaturized and compact polarimetry^93,97–102. One notable approach is the direct integration of plasmonic metasurfaces with 2D vdW material photodetectors, enabling the measurement of the full-Stokes parameters in a compact format. For instance, Wang et al.⁹⁴ developed a full-Stokes polarimeter composed of four chiral plasmonic metasurface-integrated graphene-silicon photodetectors. In this device, the full-Stokes parameters of 1550 nm infrared light are reconstructed through a defined mathematical relationship between the four output photocurrents and the corresponding Stokes parameters. Based on the reported data, the root mean square errors (RMSE) for S₁, S₂, and S₃ were estimated to be 13%, 17%, and 17%, respectively⁹⁴. However, integrating the four detectors into a single pixel structure would reduce spatial resolution for polarization imaging applications, highlighting the need for future devices to achieve full-Stokes detection with fewer output channels without sacrificing imaging quality.

Building on these insights, Dai et al.⁹² demonstrated an alternative design that further reduces the number of output terminals (Fig. 7c). Their polarimeter integrates a 2D PdSe₂-metasurface with only three output channels for full-Stokes detection at 5.3 µm. Operating at room temperature and zero bias, it attains a specific detectivity of 2.5 × 10⁵ Jones and a f_−3dB of 1.1 kHz. By carefully engineering the spatial distribution of the chiral metasurface and taking advantage of the photothermoelectric effect inherent to PdSe₂, the device generates bipolar responses to both linearly and circularly polarized light. This bipolar behavior provides an additional degree of freedom compared to traditional unipolar responses. It allows the establishment of a clear mathematical relationship between the three photocurrents and the azimuthal and ellipticity angles of the polarization state on the Poincaré sphere. As a result, the device achieved unambiguous reconstruction of the full-Stokes parameters, with estimated RMSE values of 18.5% for S₁, 20.3% for S₂, and 6.5% for S₃⁹². This example illustrates how design innovations can reduce complexity while still accurately capturing polarization information.

While metasurfaces themselves provide remarkable flexibility in controlling spatial light distributions and achieving high polarization ratios, their optical responses are typically fixed after fabrication. In contrast, 2D vdW materials offer tunability through external controls such as electrostatic fields, adding an extra layer of versatility to device design¹⁰³. By combining the fixed, engineered responses of metasurfaces with the adjustable properties of 2D vdW materials, it becomes possible to create detectors with dynamic polarization responses. For example, Cheng et al. designed a graphene-metasurface heterostructure featuring a split dual-gate configuration⁹⁵. By varying the dual-gate voltages, the device can dynamically alter its polarization response, achieving a polarization ratio that tends toward infinity. This configuration achieved high reconstruction accuracy for the S₁ and S₂ Stokes parameters, with RMSE values of 1.5% and 2.0%. Operating at room temperature and zero bias, it delivered a specific detectivity of 2.8 × 10⁵ Jones and a responsivity of 0.51 mA/W under 4.75 µm mid-infrared light illumination⁹⁵.

Furthermore, Chen et al.³² introduced optoelectronic polarization eigenvectors by integrating chiral metasurfaces with a MoS₂ channel to create four polarization-sensitive units (Fig. 7d). A linear transfer matrix relates the photocurrents from these units to the four Stokes parameters. The resulting MoS₂-metasurface detector offered a f_−3dB of 25 kHz and a responsivity of 0.6 mA/W at 1550 nm. Full-Stokes reconstruction is accomplished with machine-learning algorithms: a Gaussian-process regression model was trained on thousands of data points spanning diverse polarization states and power levels, each point pairing an input Stokes vector with its corresponding photocurrent vector. The trained model provides high-coverage, high-precision recovery of the Stokes parameters across the entire Poincaré sphere (Fig. 7d), achieving RMSE of only 0.13 % (S₀), 0.98 % (S₁), 0.96 % (S₂), and 0.58 % (S₃), matching the performance of commercial polarimeters. Moreover, its chip-scale footprint, measuring only tens of micrometres compared with the hundreds of millimetres required by commercial systems, combined with the elimination of bulky waveplates³², represents a significant step toward highly miniaturized polarimetric sensing.

Figure 7e summarizes the performance of polarimeters based on twisted 2D vdW materials and 2D vdW material-metasurface heterostructures, including RMSE values for Stokes parameter reconstruction and other detection metrics. While responsivity and detectivity vary across devices, no clear correlation is observed with reconstruction accuracy. This inconsistency likely stems from differences in material properties, device design, and reconstruction algorithms. Moving forward, efforts should focus on developing more compact devices with fewer detection channels that still cover a broad range of polarization states. Simultaneously, optimizing reconstruction algorithms will be key to improving accuracy. Together, these strategies offer a promising route to high-performance, miniaturized polarimeters.

Future challenges and perspectives

In this Review, we have summarized recent advances in 2D computational photodetectors for multidimensional optical information perception, which involves the acquisition of optical intensity, spectral composition, and polarization states. The fundamental detection capabilities include static and dynamic light intensity, spectral characteristics across a broad wavelength range, and various forms of polarization. We further discussed the key physical mechanisms and device architectures that enable these functionalities. Representative innovations including neuromorphic vision sensors, computational spectrometers, and miniaturized polarimeters based on 2D vdW materials were highlighted. These advances lay the foundation for intelligent optoelectronic systems that integrate light detection and preliminary signal processing, offering significant potential for applications in imaging, communication, and machine vision.

Although substantial progress has been made, research on multidimensional optical perception based on 2D computational photodetectors remains at an early stage. Future development can be further guided by both information demand and device innovations. Figure 8a illustrated the need for detecting intensity, spectral, and polarization dimensions. For intensity detection, natural light levels span an enormous dynamic range from overcast night to direct sunlight, varying by as much as 280 dB⁴³. This necessitates optoelectronic devices capable of perceiving subtle contrasts across multiple intensity levels, enabling accurate detection of both shadowed and brightly lit regions. In real-world scenarios such as tunnel entrances and exits, where light contrast can be extreme, the ability to extract multilevel intensity information and perform in-sensor preprocessing is critical for enhancing visual perception and ensuring safety in autonomous driving systems. From the spectral perspective, ultraviolet (200–400 nm) and mid-infrared (3–5 µm and 8–14 µm) spectral windows carry critical physical and chemical signatures that are invisible to the visible-near-infrared bands. Developing hyperspectral detection capabilities in these regimes holds significant promise for applications in biomedical imaging, environmental sensing, and biochemical analysis. On the polarization front, detecting complex vector beams such as polarized vortex beams and optical Skyrmions that carry both orbital-angular-momentum (OAM) phase and diverse polarization states, remains largely unexplored. Successfully capturing these structured fields could unlock high-capacity, tamper-resistant free-space links and drive advances in encrypted optical communication^104–106.

Fig. 8. Potential research directions for 2D computational photodetectors.

a Information demands include multilevel intensity detection with a broad dynamic range, hyperspectral imaging in the ultraviolet (UV) and mid-infrared (MIR) regions, and the detection of polarized vortex beams and optical Skyrmions carrying both orbital angular momentum (OAM) and diverse polarization states. b Device design strategies focus on enhancing performance metrics such as response speed, detectivity, and efficiency; enabling all-optical information detection, including intensity, wavelength, polarization, phase, and spatiotemporal features; and integrating multiple functional units (e.g., detection, memory, and processing) onto a single chip.

With the advent of the artificial-intelligence era, information transmission has increased dramatically, accompanied by growing demands for data security. Multidimensional optical information, including intensity, wavelength, polarization, orbital angular momentum (OAM) phase, and spatiotemporal distribution, has emerged as a promising carrier for high-capacity and encrypted transmission. By encoding various signal sources in multiple optical dimensions and transmitting them synchronously, communication channels can be significantly expanded while inherently enabling encryption. This, in turn, necessitates advanced decoders capable of simultaneously extracting and interpreting these multidimensional parameters. 2D computational photodetectors have demonstrated strong potential in this domain. Leveraging machine learning, these devices can simultaneously detect optical intensity, polarization, and wavelength. Extending this capability to even higher-dimensional parameters, however, remains an open challenge. It imposes stringent requirements on device architecture, demands precise control and multiplexing strategies, and involves solving complex nonlinear relationships between multidimensional optical inputs and electrical outputs. Large-scale artificial-intelligence models are expected to play a critical role in overcoming these challenges. On one front, inverse design and reinforcement learning can be employed to optimize programmable 2D metamaterials and vdW heterostructures, enabling fine-tuned control over complex optical field interactions. On another, end-to-end learning frameworks can effectively model the intricate nonlinear mappings between input parameters and output signals, facilitating real-time reconstruction, classification, and identification of multidimensional information. Together, these approaches will lay the groundwork for the next generation of secure, high-throughput, multidimensional optical communication and sensing technologies.

To meet these evolving demands, future device development should be oriented around three key aspects: performance, functionality, and integration, as summarized in Fig. 8b. From a performance standpoint, achieving multi-level intensity perception with sub-microsecond response times, an ultra-wide dynamic range exceeding 200 dB, and specific detectivity approaching theoretical limits, requires the co-optimization of electrical transport and optical confinement in 2D vdW materials. Embedding 2D vdW heterostructures in high-Q resonant cavities or optical metasurfaces, as well as adopting avalanche-type architectures, can substantially strengthen light-matter interaction and deliver internal gain^{97,100,101,107–116}, providing promising strategies for remarkable detection performance. As the footprint of on-chip spectrometers shrinks toward the wavelength scale, achieving high sensitivity and spectral resolution becomes increasingly critical. Bound-state-in-continuum (BIC) optical metasurfaces exhibit strong electromagnetic confinement, enabling the effective capture of weak spectral signals. Additionally, their intrinsically narrowband filtering characteristics offer improved wavelength selectivity. Therefore, patterning 2D vdW materials into BIC architectures or integrating them with BIC metasurfaces presents a promising pathway toward reconfigurable 2D computational spectrometers with enhanced sensitivity and resolution¹¹⁷. Accurate polarimetric detection relies on the precise reconstruction of the Stokes parameters. Recent progress suggests that integrating empirical models with machine-learning algorithms provides an effective strategy to resolve the complex nonlinear inverse problem, enabling high-fidelity retrieval of polarization information³².

In terms of functionality, the goal is to achieve all optical information detection that encompasses not only light intensity, wavelength, and polarization but also phase and spatiotemporal parameters. However, strong coupling among these dimensions complicates the establishment of linear mappings between optical inputs and electrical outputs, posing challenges for accurate reconstruction and increasing the risk of signal distortion. Data-driven methods, particularly those based on machine learning, offer effective nonlinear modeling solutions and enable high-precision reconstruction when integrated with detector hardware^4,31. Beyond all optical information detection, neuromorphic in-sensor processing of multidimensional optical information can be performed by integrating biomimetic perception and processing functionalities into the spectra- and polarization-sensitive devices. In nature, species such as honeybees and mantis shrimps utilize linear and circular polarization for navigation and communication. These biological strategies inspire the integration of engineered anisotropic or chiral materials with neuromorphic device architectures, such as charge-trapping synaptic devices or floating-gate memory phototransistors, to enable polarization-sensitive in-sensor processing^37,68,69. Moreover, the combination of reconfigurable spectral response and neuromorphic processing function in a 2D computational spectrometer could facilitate real-time spectral information reconstruction, encoding, and classification^118–120.

From the perspective of integration, 2D vdW materials offer significant advantages due to their vdW stacking nature, which avoids lattice-matching constraints. This characteristic enables the seamless integration of multiple functional units on a single chip, such as light detection, memory storage, and local processing. Such monolithic integration not only reduces the overall system footprint but also improves energy efficiency and processing speed^2,7,9,121. The realization of fully integrated 2D optoelectronic systems requires coordinated progress across three key domains. First, wafer-scale growth via vapor-phase epitaxy (e.g., Czochralski growth¹²²) must deliver large-area 2D vdW materials with high crystallinity, low defect densities, and excellent reproducibility, using scalable and cost-effective fabrication protocols. Second, the development of residue-free transfer (e.g. printing transfer technique¹²³), low-temperature metallization (e.g., In, Bi, Cd)^114,124 and dielectric deposition (e.g., Sb₂O₃^115,125, Ga₂O₃¹²⁶) techniques is essential to maintain 2D vdW materials quality during large-scale array fabrication processing. In situ optical metrology combined with machine-learning-assisted process control, should be employed to monitor thickness, strain, and defect distributions in real time, ensuring device-level uniformity and reliability. Third, back-end-of-line integration with CMOS or monolithic silicon-photonics wafers enables 3D stacking of sensing, memory, and logic computing layers^127–130. In parallel, machine learning algorithms are expected to facilitate the co-design of hardware and signal-processing schemes¹³¹, enabling the evaluation of key process parameters that influence the response characteristics of 2D computational photodetectors and supporting the development of compact systems for real-time multidimensional optical information perception and processing.

Author contributions

F.W. and Q.J.W. conceived this work. F.W. searched for literature data for this article. F.W. and S.F. structured and wrote the manuscript. F.W., S.F. and Q.J.W. reviewed and revised the manuscript. F.W., S.F., Y.Z. and Q.J.W. contributed to the discussion of the content.

Peer review

Peer review information

Nature Communications thanks Yang Wang and the other, anonymous, reviewer for their contribution to the peer review of this work.

Competing interests

The authors declare no competing interests.