Atomically thin heterojunction-based optoelectronic synaptic devices operable from −190 °C to 450 °C

Abstract

Exploration and operation in extreme environments (e.g., planetary surfaces) require hardware capable of sensing, storing, and processing visual information in situ, enabling intelligent perception under conditions inaccessible to humans. Optoelectronic synaptic devices (OSDs) are a promising class of such hardware, but existing OSDs fail under extreme temperatures, making thermal survivability a key bottleneck for extreme environment intelligent vision. Here, via the interface design of 2D-material heterostructures, we develop a MoS₂-based OSD operating across a 640 °C range (from −190 °C to 450 °C) with tunable neuromorphic functionality. At 450 °C, the device achieves visible detection, bridging the spectral gap of high-temperature optical sensors. It exhibits highly tunable synaptic plasticity, with a retention tunability ratio exceeding reported high-temperature OSDs by 1600×. Leveraging bio-inspired scotopic and photopic visual adaptation, the device facilitates in-sensor processing, enhancing low-light image recognition accuracy by more than 15% across multiple datasets. A conceptual robotic rover demonstration illustrates its potential applications. This work provides a route toward robust, intelligent vision perception for planetary exploration and other extreme-environmental applications.

Here, the authors report the realization of thermally resistant optoelectronic synaptic devices based on 2D MoS₂ operating across a 640 °C temperature range, showing a promising approach for the development of intelligent vision devices for planetary exploration and other extreme-environment applications.

Introduction

Exploration and operation in extreme environments pose some of humanity’s greatest technical and scientific challenges. Planetary exploration stands as one of the ultimate frontiers for expanding humanity’s cognitive boundaries. It holds the key to answering fundamental scientific questions, such as the origin of life and the evolution of the solar system, while laying the groundwork for interplanetary resource utilization. Globally, numerous missions are already underway, exemplified by NASA’s Perseverance rover (targeting Mars) and China’s Tianwen missions (targeting Mars and Venus). As exploration objectives become more complex and autonomous, traditional remote-controlled rovers face limitations, including long signal latency, uncharted climates and terrain, and stringent requirements for real-time scientific decision-making, necessitating a transition toward embodied intelligent agents with in situ perception and on-board intelligence.

Intelligent visual perception systems serve as the primary interface between such agents and their environments, requiring hardware capable of sensing, storing, and processing visual information in situ to enable autonomous decision-making under resource constraints. However, extreme thermal environments pose a fundamental constraint: planetary surfaces range from cryogenic cold to intense heat (e.g., −180 °C to 430 °C on Mercury)¹, under which sensing, memory, and computing hardware struggle to operate reliably. Under such conditions, conventional Von Neumann architectures, which implement these functions in physically separated units, suffer from inherent system fragility: additional components increase volume and power consumption and make overall operation vulnerable to single-point failures, posing a critical risk for planetary missions.

Optoelectronic synaptic devices (OSDs), which integrate sensing, memory, and processing within a single non-von Neumann architecture, address the reliability, latency, and data redundancy issues of conventional separated architectures^2–6. Nevertheless, existing OSDs show limited thermal stability, failing above ~350 °C and rarely demonstrated at cryogenic temperatures, rendering them incompatible with planetary thermal extremes. Additionally, reported high-temperature optical devices (OSDs and photodetectors) are spectrally limited because they typically rely on wide-bandgap (WBG) semiconductors (e.g., SiC, Ga₂O₃), which are insensitive to visible light, leading to the loss of critical visible-light information. Existing OSDs not only struggle to survive but also have limited sensing, memory, and processing performance in extreme environments.

Herein, we address these challenges by developing OSDs based on a designed MoS₂/ hexagonal boron nitride (h-BN)/graphene van der Waals (vdW) heterostructure. Through the synergistic protection of h-BN and graphene, we extended the operational range of OSDs to a broad 640 °C range (−190 °C to 450 °C), covering the thermal extremes of most terrestrial planets and the Moon. The devices exhibited no measurable change in electrical performance after 30 min of operation under extreme temperatures, with excellent sensing performance even after 5 h. Under extreme temperatures, they enabled broadband UV-to-visible photodetection, overcoming the intrinsic spectral blindness of WBG optical sensors. In parallel, by benefiting from the dangling-bond-free vdW interface and a low Schottky barrier, they achieved robust neuromorphic functionality at extreme temperatures, exhibiting highly tunable synaptic plasticity for memory and in-sensor processing. They achieve a high dynamic memory range at 450 °C, surpassing state-of-the-art high-temperature OSDs by a factor of ~1600. Furthermore, leveraging bio-inspired scotopic and photopic visual adaptation for adaptive in-sensor processing, the device dynamically adapts to variable illumination scenarios, enhancing low-light image recognition accuracy by more than 15% across multiple datasets. Sensing tests by integrating the OSD on a robotic dog rover were performed. This work paves the way towards robust, intelligent vision perception systems for planetary exploration and other harsh-environment applications.

Results

Structural and thermal characterization of MoS₂ OSD

To enable intelligent visual perception for planetary exploration and other extreme-environment applications (Fig. 1a), MoS₂ OSDs were fabricated by mechanically stacking the bottom h-BN, MoS₂ channel, graphene contacts, and top h-BN sequentially on SiO₂/Si substrate. The source, drain and gate electrodes were manufactured by photolithography, metal deposition, and metal liftoff. Figure 1b–e are the schematic views and optical microscopy image of the MoS₂ OSD. Atomic force microscopy (AFM) characterization confirms the thicknesses of MoS₂ channel (~3.8 nm) and graphene contact (~8 nm), along with the top (~18 nm) and bottom h-BN layers (~32 nm) (Supplementary Figs. 1, 2). Raman spectroscopy verifies the high crystalline quality of all 2D (two-dimensional) materials used in this work (Supplementary Fig. 3)^7–11. High-resolution transmission electron microscopy (HRTEM) cross-sectional image of the OSD reveals atomically flat interfaces within the h-BN/graphene/MoS₂/h-BN heterostructure, free from noticeable defects or air gaps (Fig. 1f). The measured thickness of a single MoS₂ layer is 0.7 nm, consistent with reported values¹². Built upon this robust vdW heterostructure, the devices successfully integrate broad-spectrum sensing, highly tunable memory, and adaptive in-sensor processing capabilities (Fig. 1g–i).

Fig. 1. Bio-inspired MoS₂ optoelectronic synaptic devices (OSDs) for extreme-environment applications.

a Schematic of a rover equipped with the device conducting planetary exploration under extreme temperatures (−190 °C to 450 °C). b–e Device architecture: b Three-dimensional schematic, where V_gs and V_ds denote the gate-source voltage and drain-source voltage, respectively. c Device photograph. d Cross-sectional schematic of the hexagonal boron nitride (h-BN)/graphene/MoS₂/h-BN heterostructure. e Optical microscope image, where S (source), G (gate), and D (drain) denote the respective electrodes. f High-resolution transmission electron microscopy (HRTEM) cross-sectional image of the heterostructure. g Comparison of broad-spectrum sensing performance (temperature vs. wavelength) between this work (red) and reported high-temperature optical devices (black). The dashed line indicates the trade-off between wavelength and working temperature. h Highly tunable synaptic memory behavior exhibiting transition from short-term to long-term plasticity tuned by gate voltage. i Concept of in-sensor processing demonstration. Synthetic input images (bright/dark light) are algorithmically processed using the experimentally measured linear and nonlinear photoresponse curves of the OSDs to generate high-fidelity and enhanced output images, respectively.

Via the designed 2D-material heterostructure, the device exhibits robust structural and electrical stability across a wide temperature range from −190 °C to 450 °C. As shown in Fig. 2a, its transfer curves remained virtually unchanged during continuous operation in air at −190 °C and 450 °C for 30 min. In contrast, bare MoS₂ devices without h-BN encapsulation exhibited severe degradation after only 5 min at 300 °C (Supplementary Fig. 4). Raman spectra further confirmed no oxidation or structural degradation in the MoS₂ channel following these extreme-temperature tests (Fig. 2b). Even when the operation time under extreme temperatures was increased to 5 h, the transfer curves showed only negligible shifts, with sensing performance remaining uncompromised (Supplementary Fig. 5). In addition, the device demonstrates robust electrical fatigue resistance. Its transfer curves maintained remarkable consistency without observable threshold voltage shifts even after 20 continuous sweep cycles at 450 °C (Supplementary Fig. 6). This excellent extreme-temperature tolerance arises from the engineered vdW heterostructure that combines h-BN encapsulation and graphene contacts (Fig. 2c). As revealed by HRTEM (Figs. 1f, 2c), these materials form an atomically flat and conformal interface free of voids and defects, which is critical because ambient reactive species (e.g., water and oxygen) can readily oxidize or dope MoS₂ at extreme temperatures, especially high temperatures. By effectively isolating the channel from these species, the atomically flat h-BN/graphene interface acts as a hermetic barrier that suppresses diffusion pathways, thereby preserving device stability under extreme temperatures. Typical TMDs, including MoS₂, undergo oxidation and structural degradation at temperatures around 300 °C (Supplementary Table 2). Conversely, our engineered vdW heterostructure substantially expands the operational temperature window of TMDs.

Fig. 2. Robust extreme-temperature operation and broadband sensing of MoS₂ OSD.

a Transfer curves at 450 °C and −190 °C after 30 min of continuous operation (V_ds = 0.5 V). b Raman spectra of MoS₂ channel before and after 5 h thermal treatment at 450 °C and −190 °C. c Schematic and structural characterizations of the protection mechanism. The top panel illustrates the device stability against O₂ and H₂O penetration for graphene versus metal contacts. The middle and bottom panels present the cross-sectional high-resolution transmission electron microscopy (HRTEM) images (h-BN/graphene vs. h-BN/Pt) and the surface atomic force microscopy (AFM) topography images (graphene vs. Pt electrode), respectively. d Transfer curves in dark, illuminated, and post-illumination states at 450 °C and −190 °C (V_ds = 0.5 V). e Temperature-dependent photocurrent under 254 nm, 450 nm, and 530 nm illumination. f, g Time-resolved photoresponse at 450 °C (f) and −190 °C (g). h Radar chart benchmarking performance against state-of-the-art high-temperature OSDs.

To understand the critical role of electrode and encapsulation materials in device stability, we compared graphene contacts and h-BN encapsulation with conventional metals and deposited oxides. When the graphene electrodes are replaced with conventional metal electrodes (e.g., Ti, Pt), the h-BN/metal interface exhibits significant roughness (Fig. 2c), creating nanoscale gaps that serve as diffusion channels. AFM measurements further substantiate the structural origin: the graphene surface shows a maximum height variation of only ~0.6 nm, whereas the metal electrodes exhibit peak-to-valley variations as large as ~6.8 nm. Beyond topography, the chemical and structural incompatibility at metal/h-BN interfaces further hinders conformal contact. The substantial topographical mismatch and structural incompatibility prevent metals from forming the same high-quality, conformal encapsulation achievable with graphene, leading to significantly compromised stability. Consequently, metal-contacted (Ti-contact) devices degrade rapidly at 300 °C within 3 min, indicating fast oxidation (Supplementary Fig. 7). In contrast, h-BN and graphene, as layered 2D materials, naturally form intimate vdW heterostructures that can maintain effective encapsulation even at high temperatures. When the h-BN encapsulation is replaced with conventionally deposited oxides (e.g., HfO₂, Al₂O₃ or SiO₂), the device fails to sustain high-temperature operation. The control device encapsulated with ALD-grown HfO₂ exhibits performance degradation after only 5 min at 250 °C, evidenced by a drastic shift in threshold voltage (Supplementary Fig. 8). This failure stems from the intrinsic porosity of deposited oxides¹³. Unlike the impermeable crystalline lattice of h-BN, ALD-grown films often contain pinholes and grain boundaries that serve as diffusion pathways^14,15. These defects allow ambient oxygen to penetrate and oxidize the MoS₂ channel, failing to provide the high-quality protection achievable with the all-2D vdW heterostructure. Consequently, the synergistic combination of h-BN and graphene is pivotal for this protective functionality^13,16.

To experimentally evaluate the optoelectronic functionality of the device under extreme temperatures, we characterized its photoresponse at both 450 °C and −190 °C. Transfer curves under dark, illuminated, and post-illumination conditions are shown in Fig. 2d. Upon light exposure, the device exhibited a substantial increase in source-drain current (I_ds), demonstrating its light-sensing capability at both temperatures. We further measured the temperature-dependent photocurrent (I_ph) under 254 nm, 450 nm, and 530 nm illumination (Fig. 2e). Although I_ph naturally shows a negative correlation with temperature, which is attributed to enhanced phonon scattering and carrier recombination, the device maintains a distinct photoresponse across all tested wavelengths throughout this wide thermal window. Time-resolved photoresponse measurements were subsequently performed under four different wavelengths (254 nm, 365 nm, 450 nm, and 530 nm) at 450 °C (Fig. 2f) and −190 °C (Fig. 2g). As depicted, the device exhibits distinct and repeatable I_ph across the UV-to-visible spectrum at both temperatures. Specifically, under 530 nm illumination at −190 °C, the memory response disappears and transitions into a fast response. We attribute this to carrier freeze-out and insufficient photon energy, which collectively restrict the carriers from accessing the localized states responsible for synaptic behavior (detailed discussion in Supplementary Note 1). Furthermore, the device demonstrates robust thermal stability, with its electrical and optoelectronic properties remaining highly consistent after undergoing 5 continuous temperature cycles between −190 °C and 450 °C (Supplementary Fig. 9). Conventional high-temperature photodetectors are mostly based on wide-bandgap (WBG) semiconductors (e.g., Ga₂O₃, SiC, and GaN), whose large bandgaps render them transparent to visible light, causing “spectral blindness” in the visible regime above 300 °C (Fig. 1g and Supplementary Table 1)^17–26. In contrast, by leveraging the engineered vdW heterostructure, MoS₂ can operate at high temperatures, and its relatively narrow bandgap ensures efficient absorption in the visible range, enabling broadband detection that WBG materials cannot achieve.

As summarized in Fig. 2h and Supplementary Table 3, the device exhibits high sensing performance across an ultra-wide temperature range of 640 °C (from −190 °C to 450 °C), maintaining stable operation for >30 min at extreme temperatures, whereas existing OSDs are typically limited to below 350 °C and rarely report stability beyond ~5 min. This unique combination of ultra-wide temperature adaptability, broadband visible perception, and high stability covers the environmental variations of most solar system planets (Supplementary Table 4), underscoring the device’s remarkable potential for future space exploration missions.

Gate-tunable synaptic memory at extreme temperatures

Biological synapses transmit information via neurotransmitters triggered by presynaptic action potentials, generating excitatory postsynaptic currents (PSCs) that decay over time (Fig. 3a)⁶. Mimicking this behavior, our device utilizes light pulses as presynaptic spikes, and I_ph acts as the postsynaptic response. Upon illumination, I_ph increased rapidly and decayed slowly after illumination removal, exhibiting persistent photoconductivity (PPC) arising from charge trapping in sulfur vacancies, interfacial defects, and stress-induced states in MoS₂^27–29.

Fig. 3. Emulation of synaptic plasticity and memory behaviors at 450 °C.

a Schematic diagram of a biological synapse. b Paired-pulse facilitation (PPF) index as a function of pulse interval (Δt). Inset: PPF measurement protocol, where A₁ and A₂ denote the amplitudes of the postsynaptic current (PSC) triggered by the first and second light pulses, respectively. c–f Short-term plasticity (STP) to long-term plasticity (LTP) transition at 450 °C induced by increasing pulse number (c), intensity (d), frequency (e), and width (f). g Schematic diagram of the “learning-forgetting-relearning” processes based on short-term memory (STM) and long-term memory (LTM). h Four consecutive “learning-forgetting-relearning” cycles measured at 450 °C, where ∆I_1-4 represent the current remnants extracted at the end of each forgetting phase, and the blue shaded areas represent the light stimulation periods.

We further investigated the device’s capability to emulate synaptic plasticity at 450 °C under 365 nm illumination. Synaptic plasticity, the neurobiological basis for learning and memory, is categorized into short-term plasticity (STP) for rapid information processing and long-term plasticity (LTP) for sustained memory retention⁶. Paired-pulse facilitation (PPF), a key form of STP, describes the phenomenon where the response to a second spike is enhanced relative to the first^6,26. Figure 3b presents the PPF behavior measured at 450 °C by two consecutive light pulses (3.2 mW/cm², 365 nm) with varying pulse interval (Δt). The PPF index, defined as A₂/A₁ ×100% (A₁, A₂: first and second I_ph amplitudes), decays with Δt and can be fitted using a double-exponential function³⁰:

$$PPF=c_1\times \exp \left(-\frac{\mathrm{\Delta }t}{{\tau }_1}\right)+c_2\times \exp \left(-\frac{\mathrm{\Delta }t}{{\tau }_2}\right)$$ 1

where $c_1$ and $c_2$ are the corresponding initial facilitation magnitudes, and ${\tau }_1$ and ${\tau }_2$ are the characteristic relaxation times of rapid and slow decay. The ${\tau }_1$ and ${\tau }_2$ extracted are 0.334 s and 3.41 s respectively, corresponding to the immediate response and sustained memory timescales in biological synapses.

Beyond short-term dynamics observed in PPF, the consolidation of synaptic strength into LTP is essential for permanent information storage³¹. Functionally, this plasticity evolution underpins the device’s capability to consolidate transient short-term memory (STM) into persistent long-term memory (LTM). We successfully emulated this critical transition at 450 °C by modulating optical stimulation parameters. As illustrated in Fig. 3c, increasing the number of light pulses from 1 to 30 resulted in a pronounced potentiation of the PSC from 5.53 to 15.88 nA. Similarly, enhancing the light intensity (Fig. 3d), increasing the stimulation frequency (Fig. 3e), or prolonging the illumination duration (Fig. 3f) all led to significantly extended retention times (stronger memory) and higher synaptic weights. To further evaluate the device’s capability for memory processing, we simulated the “learning-forgetting-relearning” behavior, which follows the Ebbinghaus forgetting curve (Fig. 3g)³⁰. This process involves rapid initial decay after learning, partial retention over time, and reinforcement through relearning. As shown in Fig. 3h, the device was subjected to four consecutive cycles comprising a learning state (0.5 s) and a forgetting state (10 s) at 450 °C. The residual PSC baseline increased progressively with each cycle (∆I₄ > ∆I₃ > ∆I₂ > ∆I₁). This accumulation of conductance demonstrates that relearning effectively strengthens memory retention, facilitating the consolidation of information from STM to LTM. Furthermore, the total energy consumption per synaptic event at 450 °C is calculated to be ~2.41 nJ (comprising ~2.36 nJ of incident optical energy and ~50 pJ of electrical reading energy, detailed in Supplementary Fig. 10 and Note 2). Collectively, these results demonstrate that the device reliably reproduces key cognitive functions of the human visual system from dynamic synaptic plasticity to memory consolidation and reinforcement at 450 °C, supporting its potential for neuromorphic applications in extreme environments.

Beyond optical modulation, the device also exhibits excellent gate-tunable memory (GTM) capabilities, enabling memory modulation across a wide dynamic memory range through V_gs tuning. As shown in Fig. 4a, the decay dynamics of normalized I_ph at 450 °C are highly sensitive to V_gs. As V_gs shifts from −4 to 0 V, the device transitions from a fast-decay state to a slow-decay state, referred as STM-to-LTM modulation. This GTM behavior was observed under 254, 365, 450, and 530 nm illuminations (Supplementary Fig. 11). To characterize its tunability, we applied a train of five light pulses (365 nm) to the device at V_gs = -4 V and V_gs = 0 V, respectively. The 365 nm wavelength was chosen because it provides the largest STM-to-LTM dynamic range. Shorter wavelengths (e.g., 254 nm) result in an overly prolonged STM retention, while longer wavelengths (e.g., 530 nm) make it difficult to achieve a long-duration LTM (Supplementary Fig. 11). As V_gs = −4 V, the I_ph exhibited non-accumulative photoresponse, with a decay time of 40.9 ms (Fig. 4b). In contrast, as V_gs = 0 V, I_ph accumulated with each light pulse and decayed slowly (Fig. 4b). In this state, PSC required ~660 s to return to baseline (I_dark + 3σ noise), demonstrating long signal retention. At lower temperatures, slower charge detrapping further enhanced LTM performance (Supplementary Fig. 12).

Fig. 4. Gate-tunable synaptic memory and working mechanism.

a Normalized transient photoresponse and decay curves under varying V_gs (from −4 to 0 V). b Distinct memory states: long-term memory (top, V_gs = 0 V) and short-term memory (bottom, V_gs = −4 V), where t_re denotes the retention time. c Benchmark of dynamic memory range and retention tunability ratio, comparing this work (red star) with reported high-temperature optoelectronic synaptic devices (OSDs). d Band diagrams illustrating the mechanism of gate-modulated charge trapping/detrapping dynamics under high V_gs (top) and low V_gs (bottom), where E_C, E_F and E_V represent the conduction band minimum, Fermi level, and valence band maximum, respectively; the blue circles with ‘-‘ and red circles with ‘+‘ represent electrons and holes; and the arrows indicate carrier trapping and detrapping processes. e Transfer curves of MoS₂ OSD with SiO₂ dielectric measured in dark, illuminated, and post-illumination state at room temperature (V_ds = 0.5 V). Inset: photoresponse at V_gs = −12 V and V_gs = −2 V. f Schematic illustrating the operating principle of the device with SiO₂ dielectric. The blue circles with ‘-‘ and red circles with ‘+‘ represent electrons and holes; and the red arrows indicate carrier trapping processes. g Transfer curves of MoS₂ OSD with Pt electrode measured in dark, illuminated, and post-illumination state at room temperature (V_ds = 0.5 V). Inset: photoresponse at V_gs = -3.5 V and V_gs = -0.5 V. h Schematic view of the mechanism of the device with Pt electrode. The blue circles with ‘-‘ and red circles with ‘+‘ represent electrons and holes; and the red arrows indicate carrier transport processes.

To quantify this modulation window, we define the dynamic memory range as the absolute time span between the minimum short-term memory time (t_{STM_min}) and the maximum long-term memory time (t_{LTM_max}) under identical temperature and illumination conditions. The retention tunability ratio is correspondingly calculated as t_{LTM_max}/t_{STM_min}. Under 365 nm illumination at 450 °C, our device exhibits a dynamic range from 40.9 ms (V_gs = -4 V) to 660 s (V_gs = 0 V), yielding a high tunability ratio of 16,137 (detailed extraction criteria in Supplementary Fig. 13 and Note 3). This ratio is 1600-fold higher than the typical values (<10) of state-of-the-art high-temperature OSDs^{17,26,32–34} (Fig. 4c), even operating at higher temperatures. GTM behavior was also observed at −190 °C (Supplementary Fig. 14), demonstrating that this phenomenon persists across a wide temperature range. This broad temporal dynamic range effectively emulates the biological spectrum of STM and LTM, endowing the device with the adaptive functionality essential for processing both transient and persistent visual information during planetary exploration.

The mechanism underlying the GTM behavior is attributed to the modulation of trap occupancy within the MoS₂ bandgap (Fig. 4d). MoS₂ inherently possesses defect states arising from sulfur vacancies, interfacial defects, and stress-induced states, acting as shallow and deep traps^27–29. The KPFM potential mapping clearly reveals distinct surface potential fluctuations on the MoS₂ channel, directly confirming the presence and distribution of these localized defect states (Supplementary Fig. 15). The occupation of trap states follows a statistical process, with the probability of electron trapping governed by the Fermi-Dirac distribution³⁵:

$$f_t\left(E_t\right)=\frac{1}{1+e^{\left(\frac{E_t-E_F}{k_{B}T}\right)}}$$ 2

where $E_t$ is the trap energy level, $E_F$ is the Fermi level, $k_B$ is the Boltzmann constant, and T is the absolute temperature. At higher V_gs (e.g., 0 V), the increased electron concentration shift E_F toward the conduction band (E_C). According to Eq. (2), this significantly increases the occupation probability for deep traps, allowing them to effectively capture and store photogenerated carriers. Upon light removal, carrier detrapping from these deep states is thermally activated and slow, resulting in LTM behavior. The remarkably long retention time (~660 s) is enabled by the atomically flat and trap-free h-BN interface, which suppresses carrier recombination and diffusion. As a result, photogenerated electrons trapped in these states can persist for extended periods even at high temperatures. Conversely, at lower V_gs (e.g., −4 V), E_F shifts downwards away from the trap levels, drastically reducing the occupation probability of deep traps. Consequently, photogenerated carriers interact primarily with shallow traps or recombine rapidly, leading to fast STM response. Additionally, trapped electrons can be actively released by applying a negative voltage pulse (-6 V), enabling electrical erasing (Supplementary Fig. 16).

The realization of GTM relies critically on the quality of the gate dielectric and contact barrier. Control devices utilizing conventional 3D oxides (SiO₂ and Al₂O₃) exhibited uniformly slow photoresponse across all V_gs (Fig. 4e and Supplementary Fig. 17), devoid of the tunability observed in h-BN devices. The slow response is attributed to the abundant defect states and dangling bonds characteristic of amorphous oxides (SiO₂, Al₂O₃, HfO_2, etc.)^14,15,36. These defects tend to capture photogenerated carriers, which delays the recombination process and results in a slow photoresponse. In contrast, the atomically pristine surface of 2D h-BN minimizes such extrinsic trapping, thereby facilitating efficient carrier transport and enabling the rapid photoresponse required for the STM state at low V_gs. Virtually no hysteresis was observed in the bidirectional transfer curves of h-BN device, which further confirms the low trap density at the h-BN surface (Supplementary Fig. 18)³⁷.

Furthermore, the contact barrier plays a decisive role in realizing GTM behavior. Linear I_ds-V_ds curves confirm that MoS₂ forms low-resistance contacts with graphene, corresponding to a small Schottky barrier, enabling efficient carrier injection (Supplementary Fig. 19). To verify the effect of barrier height, we fabricated control devices using high-work-function platinum³⁸ (Pt, ~5.6 eV). This created a substantial Schottky barrier at the MoS₂/Pt interface, evidenced by the suppression of electron injection and the resulting p-type behavior (Fig. 4g), which is consistent with previously reported Pt-contacted MoS₂ transistors^38,39. The Pt-contacted device demonstrated fast photoresponse and recovery across all V_gs, devoid of memory tunability (Fig. 4g). This phenomenon stems from the strong built-in electric field at the Schottky junction, which efficiently separates photogenerated electron-hole pairs and sweeps them out before they can be captured by traps. Consequently, the trap occupancy remains low, preventing LTM formation. In contrast, when MoS₂ was contacted with titanium (Ti), a low-work-function metal (~4.3 eV), the Schottky barrier was significantly reduced⁴⁰. The Ti-contacted device exhibited GTM behavior similar to graphene-contacted devices (Supplementary Fig. 20), further confirming that a low contact barrier is a prerequisite for GTM behavior. While both Ti and graphene provide a low Schottky barrier, graphene not only forms a lower barrier than Ti but also offers more stable contact characteristics, which are crucial for maintaining the electrical stability at extreme temperatures¹³.

These findings confirm that the synergy of a trap-free 2D dielectric (h-BN) and a low-barrier contact (graphene) is indispensable for enabling GTM behavior. Other factors, such as electrode position (top gate or bottom gate), substrate type (sapphire or mica), and the presence or absence of a top h-BN layer, have no discernible impact on the GTM behavior (Supplementary Figs. 21–23). Therefore, the integration of h-BN and graphene not only provides the structural guarantee for environmentally robust encapsulation at extreme temperatures but also acts as the functional prerequisite for realizing gate-tunable plasticity.

Bio-inspired scotopic and photopic visual adaptation

In biological systems, adaptive visual perception relies on the synergistic operation of two photoreceptor cells in the retina: the rods and the cones⁴¹. Rods are specialized for scotopic (low-light) vision, offering high sensitivity and nonlinear gain to detect faint signals, though they saturate easily under high intensity⁴² (Fig. 5b). Cones, by contrast, dominate in bright-light conditions (photopic vision), maintaining a linear photoresponse to ensure high-fidelity perception. By dynamically integrating these complementary modalities, the human eye achieves robust adaptation across a broad range of illumination levels, balancing sensitivity and fidelity across an enormous dynamic range (Fig. 5a).

Fig. 5. Bio-inspired gate-tunable scotopic-photopic adaptation and in-sensor image enhancement.

a Schematic of the human eye adjusting to bright (cone-dominated) and weak (rod-dominated) light conditions. b Schematic illustration of the response intensity of cone and rod cells as a function of light intensity. c Normalized photocurrent (I_ph) as a function of incident light power density (P_in) at V_gs = -4 V and V_gs = 0 V, where the data points represent the experimental measurements, and the dashed lines represent the corresponding fitting curves. d Image enhancement demonstration. Synthetic input images are algorithmically processed to show linear imaging under bright light and nonlinear enhancement under weak light, based on the experimentally measured characteristics of the OSDs. e Convolutional neural network (CNN)-based recognition workflow using weak-light and enhanced datasets. f Recognition accuracy of weak-light and enhanced CIFAR-10 datasets as a function of training epochs.

Inspired by this dual-channel visual strategy of rods and cones, our device emulates scotopic and photopic visual adaptation through gate-tunable photoresponse. Under 365 nm illumination, we measured the normalized I_ph as a function of incident light power (P_in) at V_gs = 0 V and −4 V, respectively, and fitted the results using the power law I_ph ∝ P_in^α. At V_gs = 0 V, the device exhibits a strong nonlinear response with α = 0.45 (Fig. 5c), characteristic of scotopic vision, providing high sensitivity to weak light while preventing saturation. At V_gs = −4 V, the response becomes nearly linear (α = 1.04), analogous to photopic vision, enabling accurate sensing under bright light. Gate-voltage control enables reversible switching between these regimes, producing gate-tunable scotopic-photopic adaptation (GSPA) behavior.

The GSPA behavior originates from gate-controlled activation and suppression of trap-assisted carrier trapping. In the scotopic regime, deep trap states actively participate in photocarrier dynamics and become progressively filled under illumination, thereby dynamically modulating carrier recombination and collection efficiency, leading to a sublinear photoresponse and gain saturation. In contrast, in the photopic regime, trap states are largely suppressed, enabling prompt carrier collection with minimal trapping effects and yielding an approximately linear photoresponse. Crucially, the GSPA behavior is exclusive to the h-BN/graphene heterostructure. Control devices using SiO₂ dielectrics or Pt contacts failed to exhibit GSPA behavior, instead showing either persistent nonlinearity or fixed linearity across all gate voltages (Supplementary Figs. 24, 25). This establishes the h-BN/graphene heterostructure as a necessary physical condition for gate-controlled switching between scotopic and photopic visual sensing modes.

The GSPA capability enables the device to dynamically adapt to drastic illumination variations, which is a critical challenge of planetary surface exploration. Figure 5d illustrates the device’s adaptive image processing capability under bright-light and weak-light conditions. Under bright light, it operates in photopic sensing mode (V_gs = -4 V), where its linear response ensures high-fidelity imaging without saturation. Under weak-light conditions, which are commonly encountered on planetary surfaces, it switches to scotopic sensing mode (V_gs = 0 V), leveraging high sensitivity and nonlinear gain to enhance faint signals at the sensing stage. Benefiting from its time-cumulative photocurrent and high weak-light sensitivity, this scotopic mode can amplify faint image features, achieving a physical signal-to-noise ratio (SNR) nearly 8 times higher than the photopic mode (Supplementary Fig. 26). This direct signal enhancement at the sensing stage clearly evidences the device’s robust in-sensor processing capability.

To evaluate the impact of in-sensor signal enhancement on image recognition, we constructed a convolutional neural network (CNN) evaluation framework based on the ResNet-18 architecture and utilized the CIFAR-10 dataset. To simulate extreme imaging conditions (e.g., planetary surface exploration), the original images were first linearly attenuated to 15% of their original intensity to mimic low-photon input and injected with Gaussian noise (variance σ² = 0.002) to represent the inherent readout noise of sensors. This created a weak-light dataset where visual features are heavily submerged (Fig. 5e). These degraded images were then processed via the device’s non-linear response to generate the corresponding enhanced dataset. We evaluated both systems using the CNN classifier trained for 20 epochs. As illustrated in Fig. 5f, the recognition accuracy of the conventional linear sensor was limited to 66.34% as critical texture details were lost in the noise. In contrast, the proposed in-sensor processing device achieved a remarkably higher accuracy of 81.79%. Detailed analysis via confusion matrices (Supplementary Note 5 and Supplementary Fig. 27) reveals that the device effectively recovers discriminative features for classes with fine textures and complex structures (e.g., ‘Cat’ and ‘Dog’), which were heavily misclassified due to noise-induced feature blurring.

To further assess robustness under real-world weak-light conditions beyond synthetic degradations, we extended this validation to the real-world ExDark (Exclusively Dark) dataset, observing a consistent performance advantage over conventional linear sensors (detailed results are provided in Supplementary Note 5 and Supplementary Figs. 28, 29). These results demonstrate that by modulating the signal in the analog domain, our device enhances the effective signal-to-noise ratio (SNR) prior to digitization, offering a robust hardware solution for low-light vision perception.

Integration of the OSD on a robotic dog rover

To further verify the device’s potential application, we integrated it onto a robotic dog rover to preliminarily simulate its light-sensing capability under simulated extreme environments (Supplementary Movie 1 and Supplementary Fig. 30). Liquid nitrogen and an open flame were employed to simulate low- and high-temperature scenarios, replicating the “detection-moving-detection” exploration workflow. The OSD-equipped robotic dog rover can operate in both low- and high-temperature environments, demonstrating its environmental adaptability and underscoring its promising applications for planetary exploration and other harsh-environment missions.

Discussion

By leveraging atomically thin MoS₂/h-BN/graphene heterojunctions, we have developed OSDs capable of operating across a broad temperature range of 640 °C (from −190 °C to 450 °C), establishing thermal survivability where reported OSDs fail. Building on this robustness, the device achieves UV-to-visible detection at 450 °C, effectively bridging the spectral gap of high-temperature optical sensors and demonstrating high-temperature visible-light optoelectronic synaptic response. Functionally, its excellent gate-tunable memory enables both transient response and persistent storage at thermal extremes. Furthermore, the GSPA behavior dynamically adapts to drastic illumination variations across photopic and scotopic regimes. The combination of extreme-temperature survivability, visible-light sensing, and adaptable neuromorphic computation establishes a foundational hardware platform for autonomous, intelligent vision in harsh environments. This framework opens possibilities for deploying intelligent neuromorphic systems in deep-space planetary exploration missions. Beyond aerospace applications, the robust visual perception capability demonstrated here holds significant potential for terrestrial extreme environments, including deep-well logging and industrial monitoring.

Methods

Device fabrication

We prepared clean silicon/silicon dioxide (300 nm) substrate and mica substrate by immersing them in acetone, alcohol, and deionized water successively for 2 min ultrasonic cleaning. We first fabricated 5 nm Ti /30 nm Pt electrodes using photolithography, metal deposition, and lift-off process. Then, bottom h-BN (30–50 nm), multilayer MoS₂ (3–10 nm), two graphene contacts (5–10 nm), and top h-BN (15–30 nm) were mechanically exfoliated using scotch tape and transferred onto substrate by PDMS stamp. The h-BN bulk crystals were purchased from SixCarbon Technology Shenzhen. The transfer process was carried out using a precision transfer platform (Metatest, E1-T). The annealing process for as-fabricated devices was carried out in Ar atmosphere at 200 °C for 2 h.

Device characterizations

The thicknesses of the 2D materials were characterized by AFM (Bruker, Dimension Icon) using ScanAsyst-air mode. The Raman spectra were derived by Raman spectrometer (HORIBA Jobin Yvon, LabRAM HR Evolution) with 532 nm laser (0.325 mW/cm²) and acquisition time of 120 s. The cross-sectional images of vdW heterostructures were characterized by high-resolution transmission electron microscopy (JEOL-2100 TEM) with acceleration voltage of 200 kV. The samples for HRTEM were prepared using Xenon focused ion beam (Helios G4).

Device measurements

The photoelectric characteristics of the MoS₂ FET under different temperatures were measured by semiconductor parameter analyzer (Agilent B1500A) and variable temperature probe station (HC-P621G-PM). The devices were illuminated by light-emitting diodes (254, 365, 450, and 530 nm) from JCOPTIX, China. The illumination power and time were adjusted by a signal generator (AFG 31000, Tektronix) and calibrated by using a standard light sensor (S130C, Thorlabs).

Image datasets and preprocessing

Publicly available image datasets, CIFAR-10⁴³ and ExDark⁴⁴, were used to evaluate the in-sensor processing capability of the proposed device under low-light and noisy conditions. Distinct pre-processing protocols were applied to simulate specific target environments: For CIFAR-10, images were linearly attenuated to 15% of original intensity with Gaussian noise (σ² = 0.002) injected to mimic extreme planetary exploration conditions. For ExDark, images were attenuated to 25% intensity with Gaussian noise (σ² = 0.01) to simulate the readout of conventional linear sensors under low-light scenarios.

To incorporate the device’s nonlinear enhancement, its empirical photoresponse curve was extracted and mapped into a Look-Up Table (LUT) using piecewise cubic Hermite interpolation. The attenuated images were then processed pixel-by-pixel through this hardware-derived LUT. Finally, the enhanced datasets were evaluated using a ResNet-18 convolutional neural network (optimized by SGDM, batch size = 128, learning rate = 0.001 for 20 epochs) without modifying the original labels.

Supplementary information

Supplementary Information ^{(9.3MB, pdf)}

41467_2026_76364_MOESM2_ESM.pdf ^{(77.8KB, pdf)}

Description of Additional Supplementary Files

Supplementary Movie 1 ^{(26.2MB, mp4)}

Supplementary Software 1 ^{(380.2KB, zip)}

Transparent Peer Review file ^{(3.1MB, pdf)}

Source data

Source Data ^{(4.2MB, zip)}

Author contributions

P.L. conceived the experiments. Y.Z., J.Y., and H.Z. fabricated the devices and performed the photoelectric measurements. J.Y. performed the Raman measurement. Y.Z. performed the CNN simulation. Y.Z., P.L., W.L., G.T., Z.H., and Z.Y. analyzed the results. Y.Z., P.L., G.H., W.L., G.T., and Z.Y. co-wrote the manuscript. All authors discussed the results and commented on the manuscript.

Peer review

Peer review information

Nature Communications thanks Jae-Young Choi, Zhengxun Lai and the other, anonymous, reviewer(s) for their contribution to the peer review of this work. A peer review file is available.

Funding

This work was supported by the National Natural Science Foundation of China (Grant No. 52322512 to P.L., Grant No. U2441275 and T2425018 to G.T.). G.T. thanks the support from New Cornerstone Science Foundation through the XPLORER PRIZE.

Data availability

The Source Data underlying the figures of this study are available with the paper. All raw data generated during the current study are available from the corresponding authors upon request. Source data are provided with this paper.

Code availability

The code and algorithms used for the dual-mode picture enhancement simulation (Fig. 5d), the CIFAR-10 dataset evaluation (Fig. 5e), and the ExDark dataset evaluation (Supplementary Fig. 28) are provided as supplementary files (Supplementary Software 1.zip). Any additional code is available from the corresponding author upon request.

Competing interests

The authors declare no competing interests.

Supplementary information

The online version contains supplementary material available at https://doi.org/10.1038/s41467-026-76364-z.