Extreme Illuminated Vision Processing with a Graded Alloyed Perovskite In-sensor Computing Network

Abstract

In-sensor computing is proposed to reduce energy expenditure and processing latency by unifying sensing and computation within the hardware layer, yet the application under extreme illuminating scenario remains constrained by simultaneously obtaining broadband responsivity, large linear dynamic range and fast response. Here, we report a fully vapor-deposited graded Pb-Sn alloyed perovskite heterojunction photodiode with improved crystal quality. It enables the detection of light from visible to infrared light with a 230 dB linear dynamic range and 33 ns response time. We also develop a wafer-scale imaging processor by integrating the photodiode to a reconfigurable array. With this approach, we demonstrate biomedical detection and spatiotemporal trajectory encoding. The in-sensor processor realizes low-power high-resolution visible to infrared wavelength edge detection, adaptive background suppression under dim light and noise-immune high-speed dynamic imaging. Our results extend the options for in-sensor computing hardware, and thus pave a way toward practical artificial intelligent machine vision.

Zhan et al. report a fully vapor-deposited graded perovskite photodiode for visible-to-IR imaging with 230 dB linear dynamic range and 33 ns response time. Wafer-scale reconfigurable imaging processor is developed for edge detection, adaptive background suppression under dim light, and noise-immune high-speed dynamic imaging.

Introduction

Traditional machine vision systems isolate sensing and processing units, incurring substantial energy consumption and latency from data transfer between image sensors and backend processors. This bottleneck hinders real-time applications in edge computing and artificial intelligence^1,2. In-sensor or near-sensor computing addresses this by enabling pixel-level preprocessing before data transmission, necessitating in-situ reconfigurable photoresponsivity^3,4. Recent advances leverage novel device architectures, including electrical doping⁵, ion migration⁶, ferroelectric effect⁷, floating-gate mechanism⁸, to dynamically reconfigure photoresponsivity. These approaches have demonstrated functionalities critical to machine vision, such as object recognition⁹, motion perception^5,10, high-speed data processing¹¹, spectral extraction¹², and in-sensor data compression¹³.

Practical application of in-sensor computing technology in machine vision, particularly for extreme illuminated conditions, however, remains a significant challenge, requiring the operation of a photodiode with high speed, large dynamic range and extended wavelength^12,14,15. Firstly, imaging under dim light or noisy environments is demanded yet difficult for a large number of important application contexts. Dealing with both weak and strong illuminated regions in the same image calls for sensors with a large linear dynamic range (LDR). Unfortunately, the LDR in most sensors is limited to no more than 150 dB, although a few works reported higher LDR by sacrificing the imaging speed^16,17. Secondly, a neuromorphic visual system, such as for autonomous vehicles and robotics, needs to capture dynamic images. As a fundamental prerequisite, a photodiode must have an operation speed down to the nanosecond level. However, response time is always compromised by power efficiency or detection sensitivity. And finally, it is also necessary to expand the response to ultraviolet or infrared light to capture information unavailable in the visible spectrum. For example, biomedical imaging is often operated at the infrared range.

Organic–inorganic hybrid perovskites, with high absorption coefficients and long carrier diffusion lengths, show great promise for optoelectronics. Recent advances have demonstrated the fabrication of a thin film photodetector with better performance than crystalline Si^18–21. Tin alloying narrows the bandgap and extends spectral responsivity into the near-infrared (NIR) while reducing lead-based perovskite toxicity^22,23. However, oxidation of Sn²⁺ to Sn⁴⁺ induces self-p-doping and deep traps, elevating dark current and noise as well as undermining stability²⁴. Even though sophisticated 2D/3D heterojunctions can partially suppress Sn²⁺ oxidation, the intrinsically low conductivity of 2D perovskites (~10⁻⁵ S cm⁻¹) limits the photoresponse efficiency^25,26. Accordingly, realizing scalable, high-quality, monolithically integratable perovskite units for practical in-sensor computing has remained an open challenge.

Here, we introduce a gradient composition engineering strategy that provides a protective layer for oxidation suppression of Sn²⁺. By this approach, we fabricate stable lead–tin (Pb–Sn) alloyed perovskite films via a fully vapor-deposition process. And based on these advancements, we demonstrate a scalable broadband bipolar programmable perovskite unit (BPPU). The core device comprises two sets of back-to-back cascaded perovskite photodiodes. This architecture enables rapid, stable, and precise control of photoresponse purely through circuit‑level biasing, avoiding trade‑offs in hysteresis, speed, or reliability. The uniform vapor-phase deposition ensures symmetric bipolar response tuning under low bias (±200 mV), fast response time of 33.0/37.4 ns, and a large LDR of 230 dB. By integrating the BPPU into wafer-scale arrays, we achieve diverse in-sensor computing functions such as adaptive background suppression, biomedical sensing, and spatiotemporal trajectory encoding.

Results

Design of Pb–Sn perovskite graded heterojunction photodiode

A fully vapor-deposition process is used for sequential layer-by-layer stacking of functional layers, which meets the requirements for future scalable, CMOS-compatible, and high-precision fabrication. The resulting device architecture is a vertical stack of FTO/Spiro-TTB/perovskite/C₆₀/BCP/Ag (Fig. 1a). A high-quality Pb-Sn alloyed perovskite graded heterojunction is fabricated using a synergistic strategy combining compositional engineering and self-encapsulation, as shown in Fig. 1b: First, co-evaporate SnI₂ and PbI₂ to form a homogeneous (SnI₂:PbI₂) mixing layer. Second, keep depositing PbI₂ while shutting off the SnI₂ source to deposit a thin pure PbI₂ capping layer, forming a (SnI₂:PbI₂)/PbI₂ heterojunction. The PbI₂ capping layer acts as an encapsulation layer for the underlying SnI₂. Finally, reacting the (SnI₂:PbI₂)/PbI₂ heterojunction with vaporized methylamine iodide (MAI) at 180 °C converts it into MAPb_xSn_1−xI₃/MAPbI₃ perovskite thin film. Subsequent annealing activates thermal diffusion, promotes crystallization, and establishes a Pb–Sn ratio gradient across the film.

Fig. 1. Self-encapsulated Pb–Sn alloyed perovskite graded heterojunction photodiode.

a Schematic diagram of the device structure with fully evaporated layer stacking. b Schematic diagram of the dual-source co-evaporation process for the graded Pb–Sn alloyed perovskite thin films. It highlights the gradually changed composition ratio perpendicular to the basal plane. c, d SEM micrographs of the perovskite films with the c homogeneous structure and d graded heterojunction structure ($\mathrm{\Delta }d$ = 80 nm). Scale bar: 500 nm. e, f PL intensity stability test under continuous 532 nm laser excitation (200 μW) for the perovskite sample with e homogeneous structure and f graded heterojunction structure ($\mathrm{\Delta }d$ = 80 nm). g Schematic energy diagram of the photodiode with gradient structure. HTL hole transport layer, ETL electron transport layer. h EQE of the homogeneous and the graded heterojunction perovskite photodiodes.

The effectiveness of the component engineering by the top PbI₂ layer is supported by X-ray diffraction (XRD), optical absorption, and emission shifting in Supplementary Fig. 1 and Supplementary Fig. 2. Scanning electron microscopy (SEM) analysis verifies that PbI₂ capping profoundly optimizes perovskite thin film morphology (Supplementary Fig. 3). While homogeneous structure perovskite films show a porous morphology with pinhole defects (Fig. 1c), introducing an 80 nm PbI₂ capping layer on Pb–Sn alloyed precursor films yields dense perovskite grains with reduced inter crystalline voids (Fig. 1d). This microstructural densification effectively suppresses leakage currents from 8.5 × 10⁻⁶ A to 5.6 × 10⁻⁸ A at −0.5 V bias for the device (Supplementary Fig. 4).

Effective prevention of Sn²⁺ oxidation in the graded structure is supported by the time evolution of photoluminescence (PL) (Fig. 1e, f, Supplementary Fig. 5). In homogeneous perovskite structure, the primary PL peak at 923 nm (Pb–Sn alloyed perovskite phase) suffers 50% intensity loss within 200 s (79% within 750 s), accompanied by an anomalous 702 nm emission peak emerging after 200 s, a signature of photoinduced Sn²⁺ oxidation forming MA₂SnI₆ phases^27,28. In stark contrast, the sample pre-covered 80 nm PbI₂ maintains >91% PL intensity of Pb–Sn alloyed perovskite over 750 s without secondary peak formation, demonstrating exceptional photostability. The effective encapsulation and protection of Sn²⁺ oxidation were further confirmed through femtosecond laser ablation XPS depth profiling of the aged samples (Supplementary Fig. 6). In device characterization, devices based on the pre-covered 80 nm PbI₂ layer show improved operational stability compared to the homogeneous control sample, with on-off switching characteristics remaining virtually unchanged after long-term continuous operation exceeding 700 h (Supplementary Fig. 7) and demonstrating remarkable photostability even under strong light irradiation (Supplementary Fig. 8).

To verify the formation of compositional and band-structure gradient in the perovskite film, in-depth X-ray photoelectron spectroscopy (XPS) analysis is first employed, revealing that the Sn:Pb ratio evolves gradually from 0.05:0.95 at the surface to 0.4:0.6 in the bulk (Supplementary Fig. 9). Ultraviolet photoelectron spectroscopy (UPS) measurement in Supplementary Fig. 10 is then employed to reveal the band structure, giving the detailed band alignment in Fig. 1g. It shows a type-II heterojunction between the Pb–Sn alloy region with band gap around 1.33 eV and Pb-rich surface layer with band gap around 1.50 eV. This continuous gradient results in a progressive spatial tilting of both the conduction band minimum (CBM) and valence band maximum (VBM) across the entire active layer, effectively generating a built-in electric field. The Mott-Schottky analysis in Supplementary Fig. 11 further reveals an enhancement of built-in potential from 0.58 V in the homogeneous film to 0.66 V in the gradient film. This built-in field actively drives photo-generated electrons towards the lead-rich region (orange arrow) and holes towards the tin-rich region (blue arrow) via directional drift, thereby transitioning carrier transport from diffusion-dominated to efficient drift-dominated. As a result, acceleration of charge separation and transport, substantially shortening carrier transit and residence time within the active layer, can be expected. The observation of faster rise and fall times in the photoresponse of the graded heterojunction (Supplementary Fig. 12) provides evidence for the improved carrier transport kinetics.

The thickness of the pre-covered PbI₂ layer ($\mathrm{\Delta }d$ = 20–110 nm) is a critical parameter for the compromise of device operational stability, carrier transport efficiency, and NIR sensitivity. The 80 nm PbI₂ pre-covered layer provides an optimized NIR response. When the PbI₂ thickness exceeds 80 nm, the Pb–Sn balanced region, which serves as the primary infrared absorber, is too short to preserve photoresponse beyond 850 nm. This attenuation is directly evidenced by a 33% responsivity decrease at 850 nm in $\mathrm{\Delta }d$ = 110 nm devices (0.28 A W⁻¹ versus 0.42 A W⁻¹ for 80 nm devices, Supplementary Fig. 13). Conversely, sub-80-nm pre-covered layers fail to adequately suppress oxidation-induced degradation, leading to deteriorated response. With the optimized structure, the gradient Pb–Sn alloyed perovskite film achieves a zero-bias external quantum efficiency (EQE) of 61.07% at 850 nm (Fig. 1h). This wavelength is crucial for NIR applications such as vein imaging that rely on the characteristic absorption of hemoglobin in the 760–850 nm spectral range. Although the PbI₂ coating weakens the device’s short-wave infrared response above 850 nm, the high sensitivity maintained within the key spectral window fully meets the physical requirements of the targeted biomedical applications. This structure simultaneously satisfies essential intelligent vision system requirements, including multi-wavelength adaptability, environmental stability, and rapid response capabilities.

Bipolar photoresponse perovskite in-sensor computing unit

In a machine vision system, the convolution of input optical signals and weights relies on programmable photoresponse characteristics. As shown in Fig. 2a, the designed BPPU integrates two vertically stacked photodiodes comprising a shared functional layer but adjacent electrically-isolated Ag electrodes. Detailed photocurrent-voltage characteristic of BPPU is shown in Supplementary Fig. 14. Figure 2b shows the typical bias-dependent bipolar photoresponse that is understood by the equivalent circuit in the inset. Under uniform illumination intensity P, the photocurrent of BPPU is described by $I_{\mathrm{total}}=I_{\mathrm{D}1}+I_{\mathrm{D}2}$. This relationship is verified by the photoresponse characteristic of an individual photodiode by separately illuminating each diode. It is found that the photodiode D1 contributed to the positive photoresponse and D2 contributed to the negative photoresponse. Figure 2c maps the reconfigurable photoresponse as a function of incident power across a 400 mV window (−200 to +200 mV). The net responsivity can be continuously tuned from negative to positive. This enables optoelectronic analog multiplication and accumulation (MAC) for in-sensor computing applications (see below). Notably, the low operating voltage indicates the device is operated with ultra-low power consumption.

Fig. 2. Design and performance of BPPU.

a Schematic structure of a single BPPU. b The photocurrent-bias voltage curves and responsivity-bias voltage curves of illumination on photodiode D1, photodiode D2, and photodiodes D1 with D2 simultaneously, respectively. An 860 nm laser with 5.7 mW cm⁻² is used as the light source. Inset: The BPPU equivalent circuit. c Reconfigurable photoresponse mapping. Two-dimensional current matrix as a function of incident power (0.01–1 mW cm⁻²) and bias voltage (−200 to +200 mV). d The noise current spectra of the BPPU at 0 V bias. e The photocurrent response for each photodiode in BPPU with respect to power intensity at 860 nm under 0 V. The red line is the linear fitting of the LDR. These LDR measurements were conducted by illuminating a sub-diode within the BPPU to assess its intrinsic photoresponse performance. f The bipolar switching photoresponse time of the BPPU, measured at 0 V bias under 660 nm laser illumination with a light intensity of 1 mW cm⁻² and a modulation frequency of 1 MHz. g Background (BG) suppression capability of the extended BPPU under $V_{\mathrm{A}}$ = $V_{\mathrm{B}}$ = 0 V, measured under background illumination from an 860 nm LED array. Error bars are the standard deviation. h Response to a moving 860 nm laser spot with intensities of 5.62 nW, 1.01 nW, and 146 pW under dark background illumination. i Response to the same moving laser spot (1 mm spot size) under strong background illumination provided by an 860 nm LED array (42.18 mW), with spot intensities of 1.19 μW, 719 nW, and 530 nW. The colored shading represents the spatial responsivity profile across the extended BPPU array.

Building upon this back-to-back cascading configuration, leakage currents are self-canceled as shown in Supplementary Fig. 15, suppressing the noise current to 7.44 × 10⁻¹⁴ A Hz^−0.5 under 1 Hz as shown in Fig. 2d. The photodiode exhibits exceptional 230 dB LDR from 4.75 pW cm⁻² to 1.50 W cm⁻² under 860 nm in the BPPU at zero bias (Fig. 2e, Supplementary Fig. 16). Our device demonstrates competitive advantages over recently reported broadband (visible–NIR) photodetectors, as detailed in Supplementary Table 1. The BPPU achieves an ultrafast bipolar photoresponse speed of 33.0 ns rise time and 37.4 ns fall time (Fig. 2f). It results from the cascaded architecture. The back-to-back connected diodes’ equivalent capacitance ($C_{\mathrm{eq}}=\frac{C_1{C}_2}{C_1+C_2}$) reduces by 50% compared to single-photodiode configurations. The comparative data in Supplementary Fig. 17 provide direct experimental validation for this capacitance reduction effect.

This design establishes a new architecture for in-sensor computing devices, simultaneously addressing the longstanding trade-off between detection speed, sensitivity, and power consumption in conventional photonic systems. Remarkably, our device accomplishes adjustable photoresponse with an unprecedentedly low programming voltage of ±0.2 V—a figure significantly lower than state-of-the-art bipolar programmable photodetectors (Supplementary Table 2). The 230 dB dynamic range allows unified processing of vision with both low-light and high-flux parts without any gain switching/adapting. It overcomes the critical bottleneck of saturation-induced feature loss in existing neuromorphic vision sensors.

In-sensor computing core

Under zero bias, the BPPU readily functions as a computing kernel for hardware-level background (BG) suppression, as demonstrated in Fig. 2g. The photocurrent fluctuations are constrained within <12.5 nA despite global BG illumination increasing from dark to 19.1 mW cm⁻². This stability arises from responsivity cancellation ($R_1=-R_2$) between neighboring diodes (D1 and D2 in Fig. 2b). It creates a hardware differential connection that effectively subtracts the common-mode BG. But the total current is still able to sensitively respond to locally illuminated signals. For example, a moving light spot induces a negative (positive) signal when it goes through D1 (D2) zones. Our measurements suggest that under dark BG, the array can resolve a motion signal as dim as 146 pW (Fig. 2h). Even under an intense BG illumination of 42.18 mW, a target signal down to 530 nW can still be successfully extracted (Fig. 2i), which corresponds to an interference level 98 dB stronger than the target signal. This robust performance is corroborated by the dynamic range measurements under strong background illumination (Supplementary Fig. 18). The BPPU achieves sub-microsecond fast BG suppression by directly performing analog photocurrent subtraction within the sensor hardware. This ability is not only promising for high-contrast scenes but also bypasses conventional analog-to-digital conversion and digital signal processing. It definitely saves power and prevents precision loss. It is important to note that the excellent background suppression of the BPPU relies on uniform illumination across its two sub-pixels. In practical scenarios, spatial gradients or local shadows in background illumination may reduce the suppression efficiency and introduce low-frequency spatial artifacts. These effects can be mitigated by subsequent signal processing or system-level calibration strategies.

To further demonstrate the in-sensor computing capabilities, we also designed an extended BPPU array by integrating two fundamental units through a shared Ag current output terminal, as illustrated in Fig. 3a. Figure 3b shows the equivalent circuit of this enhanced design. It provides independent bias terminals ($V_{\mathrm{A}}$, $V_{\mathrm{B}}$) to program the responsivity of two branches and sum the photocurrent at the output terminal. Therefore, it is easy to understand that each extended BPPU array serves as a minimum in-pixel MAC core for the input light intensity and weights defined by the bias voltage:

$$I_{\mathrm{output}}=\left[R\left(V_{\mathrm{A}}\right),R\left(V_{\mathrm{B}}\right)\right]\left[\begin{array}{c}\hfill P_{\mathrm{A}}\hfill \\ \hfill P_{\mathrm{B}}\hfill \end{array}\right]$$ 1

Fig. 3. Configuration of the extended BPPU array.

a Extended BPPU structure of functional extension. b Equivalent circuit diagram of the extended BPPU. c Optical images of a 4-in. wafer containing 404 photodiodes (101 BPPUs). d The PL mapping of the emission peak position of perovskite film at 17 substrate locations on the 4-in. wafer. Scale bar: 1 cm. e Magnified views of a 6 × 6 array. Scale bar: 500 mm. f Responsivity mapping of the 6 × 6 extended BPPU array. Left panel: initial responsivity matrix; right panel: functionalized responsivity matrix after separate voltage programming of each BPPU.

The MAC core can be further scaled up to form a convolution kernel. We demonstrated a chip-level implementation of 2 × 2 and 6 × 6 extended BPPU arrays on a 4-in. wafer (Fig. 3c). Quantitative verification of film uniformity was achieved by measuring PL peak positions across 17 positions (Fig. 3d, Supplementary Fig. 19), revealing a standard deviation of only 2.21 nm. To further confirm the reliability of our deposition technique, the uniformity and reproducibility of the Sn/Pb gradient structure were assessed via XPS profiling, which shows high consistency over batches (Supplementary Fig. 20). Such exceptional homogeneity ensures high device-to-device consistency (Supplementary Fig. 21). The arrays can be programmed as arbitrary matrices for convolution operation. Figure 3e, f gives an instance. Based on the aforementioned reconfigurability, extensibility, and scalability, we next demonstrate the application of BPPU in intelligent vision tasks, from biomedical imaging/processing to real-time motion tracking.

In-sensor imaging processing for biomedical applications

BPPU-based biological feature extraction was demonstrated using a home-built galvanometer-based scanner (Supplementary Fig. 22) for acquisition of both surface topography and subsurface vascular structures. For fingerprint biometrics, high-precision epidermal topography is captured under −0.2 V bias. It resolves ridge-valley structures (200–400 μm spacing) and sweat pore distributions (40–100 μm diameter, Fig. 4a). Switching to 0 V bias enables horizontal edge detection, where only the ridge boundaries are resolved as distinct intensity transitions (Fig. 4b) due to the differential output eliminating background signals. As shown in Fig. 4c, this operation manages to suppress nonlinear background variations and enables direct hardware-level extraction of Level 2 biometric features (bifurcations, ridge endings) and Level 3 identifiers (sweat pores)^29,30. Notably, the resulting discretized edge map shows a 101% improvement in edge-specific signal-to-noise ratio, alongside 95.5% agreement with conventional digital processing outputs and higher contrast than conventional Sobel or Laplacian digital filters (Supplementary Fig. 23).

Fig. 4. In-sensor imaging and processing.

a Conventional imaging mode at −0.2 V with the responsivity matrix of $\left[\begin{array}{cc}\hfill -1\hfill & \hfill 0\hfill \end{array}\right]$. The fingerprint is illuminated by a single white light LED. b Edge detection mode at 0 V with a differential responsivity of $\left[\begin{array}{cc}\hfill -1\hfill & \hfill +1\hfill \end{array}\right]$. c Cross-sectional intensity profiles. d Photograph of a hand with an inset showing the imaging part in (e, f). e, f Vein imaging via NIR: e conventional imaging, where the veins in shadow are fully blurred by the bright part; f edge detection, where the veins are clear in the same condition as (e). The dashed lines outline the localized subcutaneous veins. g Schematic of the PWV measurement method based on BPPU differential signals. h Differential time-domain signals between two sub-photodiodes in a BPPU. i Normalized PWV and heart rate during post-exercise recovery. Shaded areas are SD.

We also extended this in-sensor edge processing paradigm to deeper tissue diagnostics through NIR absorption contrast. Hemoglobin in veins exhibits stronger absorption of NIR illumination than the surrounding tissues, such as skin and subcutaneous fat. NIR is thus very suitable for vein imaging. However, post-image processing techniques such as contrast-limited adaptive histogram equalization and multi-scale filtering are required to address the weak contrast between the vein and surrounding tissues^31–33. As shown in Fig. 4d, a photograph of a hand is provided with an inset highlighting the specific region used for imaging (Fig. 4e, f). A transmission image acquired using conventional imaging mode shows that a high-intensity signal is detected between the fingers, while minimal venous detail is discernible (Fig. 4e), demonstrating the significant dynamic range challenge in conventional subcutaneous imaging. In contrast, under edge detection mode (Fig. 4f), the vein edges are clearly resolved due to abrupt transitions in hemoglobin concentration at vessel boundaries. This approach significantly enhances vein-to-tissue contrast, enabling clear visualization of vessels through differential absorption.

The device’s advantages for achieving this high-quality image are twofold. The intrinsically high LDR together with the differential architecture accommodates high-intensity background illumination from the gap between fingers with tissue scattering and weak additive absorption signals from hemoglobin. Conventional photodetectors would otherwise saturate under high surface background intensity or lose vein-specific absorption variations in quantization noise. Consequently, the BPPU enables direct extraction of venous boundaries from raw NIR matrices without dynamic range compression. Notably, in contrast to conventional imaging with separate post-computation, our in-sensor computing approach operates at 0 V, approaching nearly zero-power data processing. Furthermore, the bipolar response characteristics of the extended BPPU array, in principle, allow it to be programmed to perform arbitrary two-dimensional linear convolution operations, effectively functioning as a versatile image processing kernel for various in-sensor computational imaging effects.

Beyond static in-sensor imaging applications, the BPPU architecture also enables dynamic physiological monitoring through time-resolved differential sensing. Pulse wave velocity (PWV) is a critical parameter for assessing arterial stiffness, which has been linked to various cardiovascular diseases³⁴. Traditional PWV ($v$) measurement requires multiple sensors and complex signal processing to analyze the time delay between pulse waves at different measurement sites^35–37. However, this out-of-sensor and multi-sensor data processing can be simplified by utilizing BPPU in-sensor computing. Here, we verify it through the local measurement of the fingers.

Figure 4g schematically shows the in-sensor processing of the photoplethysmographic (PPG) signal when NIR light passes through tissues. The spatial separation ($D$ = 2.5 mm) between the two sub-photodiodes introduces a small time lag ($\mathrm{\Delta }T=\frac{D}{v}$, typically <0.5 ms for $v$ > 5 m s⁻¹) in the pulse wave arriving times. This results in phase-shifted PPG signals ($P\left(t\right)$ and $P\left(t-\mathrm{\Delta }T\right)$) with distinct temporal characteristics. The BPPU computes the differential-delay current as $I_{\mathrm{diff}}\left(t\right)=I_1\left(t\right)-I_2\left(t\right)=\alpha \left[P\left(t\right)-P\left(t-\mathrm{\Delta }T\right)\right]$, where $\alpha$ is the optoelectrical-related coefficient. Since the $\mathrm{\Delta }T$ is significantly shorter than the PPG pulse wave’s characteristic timescale, the peak differential current at the maximum pulse slope (${\left(\frac{\mathrm{d}P}{\mathrm{d}t}\right)}_{\max }$) is approximated by first-order Taylor expansion:

$$I_{\mathrm{diff},\max }\approx \alpha {\left(\frac{\mathrm{d}P}{\mathrm{d}t}\right)}_{\max }\cdot \mathrm{\Delta }T=\alpha {\left(\frac{\mathrm{d}P}{\mathrm{d}t}\right)}_{\max }\cdot \frac{D}{v}$$ 2

This yields the fundamental relationship between peak magnitude ($I_{\mathrm{diff},\max }$) and PWV:

$$v=\frac{k}{I_{\mathrm{diff},\max }}$$ 3

where the calibration constant $k=\alpha \cdot D{\left(\frac{\mathrm{d}P}{\mathrm{d}t}\right)}_{\max }$ incorporates optoelectronic and individual hemodynamic parameters. As experimentally confirmed in Fig. 4h, this inverse proportionality enables the direct translation of differential current peaks into PWV values. Figure 4i demonstrates synchronized decreasing trends in both BPPU-derived normalized PWV and heart rate during post-exercise recovery. Physiological validation of our measurements is established through covariation with heart rate, which exhibits a known correlation with PWV dynamics. Detailed signal processing is provided in Supplementary Fig. 24. Our approach thus enables efficient PWV monitoring through simple differential current measurement, offering a practical alternative to conventional multi-sensor systems.

Robust motion trajectory recognition

Finally, we use a dynamic spatiotemporal pattern recognition task to explore the potential of BPPU for intelligent vision. The task involves identifying the motion trajectory mode (32 kinds of distinct trajectories shown in Fig. 5a) of a specific target. The BPPU was scaled up to a 6 × 6 array. The trajectory-specific time-domain signals are generated via a convolution operation: $I_{\mathrm{output}}\left(t\right)={\sum }_{i=1}^N{R}_i{\varphi }_i\left(t\right)$, where $R_i$ denotes the programmable responsivity of the $i$ th pixel (Fig. 3f, right), ${\varphi }_i$ represents the time-dependent incident optical signal, and $N$ = 36 corresponds to the 6 × 6 array configuration. It enables the extended BPPU array to directly encode the moving objects’ spatiotemporal signatures into discriminative neuromorphic information. Or, in other words, the BPPU sensor array naturally supports artificial neural network vision processing as it plays the role of the first layer of a neural network. The comprehensive system and detailed recognition process are established in Supplementary Fig. 25 and “Methods”, and schematically illustrated in Fig. 5b. The original sensor output current signals for each BPPU in time-domain are shown in Fig. 5c and Supplementary Fig. 26. They were directly encoded into machine-readable features and sequentially fed into a long short-term memory (LSTM) network for robust classification^38,39. This workflow eliminates redundancy in conventional sensory systems (including data converters, digital edge detection, digital filtering, and optical flow calculation). Within the motion trajectory simulation platform measurements, the system achieves 100% classification accuracy across training, validation, and test sets (Fig. 5d) under noise-free conditions (Supplementary Fig. 27).

Fig. 5. Robust motion trajectory recognition.

a Thirty-two kinds of trajectories for validation, including noise-free paths (horizontal, vertical, diagonal, and corresponding reverse motions), and with-noise variants (4–20% random walk noise, 4–20% gaussian noise). b Schematic of the BPPU trajectory recognition system. c Distinctive time-domain signals for all noise-free trajectories. d Confusion matrix for noise-free trajectories on the test set. e LSTM network classification accuracy for with-noise trajectories under increasing noise levels. f Confusion matrix for with-noise trajectories on the test set.

Robustness against noise is another main challenge for machine vision. We validate the system rigorously by examining its performance under perturbations. Specifically, we purposely introduced two distinct types of noise into the real trajectories: Gaussian noise and random walk noise. These were selected to simulate uncorrelated and temporally correlated interference common in real-world scenarios, as shown in Supplementary Fig. 28.

Briefly, the noises are coupled to the motion target by:

$$\left\{\begin{array}{c}\hfill x\left(t_i\right)=x_0\left(t_i\right)+\alpha \cdot {\mathcal{N}}_x\left(0,{\sigma }^2\right)\hfill \\ \hfill y\left(t_i\right)=y_0\left(t_i\right)+\alpha \cdot {\mathcal{N}}_y\left(0,{\sigma }^2\right)\hfill \end{array}\right)$$ 4

for gaussian noise, and

$$\left\{\begin{array}{c}\hfill x\left(t_i\right)=x_0\left(t_i\right)+\frac{\alpha }{10}{\sum }_{k=1}^i{\mathcal{N}}_x^{\left(k\right)}\left(0,{\sigma }^2\right)\hfill \\ \hfill y\left(t_i\right)=y_0\left(t_i\right)+\frac{\alpha }{10}{\sum }_{k=1}^i{\mathcal{N}}_y^{\left(k\right)}\left(0,{\sigma }^2\right)\hfill \end{array}\right)$$ 5

for random walk noise, where $x_0\left(t_i\right)$ and $y_0\left(t_i\right)$ denote noise-free positions, $\alpha$ is the noise level ($\alpha$ = 4%, 8%, 12%, 16%, 20%), $\mathcal{N}\left(0,{\sigma }^2\right)$ represents independent Gaussian variables with 0 mean and ${\sigma }^2\left(\sigma =1\right)$ variance, and $k$ indexes discrete time steps.

Figure 5e summarizes the identification accuracy for signals subjected to a single kind of noise. The Gaussian noise in our experiments models common uncorrelated, high-frequency disturbances in real-world scenarios, such as electronic readout noise in sensors and circuits, as well as photon shot noise, which independently affects each measurement point. In this case, the system demonstrates exceptional resilience: test accuracy maintains 95.32% at 20% noise intensity, representing a marginal 4.68% degradation from baseline. This robustness stems from the uncorrelated nature of Gaussian noise. The BPPU-based system can effectively filter out this kind of transient perturbation. The random walk noise is selected to simulate cumulative, low-frequency interference, including mechanical drift or vibration and slow sensor baseline drift. Its accumulated effect introduces time-correlated deviations that distort the intended motion trajectory, thereby testing the system’s robustness against shape errors. These disturbances alter key shape features of the path and increase classification complexity. Nevertheless, the system maintains a test accuracy registers at 99.19% for 4% noise intensity and remains measurable at 81.71% under 20% intensity. This performance demonstrates robust tolerance to trajectory distortions. And finally, we also performed a comprehensive evaluation by integrating training across all noise types and intensities, which yields a 90.87% classification accuracy (Fig. 5f and Supplementary Fig. 29). These results confirm the system’s robust discrimination capability, even for perceptually similar motion patterns under challenging conditions. Fundamentally limited by a 33‑ns temporal resolution, the BPPU-based system can directly encode fast motion. The corresponding theoretical maximum detectable velocity of 7.6 × 10⁴m s⁻¹ (discussed in Supplementary Fig. 30), allowing the system to capture and classify rapid trajectories that would otherwise be blurred or missed by standard imaging systems.

Discussion

We have developed perovskite optoelectronic devices and arrays with both high performance and versatile scalability through graded vapor-phase deposition. The developed BPPU is broadband responsive, has a large LDR, and has a fast response. We also used the BPPU to implement a programmable image filter kernel for biomedical imaging and an optoelectronic convolution processor for real-time machine vision. These properties facilitate advanced in-sensor computing and processing for future artificial intelligence technology.

Methods

Materials

FTO glasses, Lead (II) iodide (PbI₂), 2,2’7,7’tetra(N,N-di-tolyl)amino-9,9-spiro-bifluorene (Spiro-TTB), and C₆₀ were purchased from Advanced Election Technology CO., Ltd. Tin (II) iodide SnI₂ (99.999%) was purchased from 3AMaterials. Methylammonium iodide (CH₃NH₃I, MAI) (99.8%) was purchased from TCI. 2,9-dimethyl-4,7-diphenyl-1,10-phenanthroline (BCP) was purchased from Xi’an Polymer Light Technology Corp. All of the chemicals were used as received.

Perovskite photodiode and BPPU fabrication

The device structure was configured with a vertically stacked structure. Prepatterned FTO substrates were cleaned, treated with oxygen plasma, and then deposited with a 6 nm Spiro-TTB layer via thermal evaporation (~0.1 Å s⁻¹). The graded perovskite heterojunction was fabricated through a sequential vapor-deposition process wherein SnI₂ and PbI₂ were first co-evaporated at a 1:1 molar ratio to form a homogeneous underlying layer with a total fixed thickness of 180 nm. The SnI₂ source was then terminated at a stage corresponding to a PbI₂ capping thickness (varied from 0 to 110 nm), depositing a pure PbI₂ overlayer. Finally, the stacked precursor film was converted into a compositionally graded perovskite structure via reaction with vaporized MAI in a 180 °C vacuum oven. After cooling, the perovskite film was IPA-spin cleaned and annealed (100 °C, 10 min). Subsequent layers included thermally evaporated C₆₀ (23 nm, ~0.1 Å s⁻¹) and BCP (8 nm, ~0.08 Å s⁻¹). The Ag electrode was finally deposited by thermal evaporation through a patterned shadow mask. For optoelectronic characterization, imaging, and trajectory recognition, the sub-photodiode areas of the BPPUs are 0.04 cm², 4 × 10⁻⁴ cm², and 0.01 cm², respectively.

Film characterization

Ultraviolet-visible absorption spectra were measured using a SHIMADZU UV-2600. The PL spectrum was measured using a confocal Raman microscope (WITec, Alpha300R) with a 532 nm laser as the excitation source. SEM images were acquired through a FEI Quanta 450 FEG field emission scanning electron microscope. XRD spectra were acquired using a BRUKER D8-Advance equipped with Cu Kα1 irradiation (λ = 1.5406 Å). UPS characterizations were performed using a Thermo Fisher Scientific K-Alpha⁺ equipped with a He I (21.22 eV) excitation source. XPS for Pb–Sn ratio characterizations was acquired on the same instrument (Al Kα radiation source). Femtosecond laser ablation (fs-LA) XPS depth profiling for Sn²⁺ oxidation analysis was performed using a Thermo Fisher Scientific Hypulse XPS system^40,41. Single femtosecond laser pulses of 40 μJ per level were employed, delivered as a horizontally linearly polarized Gaussian beam with a wavelength of 1030 nm and a pulse duration of 160 fs.

Device characterization

The electrical characterization of the photodiode and BPPU devices was conducted in a dark metal box to avoid interference from ambient light. Current–voltage characteristics were measured using a source-measure unit (Keithley 2612A Source Meters).

The LDR measurement was referred to the methodology outlined in the literature^42,43. The device was illuminated by a modulated laser, and the resulting photocurrent was detected with a lock‑in amplifier (Guangzhou Sine Scientific Instrument Co., Ltd., Model OE2031) under zero bias. The frequency spectral response was obtained by scanning the reference frequency and standardized according to the equivalent noise bandwidth determined by the lock-in amplifier’s filter order and time constant. During the experiment, the incident optical power was calibrated with an optical power meter (Daheng Optics, GCI-080302) and attenuated using a series of neutral density filters (Thorlabs).

Noise current measurements were conducted at room temperature under dark conditions, also utilizing the lock-in amplifier. EQE spectra were acquired at zero bias using an Enlitech QE-C system. The response time was acquired by an oscilloscope (Tektronix, TDS2024C) through a trans-impedance amplifier.

The BPPU array was tested using a home-built multi-channel scanning system. A Keithley 3706A Switch/Multimeter with a plug-in card (Keithley 3730 matrix card) was used to select rows and columns to determine the read-out terminal to be tested.

Imaging processing demonstration

Our home-built galvanometer scanning imager utilizes a dual-axis galvanometer driven by XY-voltage inputs for rapid, predefined trajectory scanning. The system directs reflected or transmitted scene light through a camera lens onto the photodetector. Real-time voltage programming dynamically reconfigures BPPU’s image processing functions, with processed signals acquired via TIA and lock-in amplification for subsequent image reconstruction.

Motion trajectory experimental platform

An 850 nm laser serves as the light source, with its beam dynamically steered across the BPPU device following a preset trajectory via a dual-axis galvanometer scanner. The multi-channel digital-to-analog converter provides 18 independent biases, pre-programming the 6 × 6 BPPU array to the target response matrix. A TIA and a lock-in amplifier are used to acquire the resulting time-series current signals generated by the BPPU under dynamic laser spot stimulation.

The acquired noise-free trajectories and with-noise variants were used to construct different datasets. For evaluating the impact of each noise intensity level, a dedicated dataset was assembled per intensity, consisting of 256 noise-free samples and 320 with-noise variants samples, yielding 576 samples in total. For the comprehensive evaluation involving integrated training across all noise types and intensities, a combined dataset was created by merging the noise-free data with all noisy samples, resulting in 3232 samples in total. Each dataset was split into 60% for training, 20% for validation, and 20% for testing. The BPPU-generated time-series features were fed into an LSTM network comprising a 128 hidden unit LSTM layer, a fully connected layer, and a softmax output layer, which was trained with the Adam optimizer (initial learning rate: 0.001) for up to 100 epochs.

Author contributions

Z.Z. and W.X. conceived the concept. T.S., W.X., and X.W. supervised the project. Z.Z. designed, fabricated, and measured the devices and circuits. Y.L. characterized the films, and W.L. conducted the XPS characterization. Z.Z. and Y.L. developed the imaging system and motion trajectory experimental platform. Y.Z. (Yongjian Zheng), N.P., Y.Z. (Yujian Zheng), D.L., H.L., and J.C. analyzed the data. Z.Z., T.S., X.W., and W.X. wrote the paper. All the authors discussed the results and implications and reviewed the paper.

Peer review

Peer review information

Nature Communications thanks Su-Ting Han, Xiangyue Meng and Ye Zhou for their contribution to the peer review of this work. A peer review file is available.

Data availability

The source data that support the plots within this paper are provided with this paper. Other data that support the findings of this paper are available from the corresponding author upon request. Source data are provided with this paper.

Code availability

The codes used for LSTM network training are available from the corresponding authors upon request.

Competing interests

The authors declare no competing interests.

Supplementary information

The online version contains supplementary material available at 10.1038/s41467-026-71638-y.