Phase-independent wireless data aggregation via optical in-sensor computing Authors: Shuyi Sun, Zexi Lu, Yao Wang, Zihui Liu, Zhuang Mao, Qi Shi, Fengzhi Wang, Wei Wang, Jie Jiang, Bin Lu, Haiping He, Xinhua Pan, Haoliang Qian, Zhizhen Ye Published: 2026 ABSTRACT The data-intensive nature of the Internet of Things (IoT) significantly challenges conventional communication systems. While over-the-air computation integrates communication and computation to reduce data aggregation burdens, it requires strict synchronization of transmitted signals, increasing system overhead. Here, we report an optical in-sensor wireless data aggregation paradigm that exploits photocarrier trapping in defect-dominated persistent photoconductance sensors to reduce dependence on synchronization constraints. Within a single communication window, the system reliably aggregates incoherent optical signals with relative timing offsets exceeding the signal duration by up to 120%, while achieving computation errors below 1.5% normalized mean squared error. This process achieves unbiased integration with pulsed low-frequency readout, indicating its potential for ultra-low power operation. A distributed temperature prediction model based on this paradigm shows faster convergence and competitive accuracy with a simplified architecture, demonstrating effectiveness in distributed edge sensing scenarios. This work provides a promising hardware implementation for low-power data aggregation in large-scale IoT edge networks. Introduction In the future Internet of Things (IoT), more than a billion edge intelligent devices will be deployed to collaboratively sense, learn, and communicate in real time, enabling ubiquitous intelligence across every layer of our environment1–4. Meanwhile, massive wireless connectivity and data computation in IoT networks impose significant challenges on resource-constrained edge communication systems5–12. Classical wireless data aggregation (WDA) relies on multiple access schemes for signal transmission and centralized fusion at edge or cloud servers, which inherently incur considerable communication overhead and latency13–16. This has prompted researchers to develop advanced WDA solutions that minimize latency and energy consumption, which is crucial for supporting key IoT applications such as distributed sensing and federated learning17–19. Building on these challenges, over-the-air computation (AirComp) has emerged as an effective solution that seamlessly fuses communication and computation in WDA20–23. By exploiting concurrent transmission and the inherent analog waveform superposition of multiple-access channels at the radio frequency layer, AirComp enables the direct acquisition of aggregated results, including sums or averages at the fusion center without decoding any individual signals24–26 (Fig. 1a). This principle transforms the wireless medium itself into a computational fabric, achieving ultra-fast and highly spectrum-efficient aggregation.Fig. 1Paradigms of wireless data aggregation technology.a Schematic comparison between classical data aggregation and over-the-air computation (AirComp). AirComp enables the integration of communication and computation. b Schematic of AirComp. Encoded signals are superimposed in the air, while asynchronous phases lead to the distortion in the aggregated signal. c Schematic of the proposed in-sensor wireless data aggregation (WDA) paradigm. Incoherent optical signals encoded from edge data excite photocarriers in a persistent photoconductance (PPC) sensor, resulting in the trap-assisted photocurrent for in-sensor data aggregation. Paradigms of wireless data aggregation technology. a Schematic comparison between classical data aggregation and over-the-air computation (AirComp). AirComp enables the integration of communication and computation. b Schematic of AirComp. Encoded signals are superimposed in the air, while asynchronous phases lead to the distortion in the aggregated signal. c Schematic of the proposed in-sensor wireless data aggregation (WDA) paradigm. Incoherent optical signals encoded from edge data excite photocarriers in a persistent photoconductance (PPC) sensor, resulting in the trap-assisted photocurrent for in-sensor data aggregation. Nevertheless, AirComp faces significant challenges due to signal distortions arising from channel fading27, channel estimation errors28, and carrier frequency offset29 (Fig. 1b). While numerous channel pre-compensation techniques have been proposed to mitigate such distortions30–33, they generally assume idealized conditions of perfect time synchronization. Unfortunately, in networks comprising thousands of low-cost, heterogeneous, and randomly deployed IoT devices, achieving nanosecond-level phase synchronization is both physically and economically not practically feasible. The scale and diversity of these networks make such precise synchronization highly costly34, leading to inevitable phase asynchrony that imposes fundamental limitations on the practical deployment of AirComp in real-world scenarios. In light of the inherent difficulty in achieving precise synchronization, digital AirComp has been proposed as an effective algorithmic-centric strategy for this fundamental physical limitation, which leverages advanced digital modulation and channel coding to effectively mitigate asynchrony challenges35,36. While this approach is robust, it introduces significant computational complexity at the receiver. This prompts a different, hardware-centric research question: what if computation could be divorced from the phase at the physical sensor level? Emerging neuromorphic science offers a promising solution for the asynchronous-native WDA paradigm. By leveraging the carrier dynamics within sensors, neuromorphic devices can perform a range of in-sensor computing functions37–40. In particular, a class of sensors exhibiting spontaneous carrier-trapping dynamics generates persistent photoconductance (PPC) that serves as non-volatile weights, which has been widely utilized in neuromorphic computing41–45. These sensors are referred to as PPC sensors. Interestingly, this in-sensor carrier-trapping dynamics bears a functional resemblance to the superposition of signals in the air during AirComp. The key distinction, however, is that the generation and trapping of photogenerated carriers depend solely on the flux (pulse intensity or width) of the incident light and the intrinsic response of the sensor, remaining unaffected by signal phase. Consequently, the development of WDA based on neuromorphic PPC sensors holds the promise to bypass the need for phase synchronization. Among PPC sensors, Ga2O3 is particularly notable for its large hole effective mass, resulting in a pronounced PPC effect46,47 and bridging the gap between conventional photodetectors and neuromorphic devices. Previous studies have shown that the PPC properties of Ga2O3 can be effectively tuned through oxygen vacancy engineering43,48–51, allowing controllable material synthesis and defect engineering. Furthermore, our prior work52 demonstrated that forming a Ga2O3/ZnO heterojunction introduces negative photoconductivity under illumination, enabling bidirectional photoconductive modulation. In conventional PPC sensors, a major limitation is the difficulty of eliminating the PPC effect, as the resulting drift in conductance hinders repeated operation. In contrast, the bidirectional modulation in Ga2O3/ZnO devices allows the PPC state to be readily reset, enabling repeated multi-cycle in-sensor computation. In this article, we report an in-sensor WDA paradigm based on the PPC sensor. Edge device data are encoded using incoherent optical pulses to generate photocarriers in a Si-doped Ga2O3/ZnO-based oxide semiconductor photodetector, which sequentially aggregates these signals through the sensor’s intrinsic carrier-trapping capability (Fig. 1c). Leveraging its strong PPC effect, the sensor achieves a normalized mean squared error (NMSE) below 1.5% even for 120% asynchronous signals, while performing aggregation without any externally applied bias. A distributed temperature prediction system based on this WDA paradigm effectively forecasts local temperatures under large-scale asynchronous sensing, exhibiting faster convergence and reduced model complexity compared with conventional global approaches. This paradigm demonstrates the integration of neuromorphic science and wireless communication systems, offering an efficient WDA solution for resource-constrained edge IoT applications. Results Data encoding via a persistent photoconductance sensor To validate the proposed WDA paradigm, a PPC sensor based on Si-doped Ga2O3/ZnO heterojunction was employed, which has been demonstrated to exhibit neuromorphic computing capability in our previous work52. The device is composed of a fully overlapped ZnO and Si-doped Ga2O3 heterostructure (Fig. 2a). Interdigitated electrodes are used to extract the device conductance, and a photograph of the device is shown in Supplementary Fig. 1. X-ray diffraction (XRD) confirmed the coexistence of crystalline β-Ga2O3 and ZnO films within the sensor (Supplementary Fig. 2), while transmission electron microscope (TEM) and energy-dispersive X-ray spectroscopy (EDS) indicate extensive diffusion between ZnO and Ga2O3 (Supplementary Fig. 3), resulting in a lattice distortion. Further electron paramagnetic resonance (EPR) spectra indicate that this lattice distortion introduces a higher density of defect states into the device (Supplementary Fig. 4). X-ray photoelectron spectroscopy (XPS) further reveals the abundant oxygen vacancies in Ga2O3 (Supplementary Fig. 5). Upon 255 nm UV illumination, a large number of photogenerated carriers are produced in Ga2O3 layer due to its strong absorption in the DUV band (Supplementary Fig. 6). Owing to the inherently low hole mobility and the presence of abundant defect states in Ga2O3, photogenerated holes are readily trapped, leaving free electrons in the conduction band and giving rise to PPC46,50.Fig. 2Optoelectronic characteristics of the persistent photoconductance sensor.a Schematic of the neuromorphic computing in the persistent photoconductance (PPC) sensor. Photogenerated carriers act as analogs of neurotransmitters in biological synapses, where self-trapped holes and free carriers in the valence band (VB)/conduction band (CB) modulate the device conductance. b Surface potential evolution of Ga2O3 before and after 255 nm illumination. After turning off the light, recombination between surface electrons and self-trapped holes is hindered, inducing the PPC effect. Scale bar represents 100 nm. c Dependence of photoconductance variation (ΔI, purple) on 255 nm pulse width and power intensity. The magnitude of photoconductance response can be encoded by both parameters. d Determination coefficients (R2, purple) of the linear relationship between photoconductance variation and optical power intensity under different pulse widths. R2 values approaching 1 (>0.98) indicate that the sensor’s response can be linearly encoded by light intensity at each pulse width. e Reversible photoconductance (purple) modulation of the sensor. 255 nm illumination induces PPC, while 370 nm light resets the photoconductance. f–h Carrier dynamics in the heterojunction sensor under different illumination conditions. Photogenerated holes excited by 255 nm light (f) become self-trapped and remain captured in the dark (g), and are subsequently recombined with photogenerated electrons under 370 nm illumination (h). Optoelectronic characteristics of the persistent photoconductance sensor. a Schematic of the neuromorphic computing in the persistent photoconductance (PPC) sensor. Photogenerated carriers act as analogs of neurotransmitters in biological synapses, where self-trapped holes and free carriers in the valence band (VB)/conduction band (CB) modulate the device conductance. b Surface potential evolution of Ga2O3 before and after 255 nm illumination. After turning off the light, recombination between surface electrons and self-trapped holes is hindered, inducing the PPC effect. Scale bar represents 100 nm. c Dependence of photoconductance variation (ΔI, purple) on 255 nm pulse width and power intensity. The magnitude of photoconductance response can be encoded by both parameters. d Determination coefficients (R2, purple) of the linear relationship between photoconductance variation and optical power intensity under different pulse widths. R2 values approaching 1 (>0.98) indicate that the sensor’s response can be linearly encoded by light intensity at each pulse width. e Reversible photoconductance (purple) modulation of the sensor. 255 nm illumination induces PPC, while 370 nm light resets the photoconductance. f–h Carrier dynamics in the heterojunction sensor under different illumination conditions. Photogenerated holes excited by 255 nm light (f) become self-trapped and remain captured in the dark (g), and are subsequently recombined with photogenerated electrons under 370 nm illumination (h). This carrier-trapping process resembles the operation of biological synapses, in which the release of neurotransmitters induces non-volatile modulation of the postsynaptic membrane potential53, thereby tuning synaptic weights. The surface potential distribution of Ga2O3 before and after UV illumination further evidences this carrier-trapping dynamics (Fig. 2b). After the onset of 255 nm illumination, the average surface potential of Ga2O3 increases rapidly from 362 to 649 mV, indicating substantial accumulation of photogenerated electrons on the film surface. Subsequently, the surface potential remains in an electron-accumulated state for more than 30 min after the illumination is removed (Supplementary Fig. 7), attributed to self-trapping of photogenerated holes during exposure. The persistence of these unrecombined electrons on the film underlies the observed PPC effect in the sensor. To ensure the validity of information encoding using 255 nm UV illumination, the sensor’s photoconductance responses were measured under varying light intensities and pulse widths (Fig. 2c). The sensor exhibits distinguishable conductance states at different luminous flux levels, which is critical for decoding optical information. Linear fitting of the conductance responses to varying light intensities yields the coefficient of determination (R2) approaching unity for each pulse width, indicating that the sensor maintains high linearity under 255 nm illumination at different intensities (Fig. 2d). Thus, the light intensity can serve as a reliable information carrier, establishing it as an effective medium for data encoding in this system. This property provides a solid physical basis for performing analog weighted summation with the sensor. In this scheme, the intensity of each optical pulse encoded by an edge device can be regarded as an input weight, while the net increase in the sensor’s total conductance corresponds to the sum of the data aggregation. Moreover, to mitigate potential signal weight fading in real wireless channels, effective pre-compensation can be achieved by adjusting the light intensity at the transmitting node based on the distance to the receiving node. Meanwhile, owing to the absorption of 370 nm UV light by ZnO, the photoconductance of the Ga2O3/ZnO sensor can be reversibly modulated under two UV wavelengths (Fig. 2e), offering an additional degree of freedom for signal computation. This behavior arises from the intrinsic carrier dynamics within the diffusion layer in heterojunction device, which is absent in the individual devices (Supplementary Fig. 8). Furthermore, we attribute this carrier dynamics to the high density of oxygen vacancy defects enhanced by lattice distortion in the diffusion region, since the sample with low oxygen vacancy concentrations similarly does not exhibit reversible photoresponse (Supplementary Fig. 9). Under 255 nm excitation, photocarriers generated in Ga2O3 produce a photocurrent (Fig. 2f), enabling optical signal decoding. After the light is turned off, holes captured by oxygen vacancies in diffusion region drive the aggregation progress of the decoded signals, with each optical pulse giving rise to a corresponding self-trapped hole state (Fig. 2g). When illuminated at 370 nm, photogenerated electrons in ZnO recombine with the self-trapped holes in the diffusion region, allowing free carriers to enter the external circuit and generate a transient photocurrent that persists only during exposure (Fig. 2h). Once the light is removed, the disappearance of the trapped holes restores the sensor’s conductance to its pre-aggregation level. Through this sequence of UV-driven operations, the heterojunction sensor is capable of aggregating target data while preventing interference between successive aggregation events. Implementation of single-round in-sensor data aggregation Based on the tunable photoconductance of the PPC sensor, we validated its single-shot in-sensor WDA capability experimentally. Specifically, three LEDs, acting as edge-device transmitters, generated three incoherent 255 nm UV pulses with identical pulse widths but different power intensities (Fig. 3a). The PPC sensor served as the receiver, collecting and aggregating the transmitted signals under fully synchronous or varying asynchronous conditions. The aggregated signals were recorded as device currents using a Keithley 4200-SCS semiconductor parameter analyzer in both high-time-resolution mode (for UV pulse widths above 30 ms) and high-current-resolution mode (for UV pulse widths below 20 ms). For comparison, the currents induced by individual pulses were also measured to determine the ideal aggregated response. The detailed experimental setup is illustrated in Supplementary Fig. 10. The robustness of the device toward incoherent signal aggregation was assessed by calculating the NMSE between the sum of individual pulse-induced currents and the measured current induced by aggregated input. Definitions of the asynchronous ratio and the NMSE calculation are provided in Supplementary Note 1.Fig. 3Incoherent signal aggregation based on the persistent photoconductance sensor.a Schematic of incoherent 255 nm optical signal inputs. Pulses with equal widths but different power intensities (16.8 μW, 30.7 μW, 51.6 μW) are applied to the sensor as synchronous/asynchronous inputs. b Sensor’s response (ΔI) to three incoherent synchronous inputs (90 ms width, blue, orange, and green) and aggregated input (purple). The cumulative current for individual inputs (308.8 nA) exceeds the aggregated measured value (301.42 nA), indicating the existing computation error. c Sensor’s response to three incoherent 120% asynchronization inputs (90 ms width) and aggregated input. The aggregated measured value (285.64 nA) is lower than that under synchronous input, indicating additional computation error due to asynchrony. d Sensor’s response to a single 51.6 μW input (100 μs width) in high-time-resolution mode. A dark current fluctuation of ~1.1 nA exceeds the photocurrent (0.49 nA). e Sensor’s response to a single 51.6 μW input (100 μs width) in high-current-resolution mode. The optical response exhibits a step-like behavior on the averaged baseline current, enabling accurate readout (0.32 nA). f Heatmap of device response deviations across different pulse widths and asynchronization levels, highlighting errors due to asynchronization (above 30 ms) and noise (below 20 ms). g Normalized mean squared error (NMSE) for diverse pulse widths and asynchronization levels. Each dot represents the mean of 10 repeated measurements, and error bars indicate the standard deviation (SD). Pulse widths above 30 ms were measured in high-time-resolution mode, while widths below 20 ms were measured in high-current-resolution mode. h Current spectrum density (CSD) of the device in high-time-resolution mode. 1/f noise is observed at frequencies below 3 Hz. Incoherent signal aggregation based on the persistent photoconductance sensor. a Schematic of incoherent 255 nm optical signal inputs. Pulses with equal widths but different power intensities (16.8 μW, 30.7 μW, 51.6 μW) are applied to the sensor as synchronous/asynchronous inputs. b Sensor’s response (ΔI) to three incoherent synchronous inputs (90 ms width, blue, orange, and green) and aggregated input (purple). The cumulative current for individual inputs (308.8 nA) exceeds the aggregated measured value (301.42 nA), indicating the existing computation error. c Sensor’s response to three incoherent 120% asynchronization inputs (90 ms width) and aggregated input. The aggregated measured value (285.64 nA) is lower than that under synchronous input, indicating additional computation error due to asynchrony. d Sensor’s response to a single 51.6 μW input (100 μs width) in high-time-resolution mode. A dark current fluctuation of ~1.1 nA exceeds the photocurrent (0.49 nA). e Sensor’s response to a single 51.6 μW input (100 μs width) in high-current-resolution mode. The optical response exhibits a step-like behavior on the averaged baseline current, enabling accurate readout (0.32 nA). f Heatmap of device response deviations across different pulse widths and asynchronization levels, highlighting errors due to asynchronization (above 30 ms) and noise (below 20 ms). g Normalized mean squared error (NMSE) for diverse pulse widths and asynchronization levels. Each dot represents the mean of 10 repeated measurements, and error bars indicate the standard deviation (SD). Pulse widths above 30 ms were measured in high-time-resolution mode, while widths below 20 ms were measured in high-current-resolution mode. h Current spectrum density (CSD) of the device in high-time-resolution mode. 1/f noise is observed at frequencies below 3 Hz. In high-time-resolution mode, the aggregated behavior of the sensor is clearly detectable. Upon synchronized inputs, the aggregated output of the sensor is slightly lower than the ideal value (Fig. 3b). Despite the identical total optical intensity between the independent and multiple inputs, the concurrent excitation of multiple signals produces a greater density of photocarriers within the sensor. These carriers are more likely to be trapped by defect-related recombination centers54, resulting in a reduced photoconductive gain. This non-linear aggregation, where the aggregated signal is less than the ideal sum, reveals an early onset of saturation behavior43,55,56, which was also observed in the Ga2O3/ZnO sensor under prolonged 255 nm illumination (Supplementary Fig. 11a). Nevertheless, the device retains acceptable linearity under pulse widths below 2.8 s (Supplementary Fig. 11b, c). Simultaneously, the device’s response under varying illumination intensities in the saturated state was evaluated, exhibiting a 20 dB dynamic range with a nonlinearity of 0.9 (Supplementary Fig. 11d). In the linear pulse regime, the range further extends up to 29 dB (Supplementary Fig. 12) due to the presence of numerous unfilled defect states, indicating that PPC sensors operated with shorter pulse widths can maintain accurate aggregation with minimal computational error arising from saturation effects. Compared with the synchronous case, the aggregated value under asynchronous inputs further decreased, primarily due to conductance decay during the asynchronous intervals (Fig. 3c). Owing to the strong charge-trapping capability of the Ga2O3/ZnO sensor (Supplementary Fig. 13), the limited decay introduced minor errors. Supplementary Table 1 reports the relaxation times of the device analyzed under different powers. Lower light intensities correspond to longer relaxation times, attributed to a higher contribution from photocarriers trapped in deep-level states. This also suggests that the effect of conductance decay on asynchronous errors is expected to be further reduced at short pulse widths. However, as the pulse width is further reduced and the signal intensity decreases, noise becomes increasingly significant. Figure 3d shows that at a 100 μs signal pulse width, the device exhibits a current fluctuation of ~1.1 nA, exceeding the signal response of 0.49 nA, which leads to pronounced NMSE variations at short pulse widths and limits robust operation (Supplementary Fig. 14). Notably, the persistent carrier-trapping process in PPC devices appears as variations in the baseline current. Although a signal-to-noise ratio (SNR) below 1 would theoretically bury the response, smoothing of the baseline reveals the step-like optical response of the device. This operation can be directly performed using the high-current-resolution mode of the Keithley 4200-SCS, which applies single-point averaging within a sampling window. Consequently, a more accurate computation can be extracted from the baseline, enabling device operation at a 10 kHz bandwidth (Fig. 3e). This step-like response characteristic of PPC devices helps maintain computational accuracy under low SNR conditions. The deviations shown in Fig. 3f, defined as the mean difference between the ideal WDA value and the actual device response, clearly reveal the influence of signal asynchronization and noise on the aggregation results under different WDA conditions. At longer pulse widths, slight saturation caused by overlapping signals leads the measured values to be slightly lower than the ideal values, and this deviation gradually increases with the effect of asynchronous decay. At shorter pulse widths, the influence of noise surpasses that of saturation and conductance decay, resulting in fluctuations of the deviation around zero. Thanks to the abundant defect states within the device, the NMSE under synchronous inputs remains below 0.5% (Fig. 3g). Under asynchronous conditions, the NMSE reaches a maximum of ~1.5% at 120% fully asynchronous offset for 30 ms pulses, due to both conductance decay and noise. After baseline averaging, the device demonstrates nearly asynchronous-insensitive WDA performance for pulse widths of 20 ms and below (NMSE < 0.5%). RMSE calculated in Supplementary Fig. 15 exhibits the same error trends, further demonstrating the device’s high accuracy and robustness in single-round WDA. To investigate the intrinsic noise sources in the device, the current spectral density (CSD) was measured. The results reveal a pronounced 1/f noise below ~3 Hz, whereas at higher frequencies the noise is dominated by frequency-independent white noise (Fig. 3h). Since PPC device operation relies on carrier trapping at defect states, the observed 1/f noise is attributed to random carrier capture and emission associated with abundant deep-level defects. Importantly, owing to the device’s characteristic step-like PPC response, this 1/f noise has a negligible impact on computational performance, especially after high-current-resolution averaging under low-SNR conditions. For comparison, a PPC sensor based on zinc-tin oxide (ZTO) was also fabricated, which displays a narrower linear range and faster decay speed (Supplementary Fig. 16). Consequently, while the ZTO sensor achieves accurate WDA under low photon flux and minimal asynchronization, its computational errors increase markedly with higher illumination and larger asynchronization ratios (Supplementary Fig. 17). Furthermore, a benchmarking comparison of the Ga2O3/ZnO device with reported PPC sensors is provided in Supplementary Table 2, and the calculation details are described in Supplementary Note 2. These findings highlight that selecting materials with broad linearity and strong charge-retention is crucial for precise and robust WDA based on the PPC sensor. Saturation limits in multi-round data aggregation After demonstrating the robustness and computational accuracy of the PPC sensor in single-round aggregation, we further explored its performance in continuous, multi-round data aggregation. This more complex scenario not only demands high computational precision but also imposes stringent requirements on the charge-trapping and recombination dynamics within the sensor, which ultimately define its computational limits. To investigate this, we employed three incoherent 20-ms-width LED inputs to simulate a set of edge data signals in a single communication aggregation window up to 200 ms, and the window was repeated for multiple rounds to probe the corresponding conductance response of the Ga2O3/ZnO sensor. For a single set of synchronized inputs, the sensor exhibits a photoconductive gain that subsequently decays until the next input set (Fig. 4a). Since the WDA process relies solely on the aggregated signals, the difference in photoconductive gain between consecutive input sets was taken as the readout value for each round during multi-round data aggregation. Remarkably, even with a pulsed readout voltage of 10 Hz, the device still exhibited both gain and decay processes across multiple input sets (Fig. 4b). This indicates that data aggregation within PPC sensors can occur without the need for an externally applied bias, owing to the intrinsic photocarrier dynamics of the material. Once the 200 ms window closes, the receiving center only needs to perform a single transient readout to obtain the calibrated aggregate signal from the asynchronous inputs within that round. Compared to continuously biased computation, this dynamic allocation reduces aggregation energy consumption in a single compute window from 14.45 to 0.74 nJ, approximately 5.12% of the original value (Supplementary Note 3), which is critical for WDA in energy-constrained edge computing scenarios.Fig. 4Low energy consumption aggregation over multiple communication rounds.a Photoresponses of the sensor and local magnified view (dashed graph) after multiple inputs of three synchronous inputs under continuous 0.5 V readout bias. Gain and decay behavior in photoconductance are observed during consecutive signal inputs, while the aggregation result is obtained from the difference between two consecutive conductance changes. b Photoresponses of the sensor and local magnified view (dashed graph) after multiple synchronous signal inputs under 10 Hz 0.5 V readout bias. During the computing process, the sensor still exhibits gain and decay behavior in photoconductance following signal inputs without applied bias, enabling data aggregation without external power. c Normalized mean squared error (NMSE) after multiple incoherent inputs under zero-bias computation with 50% asynchronous (asyn) signal inputs (orange), zero-bias computation with synchronous (syn) signal inputs (blue), and biased computation with synchronous signal inputs (purple). Errors increase rapidly with the number of inputs, corresponding to the depletion of the sensor’s computational memory. d Schematic of the conductance-reset scheme. After every 25 computations, 370 nm light-induced electrons (red) partially clear holes (blue) in trap states, restoring the computational memory for aggregation in the next aggregation round. e Photoresponses of the sensor and local magnified view (dashed graph) to multiple incoherent asynchronous (50%) signal inputs under the conductance-reset scheme. Instantaneous photoresponse to 370 nm light may introduce extra error, which can be avoided through coordinated readout and timing of the reset operation. f NMSE after multiple incoherent inputs under the conductance-reset scheme. NMSE gradually converges after repeated inputs, indicating incomplete reset and controllable aggregation computation error. Low energy consumption aggregation over multiple communication rounds. a Photoresponses of the sensor and local magnified view (dashed graph) after multiple inputs of three synchronous inputs under continuous 0.5 V readout bias. Gain and decay behavior in photoconductance are observed during consecutive signal inputs, while the aggregation result is obtained from the difference between two consecutive conductance changes. b Photoresponses of the sensor and local magnified view (dashed graph) after multiple synchronous signal inputs under 10 Hz 0.5 V readout bias. During the computing process, the sensor still exhibits gain and decay behavior in photoconductance following signal inputs without applied bias, enabling data aggregation without external power. c Normalized mean squared error (NMSE) after multiple incoherent inputs under zero-bias computation with 50% asynchronous (asyn) signal inputs (orange), zero-bias computation with synchronous (syn) signal inputs (blue), and biased computation with synchronous signal inputs (purple). Errors increase rapidly with the number of inputs, corresponding to the depletion of the sensor’s computational memory. d Schematic of the conductance-reset scheme. After every 25 computations, 370 nm light-induced electrons (red) partially clear holes (blue) in trap states, restoring the computational memory for aggregation in the next aggregation round. e Photoresponses of the sensor and local magnified view (dashed graph) to multiple incoherent asynchronous (50%) signal inputs under the conductance-reset scheme. Instantaneous photoresponse to 370 nm light may introduce extra error, which can be avoided through coordinated readout and timing of the reset operation. f NMSE after multiple incoherent inputs under the conductance-reset scheme. NMSE gradually converges after repeated inputs, indicating incomplete reset and controllable aggregation computation error. Despite the significant reduction in power consumption for multi-round WDA enabled by the sensor’s unbiased computation capability, its computational error increases rapidly with the number of aggregation rounds (Fig. 4c). In contrast, the error from asynchronous signal input remains negligible, indicating the saturation-dominant inherent limitations of multi-round aggregation. This asynchrony-insensitive behavior is primarily attributed to the depletion of available hole trapping sites within the sensor. In high-round computations, the signals mainly induce a diminished conductance gain and an accelerated conductance decay governed by shallow traps or free excitons (Supplementary Fig. 18). Consequently, the computation result read out at the end of the communication window remains close to the value determined by the depleted deep trapping states. The saturation effect becomes more pronounced upon successive aggregation rounds, ultimately limiting the potential of PPC sensors for multi-round WDA applications. To address this issue, it is essential to find a method that effectively mitigates the saturation of PPC sensors, thereby leaving more defect sites for aggregation. The reversible photoresponse of the Ga2O3/ZnO sensor under 370 nm illumination offers a potential solution. Building upon previous multi-round aggregation experiments with 255 nm signals, we applied a 370 nm pulse every 25 aggregation rounds to clear historical data and enhance the accuracy of subsequent calculations (Fig. 4d). The computation results (Fig. 4e) demonstrate that this reset operation successfully reduces photoconductance, allowing for more stable and repeatable performance in high-round aggregations. However, this procedure introduces additional invalid points due to the short-term forward photoconductance induced by the 370 nm light, observable during full-bias readout (Supplementary Fig. 19). Proper timing of the dynamic readout may avoid reading these invalid points. Regarding computational accuracy, although NMSE values at lower rounds remain consistent with previous results, the sensor exhibits significantly reduced and convergent errors in higher-round calculations (Fig. 4f), as further corroborated by RMSE trends (Supplementary Fig. 20). However, the reset scheme does not fully restore the sensor’s conductance to its initial level, as the photogenerated electrons from the 370 nm light are insufficient to completely recombine self-trapped holes in Ga2O3. The 370 nm illumination with 0.1 s pulse width can only suppress about 10% of the conductance within 1.5 s, while a complete conductance reset requires a larger 370 nm optical flux or a recovery time of up to 4 s (Supplementary Fig. 21). Consequently, this reset scheme offers the potential for controllable aggregation errors, but also indicates that PPC sensors inevitably incur larger computational errors, which are too large for practical high-frequency iterative applications. This incomplete recombination also implies the complex distribution of defects in PPC materials, such as varying ionization states of oxygen vacancies50, interface mismatches between heterojunctions57, and lattice distortions from doping58, which constrain the effectiveness of the reset scheme. Therefore, the enhancement of PPC sensor performance in multi-round WDA relies on defect engineering to improve the controllability of defect capture dynamics in data aggregation. Another proposed solution integrates a highly photosensitive PPC epitaxial sensing layer with a memristor core computing layer. The PPC layer enables asynchronous optical aggregation tolerant to spatial and temporal variations, while the memristor layer provides nanosecond-scale, reliable voltage-pulse switching for rapid resets and computation59. In this architecture, sensing and aggregation are powered by optical signals, whereas ultrafast forced resets are performed electrically at the backend, enabling higher-bandwidth WDA. Careful design and tuning of PPC reset schemes can significantly enhance neuromorphic device performance, supporting higher-performance phase-independent WDA via optoelectronic heterogeneous integration. Data aggregation for a large-scale distributed application The observed accuracy degradation in multi-round WDA reveals the sensor’s fundamental limitation: saturation of available trapping sites. This suggests the device’s performance is constrained by the total integrated photon flux, regardless of whether the flux arrives temporally (multi-round) or spatially (large-scale parallel). A critical question is therefore whether the device can tolerate large-scale parallel inputs before this saturation threshold is breached. This type of application-specific divergence also provides a promising approach to mitigate WDA inaccuracies arising from saturation. In typical large-scale distributed IoT scenarios, such as smart agriculture soil-moisture monitoring or temporal temperature-field reconstruction in industrial parks60, the global data aggregation cycle often spans tens of minutes to several hours. The system naturally benefits from idle periods of several seconds, allowing a full optical reset to restore conductance and thereby eliminating the effects of saturation. To validate this, multiple short-pulse incoherent inputs, simulating large-scale parallel edge sensing signals, were sequentially input into the PPC sensor. Obviously, the gain in photoconductance induced by subsequent edge signals will diminish due to the saturation effect. However, given the fixed number of available hole trapping sites within the sensor ensured by sufficient conductance 370 nm reset between communication rounds, the total gain (aggregation result) is influenced only by the total optical flux and the saturation effect in each round. Consequently, rather than using the numerical summation of multiple incoherent short-pulse inputs for aggregation, a single long 255 nm pulse is employed as the global signal to simulate the ideal aggregation result and quantify the conductance states corresponding to different defect levels. The optical flux of the global signal is calibrated to match that of the distributed signals, with identical light intensity and a pulse width corresponding to the cumulative width of those individual pulses (Fig. 5a).Fig. 5Massive parallel aggregation for distributed temperature prediction.a Schematic of wireless data parallelization with m parallel channels. The global data with a pulse width of 20 × m ms is encoded into incoherent distributed data pulses (Sn) with the parallel number of m and asynchronization of 150% (20 ms width for each pulse). b Photoresponses of the sensor (ΔI) to global (green) and distributed (purple) data inputs. The distributed aggregation introduces a certain degree of aggregation error under massive parallelization. c Normalized mean squared error (NMSE) under different numbers of parallel inputs. The computation error remains low (NMSE = 1.2%) with a large number of parallelizations (~7000), demonstrating the sensor’s robust computation for parallel data. d Schematic of the distributed temperature-prediction model. The PPC sensor aggregates feature maps (FM) generated from each local IoT device after Long Short-Term Memory (LSTM) preprocessing for further processing. e Convergence speed of global (purple), ideal (green), and asynchronous (blue) temperature-prediction models. The distributed model exhibits significantly faster convergence (100 epochs) compared to the global mode (250 epochs), while the asynchrony-induced aggregation error has minimal impact on convergence. f Actual and predicted temperature distributions obtained by the distributed sensing model. Historical weather parameters over 7 days are used to predict future temperature distributions. Massive parallel aggregation for distributed temperature prediction. a Schematic of wireless data parallelization with m parallel channels. The global data with a pulse width of 20 × m ms is encoded into incoherent distributed data pulses (Sn) with the parallel number of m and asynchronization of 150% (20 ms width for each pulse). b Photoresponses of the sensor (ΔI) to global (green) and distributed (purple) data inputs. The distributed aggregation introduces a certain degree of aggregation error under massive parallelization. c Normalized mean squared error (NMSE) under different numbers of parallel inputs. The computation error remains low (NMSE = 1.2%) with a large number of parallelizations (~7000), demonstrating the sensor’s robust computation for parallel data. d Schematic of the distributed temperature-prediction model. The PPC sensor aggregates feature maps (FM) generated from each local IoT device after Long Short-Term Memory (LSTM) preprocessing for further processing. e Convergence speed of global (purple), ideal (green), and asynchronous (blue) temperature-prediction models. The distributed model exhibits significantly faster convergence (100 epochs) compared to the global mode (250 epochs), while the asynchrony-induced aggregation error has minimal impact on convergence. f Actual and predicted temperature distributions obtained by the distributed sensing model. Historical weather parameters over 7 days are used to predict future temperature distributions. The aggregation results show that as the number of inputs increases, the aggregated result deviates from the ideal value (Fig. 5b), primarily due to the photoconductance decay in asynchronous intervals. Although the asynchronization ratio between distributed signal inputs was set as 150%, after 7000 inputs, the asynchronous interval between the first and last signals expanded by a factor of 7000. Thus, large-scale distributed inputs exacerbate asynchronous differences, challenging the receiver’s tolerance for asynchrony. Aggregation results for both sparse (200% asynchronization) and fully overlapping (synchronous) arrivals also show minor computational deviations arising from the asynchronous differences compared with the standard global signal (Supplementary Fig. 22). Nonetheless, due to its strong charge-retention properties, the PPC sensor maintains a low NMSE of 1.2% after 7000 edge inputs (Fig. 5c), outperforming the computation error in multi-round WDA. The RMSE further confirms the robust aggregation performance under large-scale inputs (Supplementary Fig. 23). Notably, the sensor exhibited anomalously high NMSE for a few edge inputs, though RMSE did not display this pattern. This discrepancy originates from the sensor’s dark current noise, which remains negligible in computations under large-scale edge inputs, as further analyzed in Supplementary Fig. 24 and Supplementary Note 4. Leveraging the distributed data aggregation capability of the PPC sensor, we developed a distributed temperature prediction model, validated using the Global Surface Summary of the Day (GSOD) dataset (2000–2023, daily temperatures for select regions in China). In this model, each edge device is equipped with a two layers Long Short-Term Memory (LSTM) preprocessing network, whose feature data is aggregated by the PPC sensor, and subsequently processed through a convolutional layer to generate temperature predictions (Fig. 5d). For comparison, a global CNN model was constructed to simulate the conventional data processing without WDA, with detailed model specifications provided in Supplementary Table 3. To validate the suitability of the Ga2O3/ZnO device in this scenario, its performance was characterized under representative environmental temperatures (Supplementary Fig. 25), exhibiting negligible drift and stable WDA computation due to the inherent stability of the oxide material system. Cycle-to-cycle variation was evaluated using up to 350,000 asynchronous optical signals, and device-to-device variation was assessed across a 30-device array (Supplementary Figs. 26 and 27). A 2 s 370 nm reset between cycles ensured that the device exhibited only ~8% conductance drift even under such a large number of inputs, demonstrating reproducible distributed computing capability. Device-to-device variation tests indicate that, although miniaturization may introduce edge-field or sidewall defect states causing slight deviations from a Gaussian distribution in both responsivity and computational performance (Supplementary Fig. 28), the devices show a mean NMSE of 0.48% with a standard deviation of 0.22%, reflecting low computational error. These inter-device differences are expected to be absorbed by backpropagation as a form of regularization in our deep learning models, further mitigating the impact of device-to-device variation. Computing results show that while the global model achieves lower error (Supplementary Table 4), it requires approximately 250 training iterations to converge. In contrast, the distributed model converges in only 100 epochs, demonstrating superior training efficiency (Fig. 5e). Furthermore, the distributed model consists of 161,537 parameters, substantially less than the 918,081 parameters in the global model, indicating the distributed model’s practicality in resource-constrained edge environments. To assess the impact of WDA with PPC sensors on model aggregation, we compared ideal data aggregation with aggregation incorporating real asynchronous error. Although the aggregation error induced by significant asynchrony decreased model accuracy (RMSE = 3.2 °C), the model still predicted a relatively accurate temperature distribution (Fig. 5f). The RMSE value is slightly higher than those of similar short-term weather forecasting models61,62 (1–3 °C), but the model demonstrates superior robustness to asynchrony and a low-power configuration. As noted in the previous section, the aggregation of distributed signals does not require a continuous external bias, and only a short bias pulse at the end of data collection is needed to read out the computational result. This approach reduces the energy consumption from 570 μJ to 4.8 nJ for 7000 distributed 20 ms signals (Supplementary Fig. 29). It is further anticipated that, for even larger-scale distributed signal sets, the energy savings could increase by several orders of magnitude. On the other hand, after this bias pulse readout, the system enters an idle period until the next round of sensing begins. In our temperature-prediction system, this interval can last several hours, fully overlapping with the time required for device reset. Therefore, the end-to-end latency of the proposed distributed WDA scheme depends solely on the time from signal emission to completion of aggregation before subsequent decisions can be initiated. Benefiting from the high SNR under large-scale post-signal inputs, the device does not require baseline averaging, and using the high-time-resolution FS-Pro semiconductor parameter analyzer, sampling can be completed in approximately 1 ms (limited by the measurement equipment). Moreover, due to the intrinsic asynchrony tolerance of PPC devices, the time required for aggregation computation is completely independent of the number of participating nodes N, maintaining O(1) temporal complexity. Asynchronous node signals induce defect-state PPC of varying magnitudes within a single aggregation window, representing a phase-independent process. This further underscores the latency-insensitive nature of the proposed WDA scheme. Supplementary Table 5 contrasts various WDA paradigms15,22,35, demonstrating that the proposed in-sensor WDA paradigm offers inherent resilience to asynchrony and reduced communication costs, highlighting its promising applicability in large-scale distributed IoT sensing systems. Discussion We have reported an in-sensor WDA paradigm that enables seamlessly integrated communication and computation. The approach employs the carrier trapping dynamics in PPC sensors, allowing the persistent storage of incoherent optical signals within the sensor for data aggregation. A UV-modulated Ga2O3/ZnO-based PPC sensor was used for principle verification. The results indicate that the in-sensor trapping process operates independently of the need for external bias, significantly reducing system power consumption. Additionally, its persistence provides tolerance to asynchronous signals, resulting in low computation error. Meanwhile, the saturation effect resulting from prolonged carrier trapping limits the paradigm’s capacity for high-frequency operations, making it more suitable for large-scale spatial parallel information aggregation. A potential solution to this trade-off lies in defect engineering, which optimizes defect density and energy levels to enhance the sensor’s dynamic range. The proposed paradigm introduces a function-driven application to neuromorphic engineering, providing an approach to mitigate asynchrony-related limitations in IoT communications. In the future, this paradigm could be further extended through device-level heterogeneous integration with memristors. A rapidly programmable memristor layer can compensate for the high-frequency limitations of a single PPC sensor, while mature memristor fabrication technology offers considerable flexibility in structural expansion. Through the design of different PPC epitaxial layers, this approach naturally supports advanced optical communication technologies such as wavelength-division multiplexing for highly parallel in-sensor computing. Together, these developments could help address the present proof-of-concept limitations, including reset speed, high-frequency operation, dynamic range under dense inputs, and scalable device-array integration, thereby supporting the development of a highly integrated and reconfigurable WDA computing platform. Methods Device fabrication Both ZnO and Si-doped Ga2O3 films were grown on Si substrates with 300 nm-thick thermally grown SiO2 via pulsed laser deposition using pure ZnO and Si-doped Ga2O3 targets (Ga2O3:SiO2 = 99.5:0.5 wt%). The distance between targets and substrates was set to 50 mm. Before deposition, the substrates were ultrasonically cleaned in acetone, ethanol, and deionized water for 15 min each before being dried with high-purity nitrogen gas. The substrates were then cleaned using a 5 Pa O2 plasma with a power of 30 W. The ZnO film was initially grown on the substrates at 300 °C for 5 min using a 300 mJ KrF excimer laser (248 nm) at a frequency of 5 Hz and an O2 pressure of 1 Pa. After that, the ZnO film was heated to 650 °C, and the Si-doped Ga2O3 film was deposited for 60 min at 0.1 Pa O2 using a 300 mJ laser with a frequency of 3 Hz. Following that, the hetero-growth film was annealed at 750 °C for 30 min in an atmosphere. The Ti/Au (20 nm/50 nm) interdigital electrode was deposited by electron-beam evaporation through a stainless steel shadow mask. Ga2O3/ZnO films with the same parameters were also fabricated on quartz substrates to avoid signal absorption from the Si substrate during EPR measurement. Device characterization Micrographs of the fabricated device were acquired using a ZEISS Smartzoom5 digital microscope. The crystallinity of the films was determined using XRD (Shimadzu XRD-7000) with Cu Kα radiation (λ = 0.154 nm). TEM and EDS (JEOL JEM-F200) were employed to capture the cross-sectional image. A UV-Vis spectrophotometer (Shimadzu UV-2600) was utilized to characterize the transmission spectrum. The content of oxygen vacancy was analyzed via XPS (Kratos Axis Supra+). Surface morphology and potential mapping were performed via Kelvin probe force microscopy on a Bruker Dimension Icon atomic force microscope. EPR measurements were conducted using a Bruker EMXplus-9.5/12 spectrometer operating at a frequency of 9.86 GHz. A semiconductor parameter analyzer, Keithley 4200-SCS, was employed to assess single-round WDA performance. In high-time-resolution mode, the device samples at 200 points per second with delay and filter factors set to 0, whereas in high-current-resolution mode, it samples at 20 points per second with delay and filter factors set to 1. The FS-Pro was employed to evaluate WDA performance in multi-round and distributed scenarios, operating at a sampling rate of 103 points per second. The light sources were LEDs with wavelengths of 255 nm and 370 nm controlled by an FPGA. Distributed temperature data preprocessing The input data consists of seven features: average temperature, average dew point, average sea-level pressure, average station pressure, average visibility, average wind speed, and precipitation probability. The spatial grid is defined by dividing the longitude range of 100°–120° and the latitude range of 20°–40° into 21 equal intervals, forming a 21 × 21 grid. Station data is then assigned to the nearest grid cells using the nearest-neighbor method. Temporal data is collected for 28 consecutive days (from day 1 to day 28) each year. For outlier handling, when multiple stations or observations exist for the same (day, feature, grid cell) combination, the average is computed by summing the values and dividing by the number of observations. The resulting data for each year is organized into a tensor with the shape (28 days, 7 climate features, 21 × 21 grid). Data is split by year (2000–2023) using a time-based extrapolation approach. The training set size is determined by a 0.96 train-test ratio. The input temperature data are expressed in Fahrenheit (°F). For consistency and readability, all model-predicted temperature values and unit-dependent error metrics reported in the manuscript and the Supplementary Information have been converted to Celsius (°C). Simulation of the global model The global model consists of an encoder and a decoder, each with two layers of ConvLSTMCell. The ConvLSTMCell processes the current input and the previous hidden state by concatenating them along the channel dimension, followed by a 2D convolution to directly compute the four gates (i/f/o/g). The current cell state and hidden state are then updated according to the LSTM equations, with the gate functions implemented using sigmoid and both the candidate cell state and output using the tanh function. The convolution is implemented with a default kernel size of 3 and padding of 1. For the encoder-decoder connection, the initial hidden state and cell state of the decoder are directly copied from the final step of the encoder’s hidden states. The input to the decoder is composed of the previous predicted temperature (1 channel) and the other 6 contextual features from the previous input frame (static reuse), making a total of 7 channels. At each step, a single predicted map is produced via the output layer, and the model performs autoregressive rolling for a specified number of output days. A masking strategy is applied to locations where no observations have been recorded within the past 7 days (sum over time equals zero), predicting these locations as zero. Finally, a 1 × 1 convolution at the output layer maps the hidden states to a single-channel temperature prediction. Simulation of a distributed model The distributed model consists of a two-layer LSTM at each edge node, which performs sequential modeling of the time series for each grid point along the feature dimension. For each edge station node, the input tensor is flattened into a spatial batch using permute and view. The LSTM then processes the data, reshaping the temporal features into a 1 × 8 × 8 tensor. Each valid station generates an 8 × 8 contribution matrix based on its own historical sequence, which is then fed into an 8 × 8 array of simulated PPC sensors. Each element of the matrix is aggregated by a corresponding PPC device in the array. After aggregation, these contribution matrices form a global 8 × 8 latent meteorological field, which is subsequently input to the following layer. The 2D convolution (Conv2d) at the computing center is applied to the final hidden representation, mapping it to pixel-level prediction channels for future temperature. The convolution operation uses padding equal to kernel//2 (kernel size of 3) to preserve spatial resolution. In the non-ideal asynchronous model, to evaluate the sensor’s performance limits in distributed sensing, the global 8 × 8 latent meteorological field is multiplied by a random tensor of the same shape, with a mean of 0.93449 and a standard deviation of 0.02218. These values correspond to the experimentally measured mean asynchronous decay and device-to-device variation of the sensor array under 500 distributed inputs (exceeding the actual 441-station allocation). Supplementary information Supplementary Information Supplementary Software 1 Transparent Peer Review file Supplementary Information Supplementary Software 1 Transparent Peer Review file Source data Source data Source data