Inhomogeneous Metasurface–CNN System for Filter-Free Single-Pixel Wavelength Recognition

Inhomogeneous Metasurface–CNN System for Filter-Free Single-Pixel Wavelength Recognition | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Article Inhomogeneous Metasurface–CNN System for Filter-Free Single-Pixel Wavelength Recognition John Yeow, Wentao Gao, Jiaqi Wang, guanxuan Lu, Yifei Yuan This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-8006942/v1 This work is licensed under a CC BY 4.0 License Status: Under Review Version 1 posted You are reading this latest preprint version Abstract High-performance visible-light detection is vital for imaging and portable sensing, but most photodetectors are broadband intensity devices without intrinsic spectral selectivity and therefore rely on filter arrays or dispersive optics. Broadband visible photodetection with filter-free wavelength recognition is enabled by an inhomogeneous metasurface integrated with a convolutional Neural Network (CNN). Inhomogeneous metasurface creates geometry-dependent Localized Surface Plasmon Resonance (LSPR) hot spots that imprint wavelength-dependent signatures, while CNN is employed to recognize the signatures. The device shows broadband response in visible region with responsivity up to 1.45×10³ V W⁻¹, noise-equivalent power (NEP) down to 8.57 × 10 –16 W Hz 1/2 , specific detectivity up to 5.71×10 14 Jones. A lightweight 1D-CNN trained on a wavelength-distance dataset achieves 98.91% accuracy in 10-fold cross-validation and is validated by raster-scanning on a laptop display and traffic light recognition. The scalable Ag inhomogeneous metasurface/poly(3,4-ethylenedioxythiophene) polystyrene sulfonate (PEDOT:PSS) process and detector–CNN integration offer a practical route to low-cost, compact, filter-free colorful imaging and spectral identification. Physical sciences/Engineering/Electrical and electronic engineering Physical sciences/Nanoscience and technology/Nanoscale devices/Nanophotonics and plasmonics Physical sciences/Optics and photonics/Applied optics/Optoelectronic devices and components Figures Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 Figure 6 Introduction Detecting visible light across a broad spectral range is central to applications such as high-resolution imaging, environmental monitoring, wearable electronics, and energy-efficient optoelectronic platforms ( 1 ). Established visible photodetectors, such as silicon and organic photodiodes, CMOS image sensors, and emerging heterojunction/2D devices, already offer high signal-to-noise ratios, fast response and wide dynamic range, reliably meeting needs in imaging, machine vision, proximity sensing, and environmental monitoring ( 2 , 3 ). However, most visible photodetectors are broadband intensity devices without intrinsic spectral selectivity and therefore rely on filter arrays or dispersive optics ( 4 , 5 ). Furthermore, conventional silicon-based photodetectors often require external bias to achieve efficient operation, which increases power demand and complicates system integration ( 6 ). These limitations are increasingly prominent for emerging applications, including self-powered sensing, portable spectroscopic devices, and flexible optoelectronics, where high responsivity, low power consumption, and wavelength tunability are critical ( 7 , 8 ). Conductive polymers such as poly(3,4-ethylenedioxythiophene) polystyrene sulfonate (PEDOT:PSS) exhibit strong and broadband absorption across the visible spectrum, in addition to mechanical flexibility and compatibility with solution processing, which support scalable, low-cost fabrication ( 9 – 15 ). Photothermoelectric (PTE) detectors convert light to voltage via the Seebeck effect: optical absorption locally heats the active region and establishes a temperature gradient, which drives an open circuit photovoltage. Compared with other mechanisms, PTE detectors offer self-powered operation with low dark current and shot noise, room-temperature use, broad spectral operation, and relatively fast response. Inhomogeneous metasurfaces can generate localized surface plasmon resonances (LSPRs), enhancing light absorption and accelerating the photothermal conversion process ( 16 – 18 ). This LSPR-driven photothermal not only gain strengthens the photovoltage but also provides the necessary features for spectral selectivity without optical filters. Convolutional neural network (CNN) learnt translation-equivariant features by convolving shared kernels across the input using far fewer parameters than fully connected models. In our voltage–time model, temporal convolutions capture edges, plateaus, and PTE decays, while pooling and striding with normalization improves robustness. Trained end-to-end, lightweight 1D CNNs reduce reliance on handcrafted features and reliably decode wavelength-inhomogeneous metasurface wavelength-dependent signatures. These attributes make PTE-inhomogeneous metasurface-CNN platforms well suited to portable, filter-free color imaging and compact spectral sensing. In this study, we investigated a hybrid visible PTE detector consisting of PEDOT:PSS films integrated with inhomogeneous metasurface on silicon substrates. Our devices show an outstanding performance at room temperature and self-powered, with responsivity up to 1.45×10³ V W⁻¹, noise-equivalent power (NEP) down to 5.54×10 -11 W Hz -1/2 , and specific detectivity up to 4.01×10 10 Jones. Silicon-based devices present notable advantages, including ease of fabrication, compatibility with established chip integration technologies, potential for device miniaturization, and inherently reduced noise ( 19 , 20 ). The device structure is designed to recognize wavelength in a cost-effective way, delivering broadband sensitivity, self-powered operation, and scalable manufacturability. A lightweight 1D-CNN trained on a wavelength-distance dataset attains 98.91% average accuracy under 10-fold cross-validation and demonstrates strong real-world application through raster-scanning on a laptop display and traffic-light recognition. Leveraging the wavelength-dependent response characteristics introduced by the inhomogeneous metasurface and integrating them with a CNN model, we demonstrate the capability to recognize light of different wavelengths and source-to-detector distances. This approach can be further extended toward filter-free colorful imaging and the development of compact, portable spectrometers. By systematically correlating the morphological characteristics of the metasurface with the photoelectric response of the detector, and further employing CNN to distinguish these spectral features, this work provides new insights into plasmon-assisted PTE photodetectors and photodetector-CNN systems, which establishes a pathway toward filter-free wavelength recognition and colorful imaging( 21 ). Results Device structure and metasurface To realize broadband visible absorption, we engineered the device comprising an inhomogeneous silver (Ag) metasurface capped by a 2µm thick PEDOT:PSS film (Fig. 1 A). A 20-nm Ag layer was first deposited on SiO₂ and transformed into an inhomogeneous metasurface by thermal annealing. A PEDOT:PSS film was then formed by spin-coating, and Ti/Pd electrodes were sequentially deposited. The cross-sectional SEM imaging of our detector is shown in Fig. S2. Full fabrication details are provided in Materials and Methods. A representative scanning electron micrograph of the metasurface is shown in Fig. 1 B. Using finite-difference time-domain (FDTD) simulations, Fig. 1 (C and D) maps the electric-field intensity at the interface between the inhomogeneous Ag metasurface and the PEDOT:PSS layer under 390–760 nm illumination. Inhomogeneous metasurface induced wavelength-dependent resonance across different islands, which modulate near-field intensity and local absorption and induce asymmetries in electron-hole transport. Because LSPRs are strongly geometry-dependent ( 16 ), the inhomogeneous metasurface generates spectrally selective and spatially distinct hot spots whose locations shift with wavelength. Moreover, wavelength-dependent variations in island multiplicity imprint distinguishable response signatures within the same PTE detection mechanism, which are exploited for CNN-assisted recognition later. The inhomogeneous metasurface enhances optical absorption and yields a stronger device response, while embedding distinct, recognizable waveform signatures in the output signal. Mechanism and Optical properties As shown in Fig. 2 A, under identical visible illumination, Ti exhibits strong optical absorption, typically converting about 50% of incident energy into heat, whereas Pd is more reflective in the same visible wavelength, with lower absorption, yielding weaker photothermal conversion ( 22 , 23 ). Thermal transport further separates their behavior: Ti’s thermal conductivity (21 W m⁻¹ K⁻¹) is much lower than Pd’s (71 W m⁻¹ K⁻¹), so heat generated in Ti spreads slowly and produces a higher local temperature, while Pd dissipates heat efficiently ( 22 , 23 ). The resulting temperature gradient between Ti and Pd drives the photothermoelectric (PTE) effect. At room temperature, Ti has a negative Seebeck coefficient (S Ti ≈ − 3 to − 7 µV K⁻¹), consistent with electrons diffusing from hot to cold; Pd has a positive Seebeck coefficient (S Pd ≈ + 7 to + 10 µV K⁻¹), corresponding to hole flow from hot to cold. Because the coefficients differ, the net thermopower ΔS ≈ S Pd − S Ti ≈ 15 µV K⁻¹, giving an open-circuit voltage V OC = (S Pd − S Ti ) ΔT. Consequently, a Ti–Pd pair delivers a stable, appreciable PTE signal under illumination ( 17 , 18 ). The device operates in a self-powered mode, delivering strong performance at zero bias and room temperature. Figure 2 (B to D) shows the device responses at different visible wavelengths. The illumination was provided by simple LED sources, driven by an Arduino with a series resistor. Emission peaks were 620 nm (red), 520 nm (green), and 460 nm (blue). All measurements were performed at room temperature and under zero bias. To mitigate thermal effects from the sources, the detector was positioned 20 cm from the sources. We repeated the acquisitions multiple times and observed no obvious decline after six months of ambient exposure. The photoresponse was recorded using a multimeter in a dark room to minimize environmental interference. Responsivity is defined as R v = \(\:\frac{V}{{P}_{in}}\) , where V is recorded average voltage under illumination and P is effective incident light power. NEP is defined by \(\:\frac{\sqrt{2q{V}_{dark}}}{{R}_{v}}\) , where q is elementary charge and V dark is recorded average voltage in the dark environment. Specific detectivity D* is defined as D * = \(\:\frac{\sqrt{A}\:\varDelta\:f}{NEP}\) , where A D is the active area of detector and Δf is the bandwidth. Under blue, green, and red illumination, respectively, the responsivity (R v ) reached 9.73 × 10 2 , 1.38 × 10 3 , and 1.45 × 10 3 V W −1 ; The NEP reached 8.57 × 10 −16 , 9.88 × 10 − 16 , and 9.09 × 10 − 16 W Hz −1/2 ; The specific detectivity (D * ) reached 5.71 × 10 14 , 4.96 × 10 14 , and 5.39 × 10 14 Jones. Rise time is defined as the time required for the output to increase from 10% to 90% of its final steady-state value after the illumination is turned on. Fall time is defined as the time required for the output to decrease from 90% to 10% of its final steady state value after the illumination is turned off. The rise and fall times are 3 ms and 5 ms, respectively. While optoelectronic measurement in this study relies on RGB primaries, the underlying mechanism of our inhomogeneous metasurface detector, LSPR’s wavelength dependence, supports broadband response and recognition. In addition, optoelectronic measurements provide distinguishing features for mechanism identification. The bolometric effect arises from photo-induced heating, which requires an external bias for efficient readout, and it is governed by thermal diffusion, yielding comparatively long-time response. Moreover, in conductive-polymer/metal composite structures, the photovoltaic effect is severely suppressed by short carrier lifetimes, interfacial defects, and metal-induced quenching, so that any photocarriers generated recombine rapidly. Based on our device structure and measurement, the detector operates at zero bias and exhibits a measured millisecond-scale response, aligning with a non-bias photothermoelectric effect, while the bias-dependent bolometric effect is negligible for the response mechanism. Moreover, the device lacks an effective p–n junction or a strong Schottky depletion region and built-in field, and the spectral response follows the localized surface plasmon resonance (LSPR) rather than a semiconductor band gap absorption. Even if the photovoltaic effect occurs, photocarriers are easy to rapidly recombine at metal/polymer interfaces, making it difficult to achieve a significant PV output. Taking together, these results indicate that LSPR-induced, temperature-gradient-driven PTE is the dominant mechanism. Convolutional neural network– inhomogeneous metasurface assisted wavelength–distance recognition. We acquired training and test data indoors under bright ambient lighting using multimeter and LED sources at 620 nm (red), 520 nm (green), and 460 nm (blue). The CNN model is trained to focus on the steady-state waveform morphology, enhancing applicability across diverse illumination scenarios. Integrating with inhomogeneous metasurface, LSPR-enhanced absorption drives wavelength-dependent variations in both the photovoltage waveform and its kinetics. Thus, the voltage–time traces recorded produce learnable features that a convolutional neural network (CNN) can exploit to distinguish classes. Each wavelength generates a time-domain signature with unique, repeatable traits during PTE conversion. In Fig. 3 (A and B), we visualize the visible-light dataset with principal component analysis (PCA) and t-distributed stochastic neighbor embedding (t-SNE) ( 24 , 25 ). PCA, a linear method that maximizes explained variance, reveals some inter-class separation in 2D but also substantial overlap, especially for same wavelength samples at different distances, which cluster nearby and stretch mainly along high-variance directions. By contrast, t-SNE, a nonlinear embedding that preserves local neighborhoods, better exposes the latent manifold, mapping classes into curved, arc-like clusters with tighter intra-class clouds and clearer inter-class spacing. These patterns indicate that conventional approaches struggle to reliably resolve wavelength in this setting. To improve recognition, we adopt deep learning as an end-to-end solution that learns discriminative features directly from data. This approach captures complex, non-linear patterns that traditional feature engineering misses, yielding higher accuracy and better robustness to environmental variation. We collected 1836 samples and applied data augmentation at the outset of training. Voltage–time traces were recorded by a multimeter under three LEDs (460, 520, 620 nm) at 20, 23, 28, 40 cm. After thorough research, CNN, a deep model that learns discriminative features by convolving shared filters, was selected for wavelength recognition. In our CNN model, temporal convolutions capture edges and plateaus, while pooling and normalization provide robustness to amplitude fluctuations and timing jitter. Trained end-to-end with backpropagation, CNNs reduce reliance on hand-crafted features and are widely adopted for signal classification and regression in sensing. These properties make lightweight 1D-CNNs well suited to our inhomogeneous metasurface-detector, where they decode wavelength-dependent waveform signatures into compact spectral labels. Coupling the CNN with the inhomogeneous metasurface enables richer analysis and prediction, enhancing practical utility and highlighting potential for advanced optical sensing. Figure 3 D outlines the workflow and our 1D-CNN architecture. We first unify sequence length and apply per-sample z-score normalization to preprocessed data. Labels are parsed as wavelength–distance composites and encoded into discrete IDs. The classifier comprises two convolutional layers and one pooling layer, optimized with adaptive moment estimation (Adam) (learning rate = 1×10⁻³) and cross-entropy loss. The model has ~ 3.18k trainable parameters, promoting generalization in small-data regimes. Training uses stratified 10-fold cross-validation with early stopping based on validation accuracy (patience = 30 epochs). For each fold, we recorded confusion matrices and classification reports, then aggregate fold-level results and export visualizations and javascript object notation (JSON) summaries to support reproducibility and audit. Considering the modest dataset size and the potential for train–test fluctuations, we employed 10-fold cross-validation to validate accuracy and eliminate the random errors. In Fig. 4 A, per-fold validation accuracies cluster tightly near 1.0, with a mean of 0.9891 and folds spanning roughly 0.985–1.000, indicating low variance across partitions. The accuracy distribution (Fig. 4 E) further shows both the mean and median concentrated around 0.989, reinforcing the stability of performance across folds. Training dynamics are summarized in Fig. 4 B–C. The first three folds’ validation accuracy curves rise rapidly and plateau, while the average loss curve with confidence intervals decrease monotonically and then stabilize. The training and validation losses almost track one another throughout. This parallel behavior of accuracy and loss demonstrates that the model is learning wavelength-distance specific features without overfitting, consistent with our intended design and hyperparameter choices. Class-level behavior across the 12 color–distance classes (3 colors × 4 distances) is shown in Fig. 4 D and Fig. 4 F. The confusion matrix (Fig. 4 D) is strongly diagonal, with only a few data off-diagonal that almost occur within the same color. Cross-color confusions are essentially absent. Correspondingly, per-class precision, recall, and F1-score (F) are uniformly high, evidencing balanced performance without bias toward particular categories. Taken together, these panels show that cross-validation yields an average accuracy of 98.91%. The training and validation curves in 10-fold exhibit marked similarity, indicating appropriate model selection, and that the classifier consistently identifies each wavelength–distance combination. The model therefore avoids common traps, and demonstrates robustness, reliability, and generalizability for wavelength–distance recognition. CNN Model Validation: color recognition and prospects for colorful imaging (A) A digital image illustrating the experimental setup designed to assess the cognitive ability in recognizing the color emitted from laptop screen. A computer screen displays three adjacent color bands, the three primary colors of RGB, which are sequentially scanned by the detector. The resulting electrical signals are recorded using a Keithley 6500 multimeter. (B) Schematic illustration of the experimental setup. (C) Comparison between the original displayed pattern (top) and the classification results obtained from the detector (bottom). Confidence levels for each predicted color are indicated below. Small discrepancies occur at the color boundaries, where classification errors are more likely to arise. To eliminate artifacts from compact LEDs, we used a laptop screen as the light source and raster-scanned the detector at a 3 cm standoff while a digital multimeter recorded the voltage traces. The results are shown in Fig. 5 C. In uniform single-color regions, the three primaries (R 620 nm, G 520 nm, B 460 nm) were classified correctly with posterior confidences ≈ 1.0. Misclassifications appear only at color boundaries where spectra mix spatially. The width of misclassifications area and length of screen are 5 mm and 350mm, so the accuracy is about 97.14% for this figure. At the blue–green boundary, portions of blue are labeled green with high confidence (~ 1.0); at the green–red boundary, portions of red are labeled green with intermediate confidence (~ 0.5). This asymmetry is consistent with the measured responsivities: the detector’s blue responsivity is markedly lower than green, so the green component dominates the composite waveform near the B–G border and the classifier confidently output green; by contrast, red and green responsivities are similar, yielding ambiguous evidence near the G–R seam and correspondingly ~ 0.5 confidence. Residual errors are therefore attributed to boundary mixing and pixel size rather than model bias. (A) Schematic diagram: a green LED luminaire simulates a traffic signal while the user faces the source. The inhomogeneous metasurface photothermoelectric detector (mounted on glasses) acquires the incident light and a lightweight CNN to process the signal, realizing the spectral recognition. (B) Laboratory demonstration of traffic-signal recognition. (C) Close-up of the glasses-mounted module. The detector is laminated on the outer surface of the eyewear, enabling a compact device that outputs signatures for downstream CNN classification and user cues. We prototyped a wearable implementation by combining the inhomogeneous metasurface detector onto the outer surface of safety glasses (Fig. 6 B). Green and red LED were used to simulate the traffic signal (Fig. 6 A); the glasses-mounted detector acquired the incident light and converted it into electrical signatures, which were classified by a lightweight CNN to determine the signal state. A simple tip was then issued to the wearer to indicate the detected state. This proof-of-concept shows that spectral recognition can be realized in a compact, wearable device under typical illumination. These results confirm that the photodetector-CNN system accurately recognizes the RGB primaries and that the waveform signatures imparted by the inhomogeneous Ag metasurface can be exploited beyond point classification. Because the signal arises from wavelength-dependent waveform morphology, the approach can be extended to filter-free colorful imaging by scanning or by array integration. In this system, the metasurface–CNN system could reduce reliance on dispersive optics in certain regimes and enable portable, low-cost spectral identification, offering a path toward compact, intelligent spectrometers. The system’s high accuracy, stable ambient operation, and robustness to modest motion suggest reliable use in dynamic scenes, lowering the cost of visible-light imaging and spectral recognition. The combination of an inhomogeneous metasurface with a lightweight deep-learning classifier thus provides a practical route to embedded optoelectronic sensing and opens opportunities for advanced visible-light communication applications. Discussion In conclusion, we designed and fabricated a self-powered visible-light photodetector that operates at room temperature and zero bias, built from an inhomogeneous Ag metasurface capped by a 30-µm PEDOT:PSS layer with Ti/Pd electrodes. The device shows broadband response in visible region with responsivity up to 1.45×10³ V W⁻¹, NEP down to 8.57 × 10 − 16 W Hz 1/2 , specific detectivity up to 5.71×10 14 Jones. Rise and fall time are 3 ms and 5 ms. The process uses scalable, cost-effective steps and maintains performance over six months under ambient exposure, indicating manufacturability and stability. Mechanistically, FDTD mapping and experiments show that inhomogeneous metasurface enabled LSPR hot spots in the metasurface both enhance responsivity and imprint wavelength-dependent waveform signatures in the PTE output. Coupling the detector with a lightweight 1D-CNN leverage these signatures. 10-fold cross-validation achieves a mean accuracy of 98.91%, with a near-perfect diagonal confusion matrix. A raster-scan validation on a laptop screen further confirms robust RGB recognition. Together, these results establish a metasurface–CNN system for filter-free color sensing and compact spectral identification, offering a potential path toward low-cost, filter-free imaging and portable spectrometer. With refined metasurface statistics, faster readout electronics, and larger training corpora, we anticipate further gains in detectivity, temporal resolution, and multi-class discrimination. The demonstrated combination of scalable fabrication, high optical response, and accurate recognition positions this platform as a promising building block for embedded optoelectronic sensing and advanced visible-light communication applications. Materials and Methods Device fabrication and Materials 4-inch, (100)-oriented Si wafers were used. The diced silicon wafers with a size of 10 mm × 10 mm were successively heated and washed in a water bath with ammonia water, hydrogen peroxide and deionized water for 15 minutes. Then, they were dried with nitrogen gas. A 500 nm SiO₂ layer was deposited by plasma enhanced chemical vapor deposition. A 20 nm Ag film was then deposited onto the SiO₂ by electron-beam physical vapor deposition and transformed into an inhomogeneous metasurface via rapid thermal annealing (500°C, 5 min). Next, 100 µL of PEDOT:PSS solution was dispensed and spin-coated at 3000 rpm for 30 s to form the polymer layer on the silicon wafer and we left the PEDOT:PSS to stand for a period of time to ensure complete drying. Finally, Ti and Pd electrodes were sequentially deposited on the PEDOT:PSS by electron-beam physical vapor deposition with shadow masks. PEDOT:PSS solution (product number: 483095, 1.3 wt%, PEDOT:PSS = 5:8) was purchased from Sigma-Aldrich®. Structure and characterization of PEDOT:PSS are shown in Fig. S1. Conductive silver glue and copper wires were purchased from Digi-Key. Measurements A scanning electron microscope (JEOL JSM-7200F SEM) at 2.0 kV voltage was used to characterize the detector. All electrical and optoelectronic measurements were performed at room temperature in a dark room to eliminate ambient light. Samples were measured unencapsulated (directly exposed to air). Illumination was provided by three compact LED sources (ELEGOO Upgraded 37 in 1 Sensor Modules Kit with Tutorial Compatible with Arduino IDE UNO R3 MEGA Nano, Amazon-sourced - https://a.co/d/33fz0am ), each operated at around 1.84 mW, driven by an Arduino with a series resistor, emitting at 620 nm (red), 520 nm (green), and 460 nm (blue), and the optical absorbing area was 24 mm². Unless otherwise noted, the detector was positioned at a fixed source–detector distance of 20 cm. Signals were recorded using a Keithley 6500 digital multimeter, while rise/fall times were measured with an Agilent Technologies DSO7104B oscilloscope since Keithey 6500 max sampling rate is around 16.67 ms. Declarations Funding: The authors acknowledge the support from the Natural Sciences and Engineering Research Council of Canada (NSERC) and the Micro-Nano Technology (MNT) program facilitated by CMC Microsystems. We also thank Auroray Labs Inc. for valuable financial contributions, which have significantly advanced the technological development of this work. Author contributions: Conceptualization: WG Methodology: WG, JW, GL, YY Investigation: WG Visualization: WG Supervision: JTWY Writing—original draft: WG Writing—review & editing: WG, JW, GL, JTWY Competing interests: Authors declare that they have no competing interests. Data and materials availability: All data needed to evaluate the conclusions are available in the main text or the supplementary materials. References P. X. Feng, E. Pacheco Cabrera, J. Chu, B. Zhou, S. Y. Flores, X. Peng, Y. Li, L. M. Diaz-Vazquez, A. F. Zhou, Self-Powered Photodetectors with Ultra-Broad Spectral Response and Thermal Stability for Broadband, Energy Efficient Wearable Sensing and Optoelectronics. Molecules 30 (2025). Y. Lin, Z. Yong, X. Luo, S. S. Azadeh, J. C. Mikkelsen, A. Sharma, H. Chen, J. C. C. Mak, P. G. Q. Lo, W. D. Sacher, J. K. S. Poon, Monolithically integrated, broadband, high-efficiency silicon nitride-on-silicon waveguide photodetectors in a visible-light integrated photonics platform. Nat Commun 13 (2022). W. Qarony, A. S. Mayet, E. Ponizovskaya Devine, S. Ghandiparsi, C. Bartolo-Perez, A. Ahamed, A. Rawat, H. H. Mamtaz, T. Yamada, S.-Y. Wang, M. Saif Islam, “Achieving higher photoabsorption than group III-V semiconductors in silicon using photon-trapping surface structures.” J. K. S. Poon, A. Govdeli, A. Sharma, X. Mu, F.-D. Chen, T. Xue, T. Liu, Silicon photonics for the visible and near-infrared spectrum. Adv Opt Photonics 16, 1 (2024). J. Ho, Z. Dong, H. Sheng Leong, J. Zhang, F. Tjiptoharsono, S. Daqiqeh Rezaei, K. Choon Hwa Goh, M. Wu, S. Li, J. Chee, C. Pei Yu Wong, A. I. Kuznetsov, J. K. W Yang, “Miniaturizing color-sensitive photodetectors via hybrid nanoantennas toward submicrometer dimensions” (2022); https://www.science.org . L. Chen, K. Preston, S. Manipatruni, M. Lipson, H. Chen, J. Zhang, N. A. Nelson, D. H. Albonesi, E. G. Friedman, P. M. Fauchet, C. Batten, A. Joshi, J. Orcutt, A. Khilo, B. Moss, C. Holzwarth, M. Popovic, H. Li, H. Smith, J. Hoyt, F. Kaertner, R. Ram, V. Stojanovic, K. Asanovic, H. Rong, R. Jones, A. Liu, O. Cohen, D. Hak, A. Fang, M. Paniccia, W. Fang, H. Park, M. J. Paniccia, J. E. Bowers, “Photonic networks-on-chip for future generations of chip multiprocessors” (2006); https://doi.org/10.1364/OA_License_v1#VOR . D. Zhang, C. Fuentes-Hernandez, R. Vijayan, Y. Zhang, Y. Li, J. W. Park, Y. Wang, Y. Zhao, N. Arora, A. Mirzazadeh, Y. Do, T. Cheng, S. Swaminathan, T. Starner, T. L. Andrew, G. D. Abowd, Flexible computational photodetectors for self-powered activity sensing. npj Flexible Electronics 6 (2022). M. Wang, M. Zhuang, L. Ge, Y. Yuan, Z. Wang, W. Liu, H. Li, A. Song, Tunable photoresponse properties of CuI/Si self-powered photodetectors through Zn doping engineering. Appl Surf Sci 654 (2024). J. Wang, Z. Xie, J. A. Liu, R. Zhou, G. Lu, J. T. W. Yeow, System design of large-area vertical photothermoelectric detectors based on carbon nanotube forests with MXene electrodes. Nanoscale Adv 5, 1133–1140 (2023). J. Wang, Z. Xie, J. A. Liu, G. Lu, R. Zhou, J. T. W. Yeow, Capillary-Assisted Self-Assembly of Carbon Nanotubes for the Self-Powered Photothermoelectric Detector. Adv Mater Technol 8 (2023). G. Lu, R. Zhou, J. Wang, Z. Xie, Y. Yuan, J. T. W. Yeow, “Flexible Graphene/PEDOT: PSS Free-Standing Infrared Photodetector” in Proceedings of the IEEE Conference on Nanotechnology (IEEE Computer Society, 2024), pp. 180–184. J. Wang, Z. Xie, J. T. W. Yeow, Two-dimensional materials applied for room-temperature thermoelectric photodetectors. Mater Res Express 7 (2020). Z. Xie, J. Wang, G. Lu, J. T. W. Yeow, Room-ambient operation of integrated and visualized photothermoelectric system with patterned Mo2C/PEDOT: PSS flexible devices. Mater Des 235 (2023). J. Wang, Z. Xie, J. A. Liu, J. T. W. Yeow, Design of room-temperature infrared photothermoelectric detectors based on CNT/PEDOT:PSS composites. J Mater Chem C Mater , doi: 10.1039/d2tc03159k (2022). J. Wang, Z. Xie, G. Lu, J. A. Liu, J. T. W. Yeow, An infrared photothermoelectric detector enabled by MXene and PEDOT:PSS composite for noncontact fingertip tracking. Microsyst Nanoeng 9 (2023). Y. Lu, L. Liu, R. Gao, Y. Xiong, P. Sun, Z. Wu, K. Wu, T. Yu, K. Zhang, C. Zhang, T. Bourouina, X. Li, X. Liu, Ultrafast near-infrared pyroelectric detector based on inhomogeneous plasmonic metasurface. Light Sci Appl 13 (2024). G. Lu, J. Wang, W. Gao, Z. Xie, R. Zhou, J. T. W. Yeow, Flexible Free-Standing Infrared Photothermoelectric Detector Based on Graphene/PEDOT. IEEE Nanotechnol Mag , doi: 10.1109/MNANO.2025.3570886 (2025). G. Lu, J. Wang, W. Gao, J. T. W. Yeow, “Flexible Infrared Photothermoelectric Detectors Based on Graphene/Metallic Junctions with Lowcost Substrates” in 2025 IEEE 25th International Conference on Nanotechnology (NANO) (IEEE, 2025; https://ieeexplore.ieee.org/document/11113664/ ), pp. 124–127. S. Shekhar, W. Bogaerts, L. Chrostowski, J. E. Bowers, M. Hochberg, R. Soref, B. J. Shastri, Roadmapping the next generation of silicon photonics. Nat Commun 15 (2024). A. Li, C. Yao, J. Xia, H. Wang, Q. Cheng, R. Penty, Y. Fainman, S. Pan, Advances in cost-effective integrated spectrometers. Springer Nature [Preprint] (2022). https://doi.org/10.1038/s41377-022-00853-1 . S. Bin Choi, J. S. Choi, H. S. Shin, J. W. Yoon, Y. Kim, J. W. Kim, Deep learning-developed multi-light source discrimination capability of stretchable capacitive photodetector. npj Flexible Electronics 9 (2025). A. D. Rakic´, R. Rakic´, A. B. Djuriš, J. M. Elazar, M. L. Majewski, “Optical properties of metallic films for vertical-cavity optoelectronic devices” (1998); https://doi.org/10.1364/OA_License_v1#VOR . C. Y. Ho; R. W. Powell; P. E. Liley, Thermal Conductivity of the Elements. J. Phys. Chem. Ref. Data 1, 279–421 (1972). I. T. Jollife, J. Cadima, Principal component analysis: A review and recent developments. Royal Society of London [Preprint] (2016). https://doi.org/10.1098/rsta.2015.0202 . L. Van Der Maaten, G. Hinton, “Visualizing Data using t-SNE” (2008). Additional Declarations There is NO Competing Interest. Supplementary Files SupplementaryMaterials.pdf Cite Share Download PDF Status: Under Review Version 1 posted You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-8006942","acceptedTermsAndConditions":true,"allowDirectSubmit":false,"archivedVersions":[],"articleType":"Article","associatedPublications":[],"authors":[{"id":543732963,"identity":"be01bc50-5c91-4f5f-ba37-6f857b03f5a1","order_by":0,"name":"John Yeow","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAA0klEQVRIiWNgGAWjYBACxgYQWQHlPSBeyxkgZgPiBOKtaiNFC/OM5IOPeecdTuyf33zsQ0KFDQN/+wECFsxISzbm3XY4ccYxtuQZCWfSGCTOELCLcUaOmXTutsO5Dcd4jBkS2w4zGBByHuOM/G/SuXMO584/xv8ZooX/AUFb2KRzGw7nbjjGwwzRIkHIlp5nxsZ/jqXXbzyWZswA9AuPxA0Cthi2Jz98OKPG2lju8OHHDB8qbOT4+wnYYjgBTQEPfvVAIM9/gKCaUTAKRsEoGOkAAAjmRTHIAZ5tAAAAAElFTkSuQmCC","orcid":"https://orcid.org/0000-0002-0740-8055","institution":"University of Waterloo","correspondingAuthor":true,"prefix":"","firstName":"John","middleName":"","lastName":"Yeow","suffix":""},{"id":543732964,"identity":"d447f0ff-d0f6-4b51-b656-68e59844af93","order_by":1,"name":"Wentao Gao","email":"","orcid":"","institution":"University of Waterloo","correspondingAuthor":false,"prefix":"","firstName":"Wentao","middleName":"","lastName":"Gao","suffix":""},{"id":543732965,"identity":"2d3abefa-58a1-49d5-9f68-b79ef57c813f","order_by":2,"name":"Jiaqi Wang","email":"","orcid":"","institution":"University of Waterloo","correspondingAuthor":false,"prefix":"","firstName":"Jiaqi","middleName":"","lastName":"Wang","suffix":""},{"id":543732966,"identity":"d63d4ac4-0904-417d-b3fe-676a676a578f","order_by":3,"name":"guanxuan Lu","email":"","orcid":"","institution":"University of Waterloo","correspondingAuthor":false,"prefix":"","firstName":"guanxuan","middleName":"","lastName":"Lu","suffix":""},{"id":543732967,"identity":"6b23d93e-12b5-4f02-8af0-1edd0cc6233b","order_by":4,"name":"Yifei Yuan","email":"","orcid":"","institution":"University of Waterloo","correspondingAuthor":false,"prefix":"","firstName":"Yifei","middleName":"","lastName":"Yuan","suffix":""}],"badges":[],"createdAt":"2025-11-01 16:25:07","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-8006942/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-8006942/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":101880550,"identity":"3183acde-08f1-4318-b678-886acd996b04","added_by":"auto","created_at":"2026-02-04 15:03:32","extension":"jpg","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":78888,"visible":true,"origin":"","legend":"\u003cp\u003eDevice structure and metasurface\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e(A)\u003c/strong\u003e Diagram of photodetector. \u003cstrong\u003e(B)\u003c/strong\u003e SEM image of the inhomogeneous metasurface, with particle sizes spanning from about 20 nm up to 375 nm. \u003cstrong\u003e(C)\u003c/strong\u003e Schematic of the simulated electric-field distribution at the interface between the inhomogeneous metasurface and the PEDOT:PSS layer obtained by FDTD. The incident wavelength ranges from 390 to 760 nm, and the maximum-to-minimum field intensity ratio is approximately 4000. \u003cstrong\u003e(D)\u003c/strong\u003e A magnified view of the simulation.\u003c/p\u003e","description":"","filename":"Picture1.jpg","url":"https://assets-eu.researchsquare.com/files/rs-8006942/v1/f40a5d6000d83ace77d3133b.jpg"},{"id":101785649,"identity":"e64ec482-bc62-4dc6-8041-0abf2f6a936b","added_by":"auto","created_at":"2026-02-03 15:37:32","extension":"jpg","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":59586,"visible":true,"origin":"","legend":"\u003cp\u003eDiagram of PTE mechanism and photoresponse\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e(A)\u003c/strong\u003e Schematic of the photothermoelectric (PTE) mechanism: a lateral temperature gradient drives carriers via the Seebeck effect, with electrons drifting toward the colder Pd contact and holes toward the hotter Ti contact.\u003cstrong\u003e(B–D) \u003c/strong\u003eZero-bias Vout responses recorded at room temperature in the dark under monochromatic illumination: \u003cstrong\u003e(B)\u003c/strong\u003e 460 nm (blue), \u003cstrong\u003e(C)\u003c/strong\u003e 520 nm (green), and \u003cstrong\u003e(D)\u003c/strong\u003e 620 nm (red). Signals were measured with a digital multimeter.\u003c/p\u003e","description":"","filename":"Picture2.jpg","url":"https://assets-eu.researchsquare.com/files/rs-8006942/v1/01dd35c735363ae806f18635.jpg"},{"id":101785652,"identity":"adecf947-4f20-492a-a2b6-137772ee7837","added_by":"auto","created_at":"2026-02-03 15:37:32","extension":"jpg","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":117668,"visible":true,"origin":"","legend":"\u003cp\u003eConventional analysis and schematic diagram of CNN structure\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e(A)\u003c/strong\u003e PCA projection of the dataset. As a linear dimensional-reduction technique, PCA identifies orthogonal components that maximize the variance explained in the original data. \u003cstrong\u003e(B)\u003c/strong\u003e t-SNE projection. This non-linear embedding emphasizes preservation of local neighborhood similarity, revealing cluster structure not captured by linear projections. \u003cstrong\u003e(C)\u003c/strong\u003e Voltage-time responses recorded with a digital multimeter under monochromatic illumination—blue (460 nm), green (520 nm), and red (620 nm)—at multiple source–detector distances.\u003cstrong\u003e (D)\u003c/strong\u003eWorkflow and model schematic showing how the signals are fed to a 1D convolutional neural network (1D-CNN) for classification; the architecture comprises two 1D convolutional layers, one max-pooling layer, and fully connected layers.\u003c/p\u003e","description":"","filename":"Picture3.jpg","url":"https://assets-eu.researchsquare.com/files/rs-8006942/v1/d45199f342b7f8dde34aadc7.jpg"},{"id":101785650,"identity":"e49d4ad1-43da-4b27-ba6b-56b41c94052a","added_by":"auto","created_at":"2026-02-03 15:37:32","extension":"jpg","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":57826,"visible":true,"origin":"","legend":"\u003cp\u003eAssessment of the feasibility of CNN for color and distance classification.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e(A)\u003c/strong\u003e Validation accuracy obtained across 10 folds, with the mean accuracy indicated by the dashed line (0.9891). \u003cstrong\u003e(B)\u003c/strong\u003e Training history for the first three folds, showing validation accuracy convergence over epochs.\u003cstrong\u003e (C)\u003c/strong\u003e Average training and validation loss curves with confidence intervals across all folds.\u003cstrong\u003e (D)\u003c/strong\u003eConfusion matrix illustrating classification performance for each color–distance class, expressed in sample counts.\u003cstrong\u003e (E)\u003c/strong\u003e Distribution of accuracy across folds, with mean and median values indicated by dashed lines. \u003cstrong\u003e(F)\u003c/strong\u003eClass-wise performance metrics (F1-score, precision, recall) demonstrate consistently high values across all categories.\u003c/p\u003e","description":"","filename":"Picture4.jpg","url":"https://assets-eu.researchsquare.com/files/rs-8006942/v1/4d8788b3efafac49def9a422.jpg"},{"id":101785654,"identity":"f8bec01d-7e6c-4ed3-b1bd-c33ff682bf55","added_by":"auto","created_at":"2026-02-03 15:37:32","extension":"jpg","order_by":5,"title":"Figure 5","display":"","copyAsset":false,"role":"figure","size":89517,"visible":true,"origin":"","legend":"\u003cp\u003eScreen color recognition and prospects for colorful imaging\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e(A)\u003c/strong\u003e A digital image illustrating the experimental setup designed to assess the cognitive ability in recognizing the color emitted from laptop screen. A computer screen displays three adjacent color bands, the three primary colors of RGB, which are sequentially scanned by the detector. The resulting electrical signals are recorded using a Keithley 6500 multimeter. \u003cstrong\u003e(B)\u003c/strong\u003e Schematic illustration of the experimental setup. \u003cstrong\u003e(C)\u003c/strong\u003e Comparison between the original displayed pattern (top) and the classification results obtained from the detector (bottom). Confidence levels for each predicted color are indicated below. Small discrepancies occur at the color boundaries, where classification errors are more likely to arise.\u003c/p\u003e","description":"","filename":"Picture5.jpg","url":"https://assets-eu.researchsquare.com/files/rs-8006942/v1/93982643491312feb4727428.jpg"},{"id":101785651,"identity":"fbd4e7ca-3740-4b3a-aa0a-3531dd820415","added_by":"auto","created_at":"2026-02-03 15:37:32","extension":"jpg","order_by":6,"title":"Figure 6","display":"","copyAsset":false,"role":"figure","size":46367,"visible":true,"origin":"","legend":"\u003cp\u003eWearable metasurface detector for traffic-signal recognition.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e(A) \u003c/strong\u003eSchematic diagram: a green LED luminaire simulates a traffic signal while the user faces the source. The inhomogeneous metasurface photothermoelectric detector (mounted on glasses) acquires the incident light and a lightweight CNN to process the signal, realizing the spectral recognition. \u003cstrong\u003e(B) \u003c/strong\u003eLaboratory demonstration of traffic-signal recognition.\u003cstrong\u003e (C)\u003c/strong\u003e Close-up of the glasses-mounted module. The detector is laminated on the outer surface of the eyewear, enabling a compact devicethat outputs signatures for downstream CNN classification and user cues.\u003c/p\u003e","description":"","filename":"Picture6.jpg","url":"https://assets-eu.researchsquare.com/files/rs-8006942/v1/43821dc77c05fb998fc133dc.jpg"},{"id":101882192,"identity":"a5067f10-3356-47bd-a799-859c22eb69cb","added_by":"auto","created_at":"2026-02-04 15:21:44","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":1062804,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-8006942/v1/c88f6739-7078-4cea-9acd-bf8977764e85.pdf"},{"id":101880870,"identity":"423d6e38-651f-4acf-ad69-a9b0bd1fa638","added_by":"auto","created_at":"2026-02-04 15:07:16","extension":"pdf","order_by":1,"title":"","display":"","copyAsset":false,"role":"supplement","size":848395,"visible":true,"origin":"","legend":"","description":"","filename":"SupplementaryMaterials.pdf","url":"https://assets-eu.researchsquare.com/files/rs-8006942/v1/f174d98a9bc91508b1afc6cc.pdf"}],"financialInterests":"There is \u003cb\u003eNO\u003c/b\u003e Competing Interest.","formattedTitle":"Inhomogeneous Metasurface–CNN System for Filter-Free Single-Pixel Wavelength Recognition","fulltext":[{"header":"Introduction","content":"\u003cp\u003e \u003cdiv class=\"BlockQuote\"\u003e \u003cp\u003eDetecting visible light across a broad spectral range is central to applications such as high-resolution imaging, environmental monitoring, wearable electronics, and energy-efficient optoelectronic platforms (\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e). Established visible photodetectors, such as silicon and organic photodiodes, CMOS image sensors, and emerging heterojunction/2D devices, already offer high signal-to-noise ratios, fast response and wide dynamic range, reliably meeting needs in imaging, machine vision, proximity sensing, and environmental monitoring (\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e, \u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e). However, most visible photodetectors are broadband intensity devices without intrinsic spectral selectivity and therefore rely on filter arrays or dispersive optics (\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e, \u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e). Furthermore, conventional silicon-based photodetectors often require external bias to achieve efficient operation, which increases power demand and complicates system integration (\u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e).\u003c/p\u003e \u003c/div\u003e \u003c/p\u003e \u003cp\u003eThese limitations are increasingly prominent for emerging applications, including self-powered sensing, portable spectroscopic devices, and flexible optoelectronics, where high responsivity, low power consumption, and wavelength tunability are critical (\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e, \u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e). Conductive polymers such as poly(3,4-ethylenedioxythiophene) polystyrene sulfonate (PEDOT:PSS) exhibit strong and broadband absorption across the visible spectrum, in addition to mechanical flexibility and compatibility with solution processing, which support scalable, low-cost fabrication (\u003cspan additionalcitationids=\"CR10 CR11 CR12 CR13 CR14\" citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e). Photothermoelectric (PTE) detectors convert light to voltage via the Seebeck effect: optical absorption locally heats the active region and establishes a temperature gradient, which drives an open circuit photovoltage. Compared with other mechanisms, PTE detectors offer self-powered operation with low dark current and shot noise, room-temperature use, broad spectral operation, and relatively fast response. Inhomogeneous metasurfaces can generate localized surface plasmon resonances (LSPRs), enhancing light absorption and accelerating the photothermal conversion process (\u003cspan additionalcitationids=\"CR17\" citationid=\"CR16\" class=\"CitationRef\"\u003e16\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e18\u003c/span\u003e). This LSPR-driven photothermal not only gain strengthens the photovoltage but also provides the necessary features for spectral selectivity without optical filters. Convolutional neural network (CNN) learnt translation-equivariant features by convolving shared kernels across the input using far fewer parameters than fully connected models. In our voltage\u0026ndash;time model, temporal convolutions capture edges, plateaus, and PTE decays, while pooling and striding with normalization improves robustness. Trained end-to-end, lightweight 1D CNNs reduce reliance on handcrafted features and reliably decode wavelength-inhomogeneous metasurface wavelength-dependent signatures. These attributes make PTE-inhomogeneous metasurface-CNN platforms well suited to portable, filter-free color imaging and compact spectral sensing.\u003c/p\u003e \u003cp\u003eIn this study, we investigated a hybrid visible PTE detector consisting of PEDOT:PSS films integrated with inhomogeneous metasurface on silicon substrates. Our devices show an outstanding performance at room temperature and self-powered, with responsivity up to 1.45\u0026times;10\u0026sup3; V W⁻\u0026sup1;, noise-equivalent power (NEP) down to 5.54\u0026times;10\u003csup\u003e-11\u003c/sup\u003eW Hz\u003csup\u003e-1/2\u003c/sup\u003e, and specific detectivity up to 4.01\u0026times;10\u003csup\u003e10\u003c/sup\u003e Jones. Silicon-based devices present notable advantages, including ease of fabrication, compatibility with established chip integration technologies, potential for device miniaturization, and inherently reduced noise (\u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e19\u003c/span\u003e, \u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e). The device structure is designed to recognize wavelength in a cost-effective way, delivering broadband sensitivity, self-powered operation, and scalable manufacturability. A lightweight 1D-CNN trained on a wavelength-distance dataset attains 98.91% average accuracy under 10-fold cross-validation and demonstrates strong real-world application through raster-scanning on a laptop display and traffic-light recognition. Leveraging the wavelength-dependent response characteristics introduced by the inhomogeneous metasurface and integrating them with a CNN model, we demonstrate the capability to recognize light of different wavelengths and source-to-detector distances. This approach can be further extended toward filter-free colorful imaging and the development of compact, portable spectrometers. By systematically correlating the morphological characteristics of the metasurface with the photoelectric response of the detector, and further employing CNN to distinguish these spectral features, this work provides new insights into plasmon-assisted PTE photodetectors and photodetector-CNN systems, which establishes a pathway toward filter-free wavelength recognition and colorful imaging(\u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e21\u003c/span\u003e).\u003c/p\u003e"},{"header":"Results","content":"\u003cp\u003eDevice structure and metasurface\u003c/p\u003e \u003cp\u003eTo realize broadband visible absorption, we engineered the device comprising an inhomogeneous silver (Ag) metasurface capped by a 2\u0026micro;m thick PEDOT:PSS film (Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003eA). A 20-nm Ag layer was first deposited on SiO₂ and transformed into an inhomogeneous metasurface by thermal annealing. A PEDOT:PSS film was then formed by spin-coating, and Ti/Pd electrodes were sequentially deposited. The cross-sectional SEM imaging of our detector is shown in Fig. S2. Full fabrication details are provided in Materials and Methods. A representative scanning electron micrograph of the metasurface is shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003eB.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eUsing finite-difference time-domain (FDTD) simulations, Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e(C and D) maps the electric-field intensity at the interface between the inhomogeneous Ag metasurface and the PEDOT:PSS layer under 390\u0026ndash;760 nm illumination. Inhomogeneous metasurface induced wavelength-dependent resonance across different islands, which modulate near-field intensity and local absorption and induce asymmetries in electron-hole transport. Because LSPRs are strongly geometry-dependent (\u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e16\u003c/span\u003e), the inhomogeneous metasurface generates spectrally selective and spatially distinct hot spots whose locations shift with wavelength. Moreover, wavelength-dependent variations in island multiplicity imprint distinguishable response signatures within the same PTE detection mechanism, which are exploited for CNN-assisted recognition later. The inhomogeneous metasurface enhances optical absorption and yields a stronger device response, while embedding distinct, recognizable waveform signatures in the output signal.\u003c/p\u003e \u003cdiv id=\"Sec3\" class=\"Section2\"\u003e \u003ch2\u003eMechanism and Optical properties\u003c/h2\u003e \u003cp\u003eAs shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e2\u003c/span\u003eA, under identical visible illumination, Ti exhibits strong optical absorption, typically converting about 50% of incident energy into heat, whereas Pd is more reflective in the same visible wavelength, with lower absorption, yielding weaker photothermal conversion (\u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e22\u003c/span\u003e, \u003cspan citationid=\"CR23\" class=\"CitationRef\"\u003e23\u003c/span\u003e). Thermal transport further separates their behavior: Ti\u0026rsquo;s thermal conductivity (21 W m⁻\u0026sup1; K⁻\u0026sup1;) is much lower than Pd\u0026rsquo;s (71 W m⁻\u0026sup1; K⁻\u0026sup1;), so heat generated in Ti spreads slowly and produces a higher local temperature, while Pd dissipates heat efficiently (\u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e22\u003c/span\u003e, \u003cspan citationid=\"CR23\" class=\"CitationRef\"\u003e23\u003c/span\u003e). The resulting temperature gradient between Ti and Pd drives the photothermoelectric (PTE) effect. At room temperature, Ti has a negative Seebeck coefficient (S\u003csub\u003eTi\u003c/sub\u003e\u0026thinsp;\u0026asymp;\u0026thinsp;\u0026minus;\u0026thinsp;3 to \u0026minus;\u0026thinsp;7 \u0026micro;V K⁻\u0026sup1;), consistent with electrons diffusing from hot to cold; Pd has a positive Seebeck coefficient (S\u003csub\u003ePd\u003c/sub\u003e\u0026thinsp;\u0026asymp;\u0026thinsp;+\u0026thinsp;7 to +\u0026thinsp;10 \u0026micro;V K⁻\u0026sup1;), corresponding to hole flow from hot to cold. Because the coefficients differ, the net thermopower ΔS\u0026thinsp;\u0026asymp;\u0026thinsp;S\u003csub\u003ePd\u003c/sub\u003e \u0026minus; S\u003csub\u003eTi\u003c/sub\u003e \u0026asymp; 15 \u0026micro;V K⁻\u0026sup1;, giving an open-circuit voltage V\u003csub\u003eOC\u003c/sub\u003e = (S\u003csub\u003ePd\u003c/sub\u003e \u0026minus; S\u003csub\u003eTi\u003c/sub\u003e) ΔT. Consequently, a Ti\u0026ndash;Pd pair delivers a stable, appreciable PTE signal under illumination (\u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e, \u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e18\u003c/span\u003e).\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eThe device operates in a self-powered mode, delivering strong performance at zero bias and room temperature. Figure\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e2\u003c/span\u003e (B to D) shows the device responses at different visible wavelengths. The illumination was provided by simple LED sources, driven by an Arduino with a series resistor. Emission peaks were 620 nm (red), 520 nm (green), and 460 nm (blue). All measurements were performed at room temperature and under zero bias. To mitigate thermal effects from the sources, the detector was positioned 20 cm from the sources. We repeated the acquisitions multiple times and observed no obvious decline after six months of ambient exposure. The photoresponse was recorded using a multimeter in a dark room to minimize environmental interference. Responsivity is defined as R\u003csub\u003ev\u003c/sub\u003e=\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:\\frac{V}{{P}_{in}}\\)\u003c/span\u003e\u003c/span\u003e, where V is recorded average voltage under illumination and P is effective incident light power. NEP is defined by \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:\\frac{\\sqrt{2q{V}_{dark}}}{{R}_{v}}\\)\u003c/span\u003e\u003c/span\u003e, where q is elementary charge and V\u003csub\u003edark\u003c/sub\u003e is recorded average voltage in the dark environment. Specific detectivity D* is defined as D\u003csup\u003e*\u003c/sup\u003e= \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:\\frac{\\sqrt{A}\\:\\varDelta\\:f}{NEP}\\)\u003c/span\u003e\u003c/span\u003e, where A\u003csub\u003eD\u003c/sub\u003e is the active area of detector and Δf is the bandwidth. Under blue, green, and red illumination, respectively, the responsivity (R\u003csub\u003ev\u003c/sub\u003e) reached 9.73 \u0026times; 10\u003csup\u003e2\u003c/sup\u003e, 1.38 \u0026times; 10\u003csup\u003e3\u003c/sup\u003e, and 1.45 \u0026times; 10\u003csup\u003e3\u003c/sup\u003e V W\u003csup\u003e\u0026minus;1\u003c/sup\u003e; The NEP reached 8.57 \u0026times; 10\u003csup\u003e\u0026minus;16\u003c/sup\u003e, 9.88 \u0026times; 10\u003csup\u003e\u0026minus;\u0026thinsp;16\u003c/sup\u003e, and 9.09 \u0026times; 10\u003csup\u003e\u0026minus;\u0026thinsp;16\u003c/sup\u003e W Hz\u003csup\u003e\u0026minus;1/2\u003c/sup\u003e; The specific detectivity (D\u003csup\u003e*\u003c/sup\u003e) reached 5.71 \u0026times; 10\u003csup\u003e14\u003c/sup\u003e, 4.96 \u0026times; 10\u003csup\u003e14\u003c/sup\u003e, and 5.39 \u0026times; 10\u003csup\u003e14\u003c/sup\u003e Jones. Rise time is defined as the time required for the output to increase from 10% to 90% of its final steady-state value after the illumination is turned on. Fall time is defined as the time required for the output to decrease from 90% to 10% of its final steady state value after the illumination is turned off. The rise and fall times are 3 ms and 5 ms, respectively. While optoelectronic measurement in this study relies on RGB primaries, the underlying mechanism of our inhomogeneous metasurface detector, LSPR\u0026rsquo;s wavelength dependence, supports broadband response and recognition.\u003c/p\u003e \u003cp\u003eIn addition, optoelectronic measurements provide distinguishing features for mechanism identification. The bolometric effect arises from photo-induced heating, which requires an external bias for efficient readout, and it is governed by thermal diffusion, yielding comparatively long-time response. Moreover, in conductive-polymer/metal composite structures, the photovoltaic effect is severely suppressed by short carrier lifetimes, interfacial defects, and metal-induced quenching, so that any photocarriers generated recombine rapidly. Based on our device structure and measurement, the detector operates at zero bias and exhibits a measured millisecond-scale response, aligning with a non-bias photothermoelectric effect, while the bias-dependent bolometric effect is negligible for the response mechanism. Moreover, the device lacks an effective p\u0026ndash;n junction or a strong Schottky depletion region and built-in field, and the spectral response follows the localized surface plasmon resonance (LSPR) rather than a semiconductor band gap absorption. Even if the photovoltaic effect occurs, photocarriers are easy to rapidly recombine at metal/polymer interfaces, making it difficult to achieve a significant PV output. Taking together, these results indicate that LSPR-induced, temperature-gradient-driven PTE is the dominant mechanism.\u003c/p\u003e \u003cp\u003e \u003cb\u003eConvolutional neural network\u0026ndash; inhomogeneous metasurface assisted wavelength\u0026ndash;distance recognition.\u003c/b\u003e \u003c/p\u003e \u003cp\u003eWe acquired training and test data indoors under bright ambient lighting using multimeter and LED sources at 620 nm (red), 520 nm (green), and 460 nm (blue). The CNN model is trained to focus on the steady-state waveform morphology, enhancing applicability across diverse illumination scenarios. Integrating with inhomogeneous metasurface, LSPR-enhanced absorption drives wavelength-dependent variations in both the photovoltage waveform and its kinetics. Thus, the voltage\u0026ndash;time traces recorded produce learnable features that a convolutional neural network (CNN) can exploit to distinguish classes. Each wavelength generates a time-domain signature with unique, repeatable traits during PTE conversion.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eIn Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e3\u003c/span\u003e (A and B), we visualize the visible-light dataset with principal component analysis (PCA) and t-distributed stochastic neighbor embedding (t-SNE) (\u003cspan citationid=\"CR24\" class=\"CitationRef\"\u003e24\u003c/span\u003e, \u003cspan citationid=\"CR25\" class=\"CitationRef\"\u003e25\u003c/span\u003e). PCA, a linear method that maximizes explained variance, reveals some inter-class separation in 2D but also substantial overlap, especially for same wavelength samples at different distances, which cluster nearby and stretch mainly along high-variance directions. By contrast, t-SNE, a nonlinear embedding that preserves local neighborhoods, better exposes the latent manifold, mapping classes into curved, arc-like clusters with tighter intra-class clouds and clearer inter-class spacing. These patterns indicate that conventional approaches struggle to reliably resolve wavelength in this setting.\u003c/p\u003e \u003cp\u003eTo improve recognition, we adopt deep learning as an end-to-end solution that learns discriminative features directly from data. This approach captures complex, non-linear patterns that traditional feature engineering misses, yielding higher accuracy and better robustness to environmental variation. We collected 1836 samples and applied data augmentation at the outset of training. Voltage\u0026ndash;time traces were recorded by a multimeter under three LEDs (460, 520, 620 nm) at 20, 23, 28, 40 cm. After thorough research, CNN, a deep model that learns discriminative features by convolving shared filters, was selected for wavelength recognition. In our CNN model, temporal convolutions capture edges and plateaus, while pooling and normalization provide robustness to amplitude fluctuations and timing jitter. Trained end-to-end with backpropagation, CNNs reduce reliance on hand-crafted features and are widely adopted for signal classification and regression in sensing. These properties make lightweight 1D-CNNs well suited to our inhomogeneous metasurface-detector, where they decode wavelength-dependent waveform signatures into compact spectral labels. Coupling the CNN with the inhomogeneous metasurface enables richer analysis and prediction, enhancing practical utility and highlighting potential for advanced optical sensing.\u003c/p\u003e \u003cp\u003eFigure\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e3\u003c/span\u003eD outlines the workflow and our 1D-CNN architecture. We first unify sequence length and apply per-sample z-score normalization to preprocessed data. Labels are parsed as wavelength\u0026ndash;distance composites and encoded into discrete IDs. The classifier comprises two convolutional layers and one pooling layer, optimized with adaptive moment estimation (Adam) (learning rate\u0026thinsp;=\u0026thinsp;1\u0026times;10⁻\u0026sup3;) and cross-entropy loss. The model has ~\u0026thinsp;3.18k trainable parameters, promoting generalization in small-data regimes. Training uses stratified 10-fold cross-validation with early stopping based on validation accuracy (patience\u0026thinsp;=\u0026thinsp;30 epochs). For each fold, we recorded confusion matrices and classification reports, then aggregate fold-level results and export visualizations and javascript object notation (JSON) summaries to support reproducibility and audit.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eConsidering the modest dataset size and the potential for train\u0026ndash;test fluctuations, we employed 10-fold cross-validation to validate accuracy and eliminate the random errors. In Fig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e4\u003c/span\u003eA, per-fold validation accuracies cluster tightly near 1.0, with a mean of 0.9891 and folds spanning roughly 0.985\u0026ndash;1.000, indicating low variance across partitions. The accuracy distribution (Fig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e4\u003c/span\u003eE) further shows both the mean and median concentrated around 0.989, reinforcing the stability of performance across folds. Training dynamics are summarized in Fig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e4\u003c/span\u003eB\u0026ndash;C. The first three folds\u0026rsquo; validation accuracy curves rise rapidly and plateau, while the average loss curve with confidence intervals decrease monotonically and then stabilize. The training and validation losses almost track one another throughout. This parallel behavior of accuracy and loss demonstrates that the model is learning wavelength-distance specific features without overfitting, consistent with our intended design and hyperparameter choices.\u003c/p\u003e \u003cp\u003eClass-level behavior across the 12 color\u0026ndash;distance classes (3 colors \u0026times; 4 distances) is shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e4\u003c/span\u003eD and Fig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e4\u003c/span\u003eF. The confusion matrix (Fig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e4\u003c/span\u003eD) is strongly diagonal, with only a few data off-diagonal that almost occur within the same color. Cross-color confusions are essentially absent. Correspondingly, per-class precision, recall, and F1-score (F) are uniformly high, evidencing balanced performance without bias toward particular categories. Taken together, these panels show that cross-validation yields an average accuracy of 98.91%. The training and validation curves in 10-fold exhibit marked similarity, indicating appropriate model selection, and that the classifier consistently identifies each wavelength\u0026ndash;distance combination. The model therefore avoids common traps, and demonstrates robustness, reliability, and generalizability for wavelength\u0026ndash;distance recognition.\u003c/p\u003e \u003c/div\u003e\n\u003ch3\u003eCNN Model Validation: color recognition and prospects for colorful imaging\u003c/h3\u003e\n\u003cp\u003e \u003c/p\u003e \u003cp\u003e \u003cb\u003e(A)\u003c/b\u003e A digital image illustrating the experimental setup designed to assess the cognitive ability in recognizing the color emitted from laptop screen. A computer screen displays three adjacent color bands, the three primary colors of RGB, which are sequentially scanned by the detector. The resulting electrical signals are recorded using a Keithley 6500 multimeter. \u003cb\u003e(B)\u003c/b\u003e Schematic illustration of the experimental setup. \u003cb\u003e(C)\u003c/b\u003e Comparison between the original displayed pattern (top) and the classification results obtained from the detector (bottom). Confidence levels for each predicted color are indicated below. Small discrepancies occur at the color boundaries, where classification errors are more likely to arise.\u003c/p\u003e \u003cp\u003eTo eliminate artifacts from compact LEDs, we used a laptop screen as the light source and raster-scanned the detector at a 3 cm standoff while a digital multimeter recorded the voltage traces. The results are shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig6\" class=\"InternalRef\"\u003e5\u003c/span\u003eC. In uniform single-color regions, the three primaries (R 620 nm, G 520 nm, B 460 nm) were classified correctly with posterior confidences\u0026thinsp;\u0026asymp;\u0026thinsp;1.0. Misclassifications appear only at color boundaries where spectra mix spatially. The width of misclassifications area and length of screen are 5 mm and 350mm, so the accuracy is about 97.14% for this figure. At the blue\u0026ndash;green boundary, portions of blue are labeled green with high confidence (~\u0026thinsp;1.0); at the green\u0026ndash;red boundary, portions of red are labeled green with intermediate confidence (~\u0026thinsp;0.5). This asymmetry is consistent with the measured responsivities: the detector\u0026rsquo;s blue responsivity is markedly lower than green, so the green component dominates the composite waveform near the B\u0026ndash;G border and the classifier confidently output green; by contrast, red and green responsivities are similar, yielding ambiguous evidence near the G\u0026ndash;R seam and correspondingly\u0026thinsp;~\u0026thinsp;0.5 confidence. Residual errors are therefore attributed to boundary mixing and pixel size rather than model bias.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003e \u003cb\u003e(A)\u003c/b\u003e Schematic diagram: a green LED luminaire simulates a traffic signal while the user faces the source. The inhomogeneous metasurface photothermoelectric detector (mounted on glasses) acquires the incident light and a lightweight CNN to process the signal, realizing the spectral recognition. \u003cb\u003e(B)\u003c/b\u003e Laboratory demonstration of traffic-signal recognition. \u003cb\u003e(C)\u003c/b\u003e Close-up of the glasses-mounted module. The detector is laminated on the outer surface of the eyewear, enabling a compact device that outputs signatures for downstream CNN classification and user cues.\u003c/p\u003e \u003cp\u003eWe prototyped a wearable implementation by combining the inhomogeneous metasurface detector onto the outer surface of safety glasses (Fig.\u0026nbsp;\u003cspan refid=\"Fig7\" class=\"InternalRef\"\u003e6\u003c/span\u003eB). Green and red LED were used to simulate the traffic signal (Fig.\u0026nbsp;\u003cspan refid=\"Fig7\" class=\"InternalRef\"\u003e6\u003c/span\u003eA); the glasses-mounted detector acquired the incident light and converted it into electrical signatures, which were classified by a lightweight CNN to determine the signal state. A simple tip was then issued to the wearer to indicate the detected state. This proof-of-concept shows that spectral recognition can be realized in a compact, wearable device under typical illumination.\u003c/p\u003e \u003cp\u003eThese results confirm that the photodetector-CNN system accurately recognizes the RGB primaries and that the waveform signatures imparted by the inhomogeneous Ag metasurface can be exploited beyond point classification. Because the signal arises from wavelength-dependent waveform morphology, the approach can be extended to filter-free colorful imaging by scanning or by array integration. In this system, the metasurface\u0026ndash;CNN system could reduce reliance on dispersive optics in certain regimes and enable portable, low-cost spectral identification, offering a path toward compact, intelligent spectrometers. The system\u0026rsquo;s high accuracy, stable ambient operation, and robustness to modest motion suggest reliable use in dynamic scenes, lowering the cost of visible-light imaging and spectral recognition. The combination of an inhomogeneous metasurface with a lightweight deep-learning classifier thus provides a practical route to embedded optoelectronic sensing and opens opportunities for advanced visible-light communication applications.\u003c/p\u003e"},{"header":"Discussion","content":"\u003cp\u003eIn conclusion, we designed and fabricated a self-powered visible-light photodetector that operates at room temperature and zero bias, built from an inhomogeneous Ag metasurface capped by a 30-\u0026micro;m PEDOT:PSS layer with Ti/Pd electrodes. The device shows broadband response in visible region with responsivity up to 1.45\u0026times;10\u0026sup3; V W⁻\u0026sup1;, NEP down to 8.57 \u0026times; 10\u003csup\u003e\u0026minus;\u0026thinsp;16\u003c/sup\u003e W Hz\u003csup\u003e1/2\u003c/sup\u003e, specific detectivity up to 5.71\u0026times;10\u003csup\u003e14\u003c/sup\u003e Jones. Rise and fall time are 3 ms and 5 ms. The process uses scalable, cost-effective steps and maintains performance over six months under ambient exposure, indicating manufacturability and stability.\u003c/p\u003e \u003cp\u003eMechanistically, FDTD mapping and experiments show that inhomogeneous metasurface enabled LSPR hot spots in the metasurface both enhance responsivity and imprint wavelength-dependent waveform signatures in the PTE output. Coupling the detector with a lightweight 1D-CNN leverage these signatures. 10-fold cross-validation achieves a mean accuracy of 98.91%, with a near-perfect diagonal confusion matrix. A raster-scan validation on a laptop screen further confirms robust RGB recognition.\u003c/p\u003e \u003cp\u003eTogether, these results establish a metasurface\u0026ndash;CNN system for filter-free color sensing and compact spectral identification, offering a potential path toward low-cost, filter-free imaging and portable spectrometer. With refined metasurface statistics, faster readout electronics, and larger training corpora, we anticipate further gains in detectivity, temporal resolution, and multi-class discrimination. The demonstrated combination of scalable fabrication, high optical response, and accurate recognition positions this platform as a promising building block for embedded optoelectronic sensing and advanced visible-light communication applications.\u003c/p\u003e"},{"header":"Materials and Methods","content":"\u003cdiv id=\"Sec7\" class=\"Section2\"\u003e \u003ch2\u003eDevice fabrication and Materials\u003c/h2\u003e \u003cp\u003e4-inch, (100)-oriented Si wafers were used. The diced silicon wafers with a size of 10 mm \u0026times; 10 mm were successively heated and washed in a water bath with ammonia water, hydrogen peroxide and deionized water for 15 minutes. Then, they were dried with nitrogen gas. A 500 nm SiO₂ layer was deposited by plasma enhanced chemical vapor deposition. A 20 nm Ag film was then deposited onto the SiO₂ by electron-beam physical vapor deposition and transformed into an inhomogeneous metasurface via rapid thermal annealing (500\u0026deg;C, 5 min). Next, 100 \u0026micro;L of PEDOT:PSS solution was dispensed and spin-coated at 3000 rpm for 30 s to form the polymer layer on the silicon wafer and we left the PEDOT:PSS to stand for a period of time to ensure complete drying. Finally, Ti and Pd electrodes were sequentially deposited on the PEDOT:PSS by electron-beam physical vapor deposition with shadow masks. PEDOT:PSS solution (product number: 483095, 1.3 wt%, PEDOT:PSS\u0026thinsp;=\u0026thinsp;5:8) was purchased from Sigma-Aldrich\u0026reg;. Structure and characterization of PEDOT:PSS are shown in Fig. S1. Conductive silver glue and copper wires were purchased from Digi-Key.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec8\" class=\"Section2\"\u003e \u003ch2\u003eMeasurements\u003c/h2\u003e \u003cp\u003eA scanning electron microscope (JEOL JSM-7200F SEM) at 2.0 kV voltage was used to characterize the detector. All electrical and optoelectronic measurements were performed at room temperature in a dark room to eliminate ambient light. Samples were measured unencapsulated (directly exposed to air). Illumination was provided by three compact LED sources (ELEGOO Upgraded 37 in 1 Sensor Modules Kit with Tutorial Compatible with Arduino IDE UNO R3 MEGA Nano, Amazon-sourced - \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://a.co/d/33fz0am\u003c/span\u003e\u003cspan address=\"https://a.co/d/33fz0am\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e), each operated at around 1.84 mW, driven by an Arduino with a series resistor, emitting at 620 nm (red), 520 nm (green), and 460 nm (blue), and the optical absorbing area was 24 mm\u0026sup2;. Unless otherwise noted, the detector was positioned at a fixed source\u0026ndash;detector distance of 20 cm. Signals were recorded using a Keithley 6500 digital multimeter, while rise/fall times were measured with an Agilent Technologies DSO7104B oscilloscope since Keithey 6500 max sampling rate is around 16.67 ms.\u003c/p\u003e \u003c/div\u003e"},{"header":"Declarations","content":"\u003cp\u003e\u003cstrong\u003eFunding:\u003c/strong\u003e The authors acknowledge the support from the Natural Sciences and Engineering Research Council of Canada (NSERC) and the Micro-Nano Technology (MNT) program facilitated by CMC Microsystems. We also thank Auroray Labs Inc. for valuable financial contributions, which have significantly advanced the technological development of this work.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAuthor contributions:\u003c/strong\u003e\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eConceptualization:\u0026nbsp;WG\u003c/p\u003e\n\u003cp\u003eMethodology: WG, JW, GL, YY\u003c/p\u003e\n\u003cp\u003eInvestigation:\u0026nbsp;WG\u003c/p\u003e\n\u003cp\u003eVisualization:\u0026nbsp;WG\u003c/p\u003e\n\u003cp\u003eSupervision:\u0026nbsp;JTWY\u003c/p\u003e\n\u003cp\u003eWriting\u0026mdash;original draft:\u0026nbsp;WG\u003c/p\u003e\n\u003cp\u003eWriting\u0026mdash;review \u0026amp; editing:\u0026nbsp;WG, JW, GL, JTWY\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eCompeting interests:\u003c/strong\u003e Authors declare that they have no competing interests.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eData and materials availability:\u003c/strong\u003e All data needed to evaluate the conclusions are available in the main text or the supplementary materials.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\u003cli\u003e\u003cspan\u003eP. X. Feng, E. Pacheco Cabrera, J. Chu, B. Zhou, S. Y. Flores, X. Peng, Y. Li, L. M. Diaz-Vazquez, A. F. Zhou, Self-Powered Photodetectors with Ultra-Broad Spectral Response and Thermal Stability for Broadband, Energy Efficient Wearable Sensing and Optoelectronics. \u003cem\u003eMolecules\u003c/em\u003e 30 (2025).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eY. Lin, Z. Yong, X. Luo, S. S. Azadeh, J. C. Mikkelsen, A. Sharma, H. Chen, J. C. C. Mak, P. G. Q. Lo, W. D. Sacher, J. K. S. Poon, Monolithically integrated, broadband, high-efficiency silicon nitride-on-silicon waveguide photodetectors in a visible-light integrated photonics platform. \u003cem\u003eNat Commun\u003c/em\u003e 13 (2022).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eW. Qarony, A. S. Mayet, E. Ponizovskaya Devine, S. Ghandiparsi, C. Bartolo-Perez, A. Ahamed, A. Rawat, H. H. Mamtaz, T. Yamada, S.-Y. Wang, M. Saif Islam, \u0026ldquo;Achieving higher photoabsorption than group III-V semiconductors in silicon using photon-trapping surface structures.\u0026rdquo;\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eJ. K. S. Poon, A. Govdeli, A. Sharma, X. Mu, F.-D. Chen, T. Xue, T. Liu, Silicon photonics for the visible and near-infrared spectrum. \u003cem\u003eAdv Opt Photonics\u003c/em\u003e 16, 1 (2024).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eJ. Ho, Z. Dong, H. Sheng Leong, J. Zhang, F. Tjiptoharsono, S. Daqiqeh Rezaei, K. Choon Hwa Goh, M. Wu, S. Li, J. Chee, C. Pei Yu Wong, A. I. Kuznetsov, J. K. W Yang, \u0026ldquo;Miniaturizing color-sensitive photodetectors via hybrid nanoantennas toward submicrometer dimensions\u0026rdquo; (2022); \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://www.science.org\u003c/span\u003e\u003cspan address=\"https://www.science.org\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eL. Chen, K. Preston, S. Manipatruni, M. Lipson, H. Chen, J. Zhang, N. A. Nelson, D. H. Albonesi, E. G. Friedman, P. M. Fauchet, C. Batten, A. Joshi, J. Orcutt, A. Khilo, B. Moss, C. Holzwarth, M. Popovic, H. Li, H. Smith, J. Hoyt, F. Kaertner, R. Ram, V. Stojanovic, K. Asanovic, H. Rong, R. Jones, A. Liu, O. Cohen, D. Hak, A. Fang, M. Paniccia, W. Fang, H. Park, M. J. Paniccia, J. E. Bowers, \u0026ldquo;Photonic networks-on-chip for future generations of chip multiprocessors\u0026rdquo; (2006); \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1364/OA_License_v1#VOR\u003c/span\u003e\u003cspan address=\"10.1364/OA_License_v1#VOR\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eD. Zhang, C. Fuentes-Hernandez, R. Vijayan, Y. Zhang, Y. Li, J. W. Park, Y. Wang, Y. Zhao, N. Arora, A. Mirzazadeh, Y. Do, T. Cheng, S. Swaminathan, T. Starner, T. L. Andrew, G. D. Abowd, Flexible computational photodetectors for self-powered activity sensing. \u003cem\u003enpj Flexible Electronics\u003c/em\u003e 6 (2022).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eM. Wang, M. Zhuang, L. Ge, Y. Yuan, Z. Wang, W. Liu, H. Li, A. Song, Tunable photoresponse properties of CuI/Si self-powered photodetectors through Zn doping engineering. \u003cem\u003eAppl Surf Sci\u003c/em\u003e 654 (2024).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eJ. Wang, Z. Xie, J. A. Liu, R. Zhou, G. Lu, J. T. W. Yeow, System design of large-area vertical photothermoelectric detectors based on carbon nanotube forests with MXene electrodes. \u003cem\u003eNanoscale Adv\u003c/em\u003e 5, 1133\u0026ndash;1140 (2023).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eJ. Wang, Z. Xie, J. A. Liu, G. Lu, R. Zhou, J. T. W. Yeow, Capillary-Assisted Self-Assembly of Carbon Nanotubes for the Self-Powered Photothermoelectric Detector. \u003cem\u003eAdv Mater Technol\u003c/em\u003e 8 (2023).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eG. Lu, R. Zhou, J. Wang, Z. Xie, Y. Yuan, J. T. W. Yeow, \u0026ldquo;Flexible Graphene/PEDOT: PSS Free-Standing Infrared Photodetector\u0026rdquo; in \u003cem\u003eProceedings of the IEEE Conference on Nanotechnology\u003c/em\u003e (IEEE Computer Society, 2024), pp. 180\u0026ndash;184.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eJ. Wang, Z. Xie, J. T. W. Yeow, Two-dimensional materials applied for room-temperature thermoelectric photodetectors. \u003cem\u003eMater Res Express\u003c/em\u003e 7 (2020).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eZ. Xie, J. Wang, G. Lu, J. T. W. Yeow, Room-ambient operation of integrated and visualized photothermoelectric system with patterned Mo2C/PEDOT: PSS flexible devices. \u003cem\u003eMater Des\u003c/em\u003e 235 (2023).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eJ. Wang, Z. Xie, J. A. Liu, J. T. W. Yeow, Design of room-temperature infrared photothermoelectric detectors based on CNT/PEDOT:PSS composites. \u003cem\u003eJ Mater Chem C Mater\u003c/em\u003e, doi: \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1039/d2tc03159k\u003c/span\u003e\u003cspan address=\"10.1039/d2tc03159k\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (2022).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eJ. Wang, Z. Xie, G. Lu, J. A. Liu, J. T. W. Yeow, An infrared photothermoelectric detector enabled by MXene and PEDOT:PSS composite for noncontact fingertip tracking. \u003cem\u003eMicrosyst Nanoeng\u003c/em\u003e 9 (2023).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eY. Lu, L. Liu, R. Gao, Y. Xiong, P. Sun, Z. Wu, K. Wu, T. Yu, K. Zhang, C. Zhang, T. Bourouina, X. Li, X. Liu, Ultrafast near-infrared pyroelectric detector based on inhomogeneous plasmonic metasurface. \u003cem\u003eLight Sci Appl\u003c/em\u003e 13 (2024).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eG. Lu, J. Wang, W. Gao, Z. Xie, R. Zhou, J. T. W. Yeow, Flexible Free-Standing Infrared Photothermoelectric Detector Based on Graphene/PEDOT. \u003cem\u003eIEEE Nanotechnol Mag\u003c/em\u003e, doi: \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1109/MNANO.2025.3570886\u003c/span\u003e\u003cspan address=\"10.1109/MNANO.2025.3570886\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (2025).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eG. Lu, J. Wang, W. Gao, J. T. W. Yeow, \u0026ldquo;Flexible Infrared Photothermoelectric Detectors Based on Graphene/Metallic Junctions with Lowcost Substrates\u0026rdquo; in \u003cem\u003e2025 IEEE 25th International Conference on Nanotechnology (NANO)\u003c/em\u003e (IEEE, 2025; \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://ieeexplore.ieee.org/document/11113664/\u003c/span\u003e\u003cspan address=\"https://ieeexplore.ieee.org/document/11113664/\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e), pp. 124\u0026ndash;127.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eS. Shekhar, W. Bogaerts, L. Chrostowski, J. E. Bowers, M. Hochberg, R. Soref, B. J. Shastri, Roadmapping the next generation of silicon photonics. \u003cem\u003eNat Commun\u003c/em\u003e 15 (2024).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eA. Li, C. Yao, J. Xia, H. Wang, Q. Cheng, R. Penty, Y. Fainman, S. Pan, Advances in cost-effective integrated spectrometers. Springer Nature [Preprint] (2022). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1038/s41377-022-00853-1\u003c/span\u003e\u003cspan address=\"10.1038/s41377-022-00853-1\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eS. Bin Choi, J. S. Choi, H. S. Shin, J. W. Yoon, Y. Kim, J. W. Kim, Deep learning-developed multi-light source discrimination capability of stretchable capacitive photodetector. \u003cem\u003enpj Flexible Electronics\u003c/em\u003e 9 (2025).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eA. D. Rakic\u0026acute;, R. Rakic\u0026acute;, A. B. Djuriš, J. M. Elazar, M. L. Majewski, \u0026ldquo;Optical properties of metallic films for vertical-cavity optoelectronic devices\u0026rdquo; (1998); \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1364/OA_License_v1#VOR\u003c/span\u003e\u003cspan address=\"10.1364/OA_License_v1#VOR\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eC. Y. Ho; R. W. Powell; P. E. Liley, Thermal Conductivity of the Elements. \u003cem\u003eJ. Phys. Chem. Ref. Data\u003c/em\u003e 1, 279\u0026ndash;421 (1972).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eI. T. Jollife, J. Cadima, Principal component analysis: A review and recent developments. Royal Society of London [Preprint] (2016). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1098/rsta.2015.0202\u003c/span\u003e\u003cspan address=\"10.1098/rsta.2015.0202\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eL. Van Der Maaten, G. Hinton, \u0026ldquo;Visualizing Data using t-SNE\u0026rdquo; (2008).\u003c/span\u003e\u003c/li\u003e\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":false,"highlight":"","institution":"","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"nature-portfolio","isNatureJournal":true,"hasQc":false,"allowDirectSubmit":false,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"","title":"Nature Portfolio","twitterHandle":"","acdcEnabled":false,"dfaEnabled":false,"editorialSystem":"ejp","reportingPortfolio":"","inReviewEnabled":true,"inReviewRevisionsEnabled":false},"keywords":"","lastPublishedDoi":"10.21203/rs.3.rs-8006942/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-8006942/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003eHigh-performance visible-light detection is vital for imaging and portable sensing, but most photodetectors are broadband intensity devices without intrinsic spectral selectivity and therefore rely on filter arrays or dispersive optics. Broadband visible photodetection with filter-free wavelength recognition is enabled by an inhomogeneous metasurface integrated with a convolutional Neural Network (CNN). Inhomogeneous metasurface creates geometry-dependent Localized Surface Plasmon Resonance (LSPR) hot spots that imprint wavelength-dependent signatures, while CNN is employed to recognize the signatures. The device shows broadband response in visible region with responsivity up to 1.45×10³ V W⁻¹, noise-equivalent power (NEP) down to 8.57 × 10\u003csup\u003e–16\u003c/sup\u003e W Hz\u003csup\u003e1/2\u003c/sup\u003e, specific detectivity up to 5.71×10\u003csup\u003e14\u003c/sup\u003e Jones. A lightweight 1D-CNN trained on a wavelength-distance dataset achieves 98.91% accuracy in 10-fold cross-validation and is validated by raster-scanning on a laptop display and traffic light recognition. The scalable Ag inhomogeneous metasurface/poly(3,4-ethylenedioxythiophene) polystyrene sulfonate (PEDOT:PSS) process and detector–CNN integration offer a practical route to low-cost, compact, filter-free colorful imaging and spectral identification.\u003c/p\u003e","manuscriptTitle":"Inhomogeneous Metasurface–CNN System for Filter-Free Single-Pixel Wavelength Recognition","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2026-02-03 15:37:27","doi":"10.21203/rs.3.rs-8006942/v1","editorialEvents":[],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"communications-engineering","isNatureJournal":true,"hasQc":false,"allowDirectSubmit":false,"externalIdentity":"commseng","sideBox":"Learn more about [Communications Engineering](http://link.springer.com/journal/44172)","snPcode":"44172","submissionUrl":"https://mts-commseng.nature.com/cgi-bin/main.plex","title":"Communications Engineering","twitterHandle":"@commseng","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"ejp","reportingPortfolio":"Springer Hybrid","inReviewEnabled":true,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"40cbfb03-5531-4a97-8f5d-62104a5712e2","owner":[],"postedDate":"February 3rd, 2026","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"under-review","subjectAreas":[{"id":57848038,"name":"Physical sciences/Engineering/Electrical and electronic engineering"},{"id":57848039,"name":"Physical sciences/Nanoscience and technology/Nanoscale devices/Nanophotonics and plasmonics"},{"id":57848040,"name":"Physical sciences/Optics and photonics/Applied optics/Optoelectronic devices and components"}],"tags":[],"updatedAt":"2026-02-03T15:37:28+00:00","versionOfRecord":[],"versionCreatedAt":"2026-02-03 15:37:27","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-8006942","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-8006942","identity":"rs-8006942","version":["v1"]},"buildId":"XKTyCvWXoU3ODBz1xrDgd","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}