Towards real-time additive-free dopamine detection at 10−8 mM with hardware accelerated platform integrated on camera Authors: Ning Li, Qizhou Wang, Zhao He, Arturo Burguete-Lopez, Fei Xiang, Andrea Fratalocchi Published: 2026 ABSTRACT AbstractTracing physiological neurotransmitters such as dopamine (DA) with detection limits down to 1×10−8 mM is a critical goal in neuroscience for studying brain functions and progressing the understanding of cerebral disease. Addressing this problem requires enhancing the current state-of-the-art additive-free electrochemical workstation methods by over two orders of magnitude. In this work, we implement an ultra-sensitive, additive-free platform exploiting suitably engineered light-scattering membranes and optical accelerators integrated into commercial vision cameras, reporting real-time detection of DA in uric and ascorbic acid below the concentration of 10−8 mM. These performances improve the current best technology by over two orders of magnitude in resolution while providing continuous, real-time detection at video rates. This technology also upgrades the bulk form factor of an electrochemical workstation with an imaging camera’s compact and portable footprint. The optical accelerator implemented in this work is universal and trainable to detect a wide range of biological analytes. This technology’s wide adoption could help enable early disease detection and personalized treatment adjustments while improving the management of neurological, mental, and immune-related conditions.Subject terms: Nanobiotechnology, Sensors and probes, Medical research, Optical spectroscopy, Neuroscience Introduction Tracing analyte concentrations at ultra-low quantities is essential across various scientific areas, including security1–3, medicine4–8, food safety9, and environmental protection10,11. In disease diagnosis, ultra-sensitive detection enables tracking subtle changes of biomarker concentration, which indicate early onset or progression of debilitating conditions12. Dopamine (DA) is a fundamental neurotransmitter biomarker whose microscopic change within the 10−7 mM range13–16 reveals various neurological disorders, including Parkinson’s17,18, Alzheimer’s19, depression20, schizophrenia21,22, and attention deficit hyperactivity disorder23. Practical strategies for ultra-sensitive detection of DA require identifying the analyte in interference media such as uric acid (UA) and ascorbic acid (AA) without any additive incorporation of enzymes, polymers, aptamers, etc., which introduce impurities and instabilities24,25. The state-of-the-art in additive-free biosensors that work in physiological conditions reaches the limit of 5 × 10−6 mM with Mn-MoS2/PGS sensor26. These biosensors do not work in real-time and necessitate measurement times from tens of seconds to several minutes27–29, while requiring an enhancement of more than two orders of magnitude to reach the range between 10−8 and 10−6 mM characteristic of a majority of brain disorders30,31. Devising pathways to enhance these performances is challenging due to background noise interference with the signal that, at ultra-low sensitivities, yields variation of measurement results across different manufactured devices32,33, making it difficult to deploy an efficient decision-making strategy that minimizes human bias34,35. Addressing these issues provides a significant opportunity in research, promising early diagnosis and low-cost routine monitoring that could help progress against elusive disorders affecting tens of millions of people worldwide. This work implements an ultra-sensitive detection platform leveraging hardware artificial intelligence (AI), incorporating data acquisition and feature extraction into a single process. The hardware accelerator integrates with commercially available monochrome cameras, retaining their footprint, portability, and video rate processing speed. In experimental tests, the system reports DA detection at a concentration of 10−8 mM with interference media including UA, AA, phosphate buffer solution (PBS), and artificial cerebrospinal fluid (aCSF). The proposed integrated hardware AI platform enhanced the detection limit of DA by six orders of magnitude compared to existing AI-based biochemical sensing approaches, which rely solely on software-implemented models such as support vector machine36 and neural networks (NNs)37,38. Furthermore, the proposed platform achieves two orders of magnitude compared to state-of-the-art additive-free DA detection technologies26, while providing a compact, fully integrated form factor that potentially enables point-of-care deployment. Results Figure 1 provides a general scheme of the ultra-sensitive detection platform we implement in this work. Input light with spectral intensity distribution Ii(ω) impinges on an analyte with unknown concentration y, placed inside a suitably designed scattering surface (Fig. 1a). When illuminated by a broadband input Ii(ω), the surface outputs a scattering response Io(ω, y) that could leverage different forms of light-matter interaction, including plasmonic resonances in metallic nanostructures39, Mie or Fano resonance in high-refractive-index dielectrics40, and chemical bonding between the analyte and molecules with high scattering cross section26. Schematics of the general sensing platform. a Spectral dataset generation using a scattering surface with broadband illumination. b Example of a generated spectral dataset for a varying concentration of analyte. c Encoder-decoder unsupervised training model with explainable-AI (XAI) interface for selecting latent features. d Implementation design of a hardware encoder with sub-pixels of inverse-designed nanoresonators. e Integration of the hardware encoder into a CMOS camera sensor. f Software post-processing network processes the extracted latent features into the analyte concentration. g Each sub-pixel encoder’s trained spectral transmission response implements the multiply and accumulate (MAC) neural network operation required to extract the selected latent space features. The spectra obtained at different analyte concentrations y compose a training dataset (Fig. 1b). The dataset contains sparse spectral features relating the concentration y with the input beam Ii(ω) spectrum. Instead of characterizing the data with prior-known metrics, the platform uses the spectral dataset to train a hardware encoder E, which extracts spectral features using AI hardware (Fig. 1c). The encoder E follows the methodology of unsupervised dimensionality reduction41, and transforms high-dimensional scattered spectra Io(ω, y) into a latent vector z=E(Io(ω,y)), representing a low-dimensional, refined feature space of the input spectra. In practical tasks such as regression and classification, a machine learning model can leverage the latent vectors to give predictions beyond normal metrics. A decoder model D employs the information contained in the latent vector z to reconstruct the input spectra, providing at the output the set of spectra Io′(ω,y) as close as possible to the input ones. Training of the E- D systems comprises minimizing the mean square error (MSE) loss42 between the input and reconstructed spectra: Lrec=1N∑y∥Io(ω,y)−Io′(ω,y)∥, where N represents the total number of spectra samples. As the training converges, the encoder E provides an information-preserving projection from the spectral scattering responses to the distribution of all possible low-dimensional representations, yielding a latent space encompassing the important sparse features of the input spectra Io(ω, y). The proposed framework consequently applies a supervised feature selection process in the latent space using explainable AI (XAI)43, which optimizes the number of latent features for predicting the analyte concentration y (arrow under Fig. 1c). We implement the encoder E in optoelectronic hardware using nanostructured encoder elements with engineered transmission responses in the frequency domain Λi(ω) (Fig. 1d) placed over the sensor of a conventional monochrome camera. In this process (Fig. 1d, e), each component zi in the latent vector z is hardware generated from input light Io transmitted through a surface Si with transfer function Λi(ω), and then converted into electronic signal in the photo-detector: with Δω = ω1−ω0 being the sensitivity bandwidth of the detector and σ its nonlinear saturable transfer-function44, or an additional post-conversion nonlinearity computed on the electronic stream. Equation (1) implements a multiply and accumulate (MAC) operation of NNs, with the camera conversion function σ resembling the corresponding activation function in NN. In the projection process described by Eq. (1), we inverse design the transmission Λi with an open-source inverse design software45 to yield the closest representation of zi in the least-square sense (Fig. 1f). To ensure the solution to this problem exists, we implement Si with a single metasurface containing a layer of physical universal approximators represented by nanoscale resonators with tailored resonance responses46. We arrange the set of S1, …SN projecting surfaces into a matrix of sub-pixel arrays projecting on the camera pixels (Fig. 1g). The image readout from the camera acquires the latent vector z and flattens it into an electric signal, which the regression network processes to predict the analyte concentration y (Fig. 1g). Figure 2 details the hardware encoder E feature extraction process, using an input spectral dataset X = [Io(ω, y)1, Io(ω, y)2, …, Io(ω, y)N] obtained from DA dissolved in PBS. In this work, we acquire spectral datasets with a commercial Ocean Optics UV-VIS spectrum analyzer equipped with an integrating sphere. The dataset X comprises N = 240 reflection spectra equally measured in 8 concentrations varying from 10−1 to 10−8 mM (Fig. 2a). In this work, we perform dimensionality reduction by principal component analysis (PCA) (Fig. 2b). PCA is a widely used technique in spectral material analysis and classification47–49 that assists in finding the smallest group of uncorrelated and repeated spectral shapes, which linearly combine to reconstruct the dataset X. PCA encodings implement a projection camera process with an identity transfer function, which maximizes the data processing speed. The PCA eigenvector components represent the MAC weights Λ(ω) = [Λ(ω)1, …, λM(ω)] for training the hardware accelerator encoder in Fig. 1f. The first M = 8 components represent the spectral data up to a total of 99.8% explained variance and result in a dataset reconstruction loss Lrec=1.2×10−6. Figure 2b projects the input spectral dataset X along the first three principal components. The projected principal components provide a new set of coordinates to represent the original reflectance spectra, serving as a basis for feature separation and identification. The principal features provide a new set of coordinates to represent the data with successive refinement, serving as a basis for feature separation and identification. The projected points in Fig. 2b show this property quantitatively, separating data distribution patterns in the space where high-concentration and low-concentration data points cluster in different areas. Design workflow of the optical hardware encoder. a Example of a dataset comprising scattered spectra from DA in UA/AA. b The dataset was projected along the first three principal components extracted from the data. c Explainable-AI integrated gradient method for extracting the essential features in the latent space for predicting the analyte concentration y. d Principal component’s importance score vs. explained variance. Components 1,4,5 and 6 are the most important in predicting the concentration y. We identify the most useful principal components for analyte concentration prediction using the XAI integrated gradient (IG) method50. The IG trains an auxiliary regression network F to predict the analyte concentration (green block in Fig. 2c). The network F uses input data comprising the spectral features z = [z1, …, zM], with zi being the projection of the input spectra along the ith principal component ui. The training of F aims to minimize the regression loss Lreg=∣∣y^−y∣∣ between its prediction output y^=F(z) and the ground truth concentration y obtained from measurements. The IG assigns to each input feature zi an importance score (IS), computed from: with z′=0 representing a “void sample,” and the integral accumulating the partial derivatives of prediction output y^ with regard to the input feature zi. The IG method constructs an integral path from z′ to z by linearly interpolating between them (Fig. 2c). Along this path, it computes the gradient on inputs z′+α(z−z)′ via backpropagation, with α ∈ [0, 1] being the interpolation factor. The integral value in (2) is rescaled by zi−zi′ to ensure that the IS is proportional to the distance between the input and the void sample. The larger the IS deviates from 0, the greater significance zi carries for predicting the analyte concentrations y. Figure 2d plots the absolute values of the calculated IS and the explained variances of each principal component. While consecutive principal components represent the input dataset with accumulating total variance (Fig. 2d dashed line), the IG analysis highlights that the four most informative components for predicting an analyte concentration are in a nonconsecutive order (Fig. 2d, red solid lines). Further reducing the PCA components to the four with the highest IS projects X to a feature space with an explained variance of 83.92% for the input data while carrying a total IS of 86.99%. Figure 3a shows the encoder integration on top of a commercial DMM 36AX290-ML board monochrome camera from The Imaging Source. Figure 3b shows a scanning electron microscope (SEM) image of the hardware encoders, with a one-to-one mapping between the camera sensor pixels and the metasurface sub-pixels (see fabrication details in “Methods”). Hardware implementation on the camera chip of XAI-selected encoders. a Integration of the hardware encoders on a CMOS camera sensor. b SEM image of the fabricated encoder nanostructures. Figure 4a illustrates the scattering device implemented in this study (details in the “Methods” section). The system integrates a scattering plasmonic system in a microfluidic cell, comprising a central cylindrical cavity with a diameter of 10 mm, capable of accommodating 1 mL of solution. The device includes two input and output analyte channels, each with a diameter of 0.5 mm, positioned on either side of the cavity (Fig. 4a, solid arrows). Figure 4b shows the schematic of the device’s assembly. The bottom scattering layer comprises a nanostructured silicon (Si) wafer decorated with metallic nanoparticles (Fig. 4c). We fabricate nanostructured Si using wet-chemistry self-assembly followed by nanostructuring via reactive ion etching (RIE). The RIE process is separately optimized for each analyte, leveraging the available RIE parameters comprising the gas flow of SF6, CHF3, O2, the set pressure, the power of inductively coupled plasma, the power of radio frequency, and etching time. The optimization process maximizes the separation of features in the latent space generated by the hardware encoder E (details in the “Methods” section). The metal decoration supports strong surface plasmon resonance at the metal-dielectric interface, where the collective oscillation of free electrons generates an intensified local electromagnetic field51,52. This plasmonic field enhancement amplifies subtle refractive index changes near the surface, leading to strong optical signals and high sensitivity53,54. The device’s top layer is a hydrophobic multilayer structure. This system implements a bottom layer of 1H,1H,2H,2H- Perfluorodecanethiol (PFDT, CF3(CF2)7CH2CH2SH), an interlayer of Polydimethylsiloxane (PDMS), and cover glass treated by Perfluorodecyltrichlorosilane (FDTS, CF3(CF2)5CH2CH2SiCl3). The hydrophobic inner design prevents the analyte from contaminating the scattering substrate, allowing the sensor to be reusable. This property can facilitate the sensor achieving a cleaner baseline, reduced refractive index drift, and enhanced measurement stability, which together lower background noise and improve detection limits. We further validate the overall stability of the scattering device over time. To assess baseline consistency, we measure the reflectance spectra of the device filled with deionized water at multiple time points over 17 days. Figure 4d shows the mean spectra (solid lines) and the corresponding ±3σ interval (shaded area) for each measurement day, where σ represents the standard deviation. The mean intra-day correlation coefficients are 0.9997, 0.9989, 0.9972, 0.9994, and 0.9986 for Day 1 through Day 17, respectively, confirming highly consistent spectral responses across all measurements. Figure 4e further visualizes the inter-day correlation coefficients, showing a gradual decline over time. Nevertheless, the baseline measurements maintain correlation coefficients above 0.994 after 17 days, which reveals subtle variations of the spectra and demonstrates long-term stability of the scattering device. Implementation of the scattering device. a Optical image of an experimentally assembled scattering device comprising a reusable microfluidic cell with input and output channels. b Schematic assembly of the microfluidic device composed of the scatterer, PFDT, PDMS, FDTS, and the cover glass. c Schematic of one scattering pillar. We engineer each pillar’s shape and metallic covering by optimizing the scattering process for predicting the analyte’s concentration. d Visualization of measured baseline spectra across 17 days. e Correlation matrix of inter-day measurements. Figure 5a shows the light path and the device assembly for measuring DA in PBS, with Fig. 5b, c detailing the nanoscale images of the scattering substrate engineered in the microfluidic cell. The 3D molecular model is generated with an open-source molecular editor, Avogadro. The SEM image in Fig. 5b shows that the RIE-optimized nanostructure for DA appears as elongated pillars on a hexagonal lattice with an average period of 200 nm and 535 nm height, characterized by a disordered profile modulation along the transverse z axis. Figure 5c provides additional details by showing scanning transmission electron microscopy-energy-dispersive X-ray spectroscopy (STEM-EDS) mapping of a material section comprising two consecutive nanopillars. Metallic Au nanoparticles deposited via sputtering accumulate primarily on top of the pillars, providing mushroom-like scattering shapes. Supplementary Fig. S1 shows individual STEM-EDS mapping of each component analyzed, including Si, Au, Sulfur (S), and Fluorine (F). Elements S and F, originating from assembled PFDT molecules, align with the regions where Au appears on the scatterer. This feature indicates that assembled PFDT via the liquid-phase growth method contributes to the inhomogeneous surface metal decoration observed in Fig. 5c. Detection of DA in buffer solution (PBS). a Schematic of the setup for measuring DA in PBS. b Tilted SEM and c overlay EDS elemental mapping of the fabricated optical scattering device. d CA image of the scatterer treated by PFDT. e SEM image of the first four hardware encoders and f trained transmission spectra weight functions. g Calibration of the model on 6 test samples per concentration, including mean predicted concentration versus true concentration with 95% CI at each level. h LOD and dynamic range of DA in PBS compared against the state-of-the-art approaches. Images in (b, c, e) are representative of at least three independent experiments. Supplementary Figs. S2 and S3 use X-ray photoelectron spectroscopy (XPS) to investigate the mechanisms of PFDT layer deposition over the substrate. In Fig. S2, we present the survey spectra of the scatterer and the PFDT-decorated scatterer, and we normalize them to their respective peak maximum to enable a direct comparison between the samples. The unexpected F 1s signal presents in the scatterer, which is due to fluorocarbon polymer formed during the etching process involving reactions of SF6 and CHF3 with Si55. Compared with the pristine scatterer, the Au 4f signal intensity decreases significantly after PFDT modification. The PFDT-decorated scatterer exhibits a stronger F 1s signal intensity than the control sample. Moreover, sulfur peaks appear in the PFDT-decorated scatterer but are absent in the pristine scatterer. These results indicate that the PFDT chains remain on the surface even after rinsing with solvents. Figure S3 shows the results of the XPS fitting peak for C 1s, Au 4f, and S 2p. Figure S3a shows that the C 1s spectrum contains peaks corresponding to C–C, C–S, –CF2, and –CF356,57, demonstrating the grafting of long fluorocarbon chains onto the surface of the scatterer. Figure S3b shows the XPS spectrum of Au 4f. The Au 4f spectrum shows two main peaks at 83.9 eV (Au 4f7/2) and 87.6 eV (Au 4f5/2), respectively. Fitting the Au 4f peaks reveals two states, Au(0) with Au 4f7/2 binding energy at 83.9 eV, and Au(I) at 84.4 eV58,59. The presence of Au(I) indicates the formation of interfacial covalent bonds between the scatterer and the PFDT ligands, specifically through Au–S bonding. Figure S3c shows the XPS spectrum of S 2p. We deconvoluted the S 2p spectrum into three components, and each component includes a spin-orbit splitting doublet (S 2p1/2 and S 2p3/2, △E ≅ 1.2 eV). The first component represents chemisorbed thiols (regions marked with S 2p1/2B and S 2p3/2B in Fig. S3c), while the second corresponds to weakly bound (physisorbed) thiol species (regions marked with S 2p1/2A and S 2p3/2A in Fig. S3c) that still remain on the surface even after thorough rinsing with ethanol and water60. In addition to these two components, a weak but distinct S 2p doublet (regions marked with oxidized sulfur in Fig. S3c) appears at around 168.3 eV and 169.6 eV. Binding energies in the range of 168.0–170.0 eV are commonly attributed to highly oxidized sulfur species such as sulfonates or sulfates61,62, indicating that a fraction of the thiol groups in the PFDT monolayer underwent oxidation during the assembly or subsequent exposure to air. Figure 5d employs contact angle (CA) measurements to quantify the hydrophobicity of PFDT-treated scatterers. The CA observed lies above 152.3 ± 0. 6∘ with the relative standard deviation of 0.36% computed from 10 different measurements, showing super-hydrophobic behavior (Fig. S4). It implies that no significant desorption or reorganization of the chemisorbed, physisorbed, or oxidized PFDT species occurs on the timescale of the measurements. Therefore, these surface components do not compromise the stability of the interface or the accuracy of the analyte detection. The CA observed for the FDTS-treated cover glass is 102.2 ± 0. 5∘ (Fig. S5), slightly lower than PFDT but still well above the hydrophobic threshold of 90∘. Figure 5e shows the morphology of the nanostructured metasurfaces representing the four XAI-selected principal components trained by the procedure of Fig. 2 in the case of DA analyte in PBS. Figure 5f illustrates the measured spectral responses Γi(ω) (i = 1, 2, 3, 4) of the trained MAC encoders. Figure 5g visualizes the calibration for evaluating the consistency of the predicted and true DA concentrations. From the measured 240 spectra with dopamine concentration across 10−8 to 10−1 mM, we leverage 80% of the samples to train the sensing framework and use 20% of the measurements to calibrate and evaluate. Following the common practice in AI-assisted biochemical detection works, the plot visualizes the distribution of predicted concentrations around true concentration levels, defined by their mean values and the 95% confidence interval (CI). As part of the curve calibration process, we apply linear regression to the predicted and true values, resulting in a fitted slope of 0.949, an intercept of −0.372, and a root mean square error (RMSE) value of 0.3113. Figure 5h summarizes the detection results of the proposed platform and compares them against the state-of-the-art26,63,64. The dynamic range achievable with this technique spans eight orders of magnitude, from 10−8 to 10−1 mM. We further evaluate the model’s generalization ability by calculating the R2 score of the prediction on the test dataset, which yields R2 = 0.9845. To align with common practice in AI-based sensing36,65, we define the limit of detection (LOD) as the lowest concentration that is validated under the calibration framework, which is 10−8 mM. The LOD is comparable to the best integrated and non-integrated state-of-the-art, respectively. It also provides the broadest dynamic range, surpassing the best Mn-MoS2/PGS electrochemical device by two orders of magnitude26. Figure 6 summarizes detection results for measuring DA in a PBS solution with the interference of UA and AA. Addressing performances with these interference media is necessary for the reliability of DA reading under physiological conditions and crucial for point-of-care diagnostics26. Figure 6a shows the SEM image of the nanostructures from the four XAI-selected light encoders that extract the required features for this type of analyte in the presence of UA and AA. It also depicts the corresponding trained MAC spectral weights in Fig. 6b. Similarly to the analysis of PBS, we collected 180 spectra data reflected from samples with different concentrations from 10−8 to 10−3 mM, and split the training/testing dataset as 4:1. We further extracted four PCA components carrying 87.12% of IS and 96.40% of the total variance of the training dataset. We conducted calibration and the 95% CI visualization in Fig. 6c, yielding fitted slope, intercept, and RMSE of 0.975, −0.215, 0.0979, respectively. Analysis of the test dataset reports an R2 score of 0.9971. At last, we repeat the experiments on DA in aCSF to emulate a realistic human body fluid environment. It reveals an R2 score of 0.9982 at the detection range from 10−8 to 10−3 mM. Figure 6d reveals a near-ideal calibration between the predicted and true concentration values, with a slope of 0.975, an intercept of −0.216, and a fitting RMSE of 0.0770. Figure 6e visualizes the detection performance compared to the state-of-the-art by using a hardware encoder implementing the four selected PCA components. We here obtain LOD of 10−8 mM and a dynamic range spanning from 10−8 to 10−3 mM in both UA/AA and aCSF medium. We also visualize the confusion matrix of the model’s prediction in Supplementary Fig. S6. Comparable to available technologies26,66–69, the integrated technique presented succeeds in overcoming the threshold limit range from 10−8 to 10−6 mM (gray rectangle region) required for monitoring DA fluctuations in neurological disorders13,15,16. To complement the graphical comparison, we additionally provide Table 1 that details the LOD, dynamic range, and R2 value of the state-of-the-art and our work. Supplementary Tables S1–S3 provide detailed information on the performance metrics of medium-specific approaches26,63,64,66–81. Detection of DA in interference media (UA/AA and aCSF). a SEM image of four hardware encoders with corresponding b measured trained spectral weights. Calibration of the detection model on c UA/AA and d aCSF, including mean predicted concentration versus true concentration with 95% CI on six test samples at each concentration level. e LOD and dynamic range of DA detection compared against the state-of-the-art. Images in (a) are representative of at least three independent experiments. Table 1Comparison of DA detection approaches: limit of detection, dynamic range, and performance across different mediaApproachMediumLOD (mM)Dynamic Range (mM)R2PEDOT-modified SPE63PBS8.87 × 10−71 × 10−6 to 2 × 10−10.99872SWNT64PBS5 × 10−95 × 10−6 to 1 × 10−30.9996Mn-MoS2/PGS26PBS5 × 10−85 × 10−8 to 5 × 10−2–Integrated hardware AI (This Work)PBS1 × 10−81 × 10−8 to 1 × 10−10.9845SnO2/AuNPs/PDA-RGO/GCE66UA/AA5 × 10−68 × 10−6 to 2 × 10−20.9986GQDs/IL-SPCE67UA/AA6 × 10−52 × 10−4 to 1.5 × 10−20.9963Mn-MoS2/PGS26UA/AA5 × 10−65 × 10−6 to 5 × 10−4–Integrated hardware AI (This Work)UA/AA1 × 10−81 × 10−8 to 1 × 10−30.9971Ag film/Nafion/CNT-PPy MIP68aCSF1.89 × 10−81 × 10−6 to 1 × 10−2–MIPs/pThi/NPG69aCSF–3 × 10−4 to 3 × 10−20.998Mn-MoS2/PGS2610% serum5 × 10−65 × 10−6 to 5 × 10−1–Integrated hardware AI (This Work)aCSF1 × 10−81 × 10−8 to 1 × 10−30.9982PEDOT poly(3,4- ethylenedioxythiophene), SPE screen-printed electrode, SWNT single-walled carbon nanotube, PGS pyrolytic graphite sheet, PDA polydopamine, RGO reduced graphene oxide, GCE glassy carbon electrode, GQDs graphene quantum dots, IL ionic liquid, SPCE screen-printed carbon electrode, CNT carbon nanotube, PPy polypyrrole, MIP molecular imprinting polymer, pThi polythionine, NPG nanoporous gold. Comparison of DA detection approaches: limit of detection, dynamic range, and performance across different media PEDOT poly(3,4- ethylenedioxythiophene), SPE screen-printed electrode, SWNT single-walled carbon nanotube, PGS pyrolytic graphite sheet, PDA polydopamine, RGO reduced graphene oxide, GCE glassy carbon electrode, GQDs graphene quantum dots, IL ionic liquid, SPCE screen-printed carbon electrode, CNT carbon nanotube, PPy polypyrrole, MIP molecular imprinting polymer, pThi polythionine, NPG nanoporous gold. Discussion This work implements a hardware-accelerated ultra-sensitive platform technology to detect biochemical analytes in real time. The proposed technology integrates conventional camera sensors and does not require extra additives, avoiding traditional impurity treatment requiring between tens of minutes and several hours82 that make continuous motoring impossible. The entire manufacturing process for a single detection system is within 1000 USD, evaluated with clean room actual costs. This figure represents 1/10 of the price of a state-of-the-art electrochemical workstation, which also requires a significantly larger footprint than a portable integrated camera. By reaching the threshold of DA detection at 10−8 mM, the proposed technology can enable real-time tracking of addiction-related DA dynamics83, and early diagnosis of Parkinson’s disease, facilitating the study of DA’s influence on emotional regulation84,85, and neurotoxic effects86. The universal design of this sensing platform, combining trained optical accelerators with software NNs, can extend this study on a wide range of additional analytes relevant to medical research, including serotonin and norepinephrine for studying depression and anxiety87, glutamate and gamma-aminobutyric acid for understanding neurodegenerative diseases like Alzheimer’s and epilepsy88,89, and cytokine indicators for cancer and immune system disorders90. Methods Materials and reagents We purchase polystyrene (PS) latex particles (PS200NM, 10% w/v) from MAGSPHERE Inc. We order SylgardTM 184 from Electron Microscopy Sciences. We acquire ethanol, toluene, and sodium dodecyl sulfate solutions (SDS, 10% in H2O), PFDT, PBS, DA, UA/AA from Sigma-Aldrich Inc. We obtain Si wafers from University Wafer Inc. We purchase the artificial cerebrospinal solution (BZ453) from BiochemazoneTM. We use deionized water (Milli-Q® Advantage A10) with a resistance of 18 MΩ cm in all experiments. Fabrication of the scattering surface We clean the Si surface with isopropanol, acetone, and ultra-pure water and then dry it with nitrogen gas. Next, we prepare the PS solutions by mixing the ethanol and water with a volume ratio of 1:1 and make these latex particles disperse in the solutions homogeneously by ultrasonic process for 30 min. We then perform self-assembly of PS on the water with the assistance of SDS. After the air-drying process, we etch the sample in the RIE instrument (Oxford Instruments, Plasmalab 100 - ICP 380). We use an oxygen plasma process (JST Manufacturing, Inc., ST0122A0) to remove the residual PS. We finally deposit 80 nm gold film on the structured Si substrate in the sputtering process (Equipment Support Company Ltd, ESCRD4), followed by the annealing process (Qualflow Them/Jipelec, JETFIRST) at 200 °C for 5 min. Assembly of the scattering device We perform an oxygen plasma process on the PDMS layer for 1 min. We then assemble it with a PFDT decorated scatterer and an FDTS modified cover glass. After stabilizing the structure for 2 days, the cell is ready for experiments. Fabrication of the sub-pixel encoders Using plasma-enhanced chemical vapor deposition, we deposit amorphous silicon on a fused silica substrate. We consequently spin coat and thermal-process the AR-P 6200.09 (Allresist GmbH) positive electron beam resist and a conductive polymer AR-PC 5090.02 layer on the substrate. We employ a JEOL JBX-6300FS electron-beam lithography system to write the nanostructure pattern, followed by a resist development process. We then deposit a 22 nm layer of chromium on the substrate to form a protective layer in the shape of the patterns to be unetched. We use a RIE system to expose the patterns of silicon. As a final step, we remove the chromium protection mask with perchloric acid and ceric ammonium nitrate solution (TechniEtch Cr01 from MicroChemicals). Physical characterization We collect the SEM images on a Carl Zeiss Merlin field-emission SEM with an In-Lens detector. We collect the high-angle annular dark-field scanning transmission electron microscopy (HAADF-STEM) images and STEM-EDS mapping of the FIB-prepared specimen with Titan ST transmission electron microscope (FEI Company) at an operating voltage of 300 kV. We perform the XPS measurements on the Kratos AXIS Supra system. We use a UV-vis spectrometer (Ocean Optics SR-6XR250-10) alongside the OceanView 2.0.18 software to characterize the reflectance of the samples in the 400–800 nm wavelength range. Reporting summary Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article. Supplementary information Supplementary Information Reporting Summary Transparent Peer Review file Source data Source data Acknowledgements The authors conducted the hardware encoder’s design using computing resources provided by KAUST Supercomputing Core Lab. The authors carried out the hardware encoder’s fabrication, scatterer fabrication, scanning electron microscopy, and transmission electron microscopy analysis at KAUST’s Nanofabrication, Imaging, and Characterization Core Lab, respectively. The authors thank KAUST’s Core Labs staff for their help in the design and fabrication processes. Author contributions N.L., Q.W., and A.F. initiated, managed, and planned the overall project. N.L. and Z.H. completed the fabrication process. Q.W. finished the machine learning and related theory. A.B.L. helped calibrate the spectrometer. F.X. performed FIB specimen preparation and STEM/EDS characterizations. N.L. and Q.W. contributed to this work equally. All authors contributed to data analysis, manuscript refinement, and preparation. Peer review Peer review information Nature Communications thanks the anonymous reviewer(s) for their contribution to the peer review of this work. A peer review file is available. Funding The authors acknowledge funding from KAUST (Baseline Grant BAS/1/1628-01-01) and KAUST Smart Health Center of Excellence (Grant FCC1593207). Data availability Source data are provided with this paper. Code availability The code91 used in this work is available at https://github.com/QizhouW/Hardware-accelerated-Dopamine-Sensing. Competing interests The authors declare no competing interests. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. These authors contributed equally: Ning Li, Qizhou Wang. Supplementary information The online version contains supplementary material available at 10.1038/s41467-026-73932-1. References Associated Data Supplementary Materials Supplementary Information Reporting Summary Transparent Peer Review file Source data Data Availability Statement Source data are provided with this paper. The code91 used in this work is available at https://github.com/QizhouW/Hardware-accelerated-Dopamine-Sensing.