Physical echo state network based on the nonlinearity and dynamic response of ambipolar heterostructure transistors.

Skip to main content An official website of the United States government Here's how you know Here's how you know Official websites use .gov A .gov website belongs to an official government organization in the United States. Secure .gov websites use HTTPS A lock ( Lock Locked padlock icon ) or https:// means you've safely connected to the .gov website. Share sensitive information only on official, secure websites. Search Log in Dashboard Publications Account settings Log out Search… Search NCBI Primary site navigation Search Logged in as: Dashboard Publications Account settings Log in Search PMC Full-Text Archive Search in PMC Journal List User Guide PERMALINK Copy As a library, NLM provides access to scientific literature. Inclusion in an NLM database does not imply endorsement of, or agreement with, the contents by NLM or the National Institutes of Health. Learn more: PMC Disclaimer | PMC Copyright Notice Nat Commun . 2026 Feb 28;17:3321. doi: 10.1038/s41467-026-70171-2 Search in PMC Search in PubMed View in NLM Catalog Add to search Physical echo state network based on the nonlinearity and dynamic response of ambipolar heterostructure transistors Wen-Min Zhong Wen-Min Zhong 1 College of Civil and Transportation Engineering, Shenzhen University, Shenzhen, P. R. China 2 Institute for Advanced Study, Shenzhen University, Shenzhen, P. R. China Find articles by Wen-Min Zhong 1, 2 , Wenbin Zhang Wenbin Zhang 2 Institute for Advanced Study, Shenzhen University, Shenzhen, P. R. China Find articles by Wenbin Zhang 2 , Yu-Xiang Zeng Yu-Xiang Zeng 2 Institute for Advanced Study, Shenzhen University, Shenzhen, P. R. China Find articles by Yu-Xiang Zeng 2 , JiYu Zhao JiYu Zhao 3 State Key Laboratory of Fine Chemicals, Frontiers Science Center for Smart Materials, Dalian University of Technology, Dalian, P. R. China Find articles by JiYu Zhao 3 , Ziqi Jia Ziqi Jia 2 Institute for Advanced Study, Shenzhen University, Shenzhen, P. R. China Find articles by Ziqi Jia 2 , Guanglong Ding Guanglong Ding 4 State Key Laboratory of Radio Frequency Heterogeneous Integration, Shenzhen University, Shenzhen, P. R. China 5 College of Electronics and Information Engineering, Shenzhen University, Shenzhen, P. R. China Find articles by Guanglong Ding 4, 5 , Su-Ting Han Su-Ting Han 6 Department of Applied Biology and Chemical Technology, The Hong Kong Polytechnic University, Hung Hom, Hong Kong SAR P. R. China Find articles by Su-Ting Han 6 , Vellaisamy A L Roy Vellaisamy A L Roy 7 School of Science and Technology, Hong Kong Metropolitan University, Ho Man Tin, Hong Kong SAR P. R. China 8 James Watt School of Engineering, University of Glasgow, Glasgow, UK Find articles by Vellaisamy A L Roy 7, 8 , Ye Zhou Ye Zhou 2 Institute for Advanced Study, Shenzhen University, Shenzhen, P. R. China 4 State Key Laboratory of Radio Frequency Heterogeneous Integration, Shenzhen University, Shenzhen, P. R. China Find articles by Ye Zhou 2, 4, ✉ Author information Article notes Copyright and License information 1 College of Civil and Transportation Engineering, Shenzhen University, Shenzhen, P. R. China 2 Institute for Advanced Study, Shenzhen University, Shenzhen, P. R. China 3 State Key Laboratory of Fine Chemicals, Frontiers Science Center for Smart Materials, Dalian University of Technology, Dalian, P. R. China 4 State Key Laboratory of Radio Frequency Heterogeneous Integration, Shenzhen University, Shenzhen, P. R. China 5 College of Electronics and Information Engineering, Shenzhen University, Shenzhen, P. R. China 6 Department of Applied Biology and Chemical Technology, The Hong Kong Polytechnic University, Hung Hom, Hong Kong SAR P. R. China 7 School of Science and Technology, Hong Kong Metropolitan University, Ho Man Tin, Hong Kong SAR P. R. China 8 James Watt School of Engineering, University of Glasgow, Glasgow, UK ✉ Corresponding author. Received 2025 Jan 8; Accepted 2026 Feb 19; Collection date 2026. © The Author(s) 2026 Open Access This article is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License, which permits any non-commercial use, sharing, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if you modified the licensed material. You do not have permission under this licence to share adapted material derived from this article or parts of it. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by-nc-nd/4.0/ . PMC Copyright notice PMCID: PMC13065983  PMID: 41764186 Abstract In the field of neuromorphic computing, time-series prediction poses a significant challenge to recurrent neural network architectures, often requiring task-specific customization that limits the development of general-purpose computing platforms. In this work, we implement a physical echo-state network (ESN) using ambipolar organic–inorganic heterostructure transistors to form its reservoir layer. Leveraging the ambipolar nature of the transistor, its variable-resistance region enables sparse matrix operations, while the saturation region provides tanh-like nonlinearity, making it well-suited for implementing both synaptic weighting and neuronal activation in an ESN. Additionally, its dynamic response naturally introduces temporal attributes. Thus, it can serve as a neuromorphic computing model for time-series tasks. Without the involvement of dynamic mechanisms, it is capable of performing image recognition, time-series prediction, and multimodal recognition tasks. When dynamic mechanisms are incorporated, the model achieves an accuracy of 96.98% on the MNIST handwritten digit dataset and 86.67% on the Fashion-MNIST dataset. This work offers a neuromorphic computing architecture, providing insights for tasks such as nonlinear mapping and time-series prediction. Subject terms: Electronic devices, Electronic and spintronic devices In neuromorphic computing, time-series prediction is a significant challenge to recurrent neural network architecture. Here, the authors present ambipolar transistors implementing a physical echo-state network for efficient time-series and image tasks, achieving high accuracy via static and dynamic computing modes. Introduction With the rapid development of artificial intelligence, computer chips are increasingly presenting challenges to humanity 1 – 5 . With the advancement of manufacturing processes, the size of transistors is approaching physical limits, meaning it is difficult to increase the density of transistors per unit area 6 . This has led to rising chip manufacturing costs, significant heat generation, and increased energy consumption 7 , 8 . In the von Neumann architecture, the separation of memory and processor causes memory bottleneck issues. To solve these problems, neuromorphic architectures have been proposed 9 – 11 . These novel computing architectures are inspired by the human brain and train neural networks in a manner that mimics brain function 12 – 15 . Recently, researchers have utilized physical devices to simulate biological synapses and neurons, connecting these devices to construct neuromorphic neural networks. These networks emulate the brain’s working mechanisms, significantly enhancing efficiency and reducing energy consumption. In a traditional sense, physical devices have been used as synapses to implement physical weights and as neurons to perform data mapping functions 16 – 22 . Neural networks composed of these devices have significantly advanced the field of image recognition. However, there is relatively limited research on using recurrent neural networks to address time series problems. This is partly because the training process is more complex, requiring a greater range of conductive states from physical devices. To overcome this challenge, researchers have begun exploring physical reservoir computing 23 – 28 . This approach involves constructing a reservoir with memory to replace the memory units in recurrent neural networks. Previous studies have utilized the nonlinear conductive states of physical devices to capture information from time series 29 . While this method is effective, it also presents limitations, such as difficulties in addressing binarization and multi-bit issues. An echo state network (ESN) is a specialized type of recurrent neural network that includes a reservoir layer 30 – 32 . The reservoir layer is constructed with a sparse matrix with fixed weights and neurons, which do not require training. The data is processed by the reservoir, resulting in reservoir states with high-dimensional features. The neural network excels at time series prediction tasks and can also classify image data that has been transformed into a time series. Consequently, reservoir layers constructed with physical devices hold significant potential. In this work, we design ambipolar transistors based on poly(3-hexylthiophene) (P3HT)–zinc oxide (ZnO) heterostructures for implementing a physical echo-state network (ESN). Serving simultaneously as synaptic and neuronal elements within the ESN reservoir, these transistors generate high-dimensional reservoir states from electrical inputs, which are subsequently classified by a simple readout layer for recognition or prediction tasks. Across diverse machine learning benchmarks, the reservoir states produced by the ambipolar ESN consistently surpass those of unipolar counterparts, highlighting the significance of ambipolar conduction in reservoir construction. We evaluate the physical ESN in both static and dynamic operational modes. In static mode, the system achieves 95.27% accuracy on the MNIST handwritten digits dataset and 86.10% on Fashion-MNIST. For time-series prediction, it delivers normalized root-mean-square errors (NRMSE) of 0.15%, 0.22%, and 0.19% along the x-, y-, and z-axes of the Lorenz attractor, respectively. In typhoon trajectory forecasting, the ESN exhibits optimal performance with a 500 × 500 reservoir, and attains 89.7% accuracy in multimodal classification tasks. In dynamic mode, data are preprocessed through the intrinsic temporal response of the device, embedding temporal correlations before ESN training. This approach enhances system performance, yielding 96.98% accuracy on MNIST digits and 86.67% on Fashion-MNIST. Moreover, in Lorenz attractor prediction, the dynamic mode achieves NRMSEs of 0.165, 0.28, and 0.206% across the three axes—performance comparable to that attained with a forgetting factor in conventional ESNs. Together, these results demonstrate that ambipolar transistors constitute a promising hardware platform for high-performance neuromorphic computing. Result and discussion The neural network comprises three distinct layers: the input layer, the reservoir layer and the readout layer (Fig. 1a ). The input layer is used for inputting data and transforming it to other dimensions. It allows the data to be fed into the reservoir layer. The reservoir is a layer constructed with a sparse matrix and neurons. It has the function of extracting features to map data to high dimensions and can also record historical information. It is responsible for outputting the reservoir states of the ESN. The readout layer is the layer that interprets the reservoir states. The reservoir states can be quickly solved through linear regression 33 – 35 . At the same time, other machine models can also be utilized for processing, such as Support Vector Machines (SVM) or deep neural networks. The cross-sectional morphology of the transistors with various thicknesses of ZnO are shown(Fig. 1b ). The thickness of P3HT deposited on ZnO of varying thicknesses is almost consistent, close to 70 nm. The thickness of ZnO has a minor effect on the deposition of P3HT, and the semiconductor performance of P3HT is almost uniform. The high-resolution image shows that the thicknesses of ZnO are 17.8, 12.2, and 8.0 nm, respectively, corresponding to the n-type, ambipolar and p-type transistors. The surface morphology of the ZnO and P3HT layers of the transistors are captured(Fig. 1c ). The AFM images reveal that both the ZnO and P3HT layers are relatively smooth. The root mean square (RMS) surface roughness of the ZnO layers for the n-type transistor, ambipolar transistor and p-type transistor are 935, 718 and 704 pm, respectively. Correspondingly, the surface roughness of the P3HT layers are 4.67, 3.36 and 4.12 nm, respectively. Fig. 1. Structure and characterization of the ambipolar transistor. Open in a new tab a Schematic diagram of ESN. b TEM image of the cross-sectional morphologies under 100,000 x magnification: n -type transistor, ambipolar transistor and p -type transistor cross-section. High-resolution cross-sectional morphologies of ZnO layers in n-type transistor, ambipolar transistor and p -type transistor. c Surface morphologies observed under AFM: ZnO layers in n -type, ambipolar, and p -type transistors, as well as P3HT layers in n-type, ambipolar, and p -type transistors. d Output curves of n -type, ambipolar, and p -type transistors using 100 nm SiO 2 as the dielectric layer. e Optical image of the ambipolar transistor array. f Optical image of the physical ESN system. g Output characteristic of the ambipolar transistor in the array and tanh function. The output performance of the transistors is illustrated(Fig. 1d ). The p -type silicon wafer acts as the gate, and the 100 nm SiO 2 layer serves as the gate dielectric. The ZnO layer is deposited on top, serving as the n-type channel material for the transistor. Subsequently, the P3HT layer is deposited on the ZnO layer, acting as the p-type channel material. An Au electrode is then deposited on the P3HT layer. The ambipolar transistor can transport electrons or holes depending on the applied gate voltage. For the n-type transistor, when the gate voltage (V gs ) is less than 0, electrons in the ZnO layer are captured by O 2− ions, and holes in the P3HT layer are released by the conjugated chain 36 – 38 . Given that the ZnO layer is 17.8 nm thick, the electron concentration is higher than the hole concentration, resulting in the transistor being in an off state. When V g is greater than 0, electron pairs in the ZnO layer are released by O 2− ions, and holes in the P3HT layer are captured by the conjugated chain. As the drain-source voltage (V ds ) increases, the drain-source current (I ds ) rises linearly until it reaches the pinch-off point, after which I ds tends to saturate, indicating the transistor is in an on state. Ids reaches approximately 0.93 µA at V g = 30 V and V ds = 30 V. For the ambipolar transistor, when V g is less than 0, electrons in the ZnO layer are captured by O 2− ions, and holes in the P3HT layer are released by the conjugated chain. As V ds increases, I ds rises linearly until the pinch-off point, after which it saturates, indicating an on state. I ds is approximately −0.33 µA when V g = − 30 V and V ds = − 30 V. When V g is greater than 0, electrons in the ZnO layer are released by O 2− ions, and holes in the P3HT layer are captured by the conjugated chain. I ds again rises linearly with V ds until saturation, indicating an on state. I ds reaches about 0.33 µA at V g = 30 V and V ds = 30 V. The balanced performance of the ZnO and P3HT layers allows the transistor to exhibit ambipolar properties. For the p-type transistor, when V g is less than 0, electrons in the ZnO layer are captured by O 2− ions, and holes in the P3HT layer are released by the conjugated chain. As V ds increases, I ds rises linearly until the pinch-off point, after which it saturates, indicating an on state. I ds is approximately − 0.34 µA when V g = − 30 V and V ds = − 30 V. When V g is greater than 0, electron pairs in the ZnO layer are released by O 2− ions, and holes in the P3HT layer are captured by the conjugated chain. Due to the ZnO layer being 8.0 nm thick, its electron concentration is lower than the hole concentration in the P3HT layer, resulting in the transistor being in an off state. Due to the presence of traps within P3HT and ZnO, its transfer characteristic exhibits hysteresis (Supplementary Fig. S1 ). In fabricating the array device, we first employed a 100-nm-thick Al₂O₃ layer as the gate dielectric. The array structure (Fig. 1e ) and the corresponding hardware system (Fig. 1f ) are shown, along with the memory characteristics of the array (Supplementary Figs. S2 – S6 ). To further optimize device performance, we reduced the dielectric thickness to 30 nm and 10 nm (Supplementary Fig. S7 ). Thinning the dielectric layer effectively lowers the operating voltage while preserving ambipolar behavior. Transistors fabricated with the 100 nm Al₂O₃ dielectric also exhibit well-defined output characteristics consistent with their device type (Supplementary Fig. S8 ). To advance integration density, we further scaled the channel length down to 2 µm (Supplementary Fig. S9 ). With the dielectric thickness reduced to 10 nm, the device achieves an ultralow switching energy of 202 fJ (Supplementary Fig. S10 and Supplementary Table S1 ), providing a strong foundation for future low‑power electronics. Moreover, the device shows clear dynamic response (Supplementary Figs. S11 , S12 ), supporting its potential in neuromorphic computing applications. Notably, the device’s transfer characteristic resembles the tanh function (Fig. 1g ), enabling simultaneous data mapping and signal bounding. Specifically, in the linear region, it can perform sparse matrix multiplication, whereas in saturation, it offers nonlinearity and amplitude constraint. Its dynamic properties further introduce temporal correlation processing. Collectively, these features highlight the advantages of using ambipolar transistors to implement physical echo‑state networks for neuromorphic computing. To demonstrate the capabilities of the physical ESN, we first conduct a letter recognition task involving six letters (O, Z, W, T, X, U). Each letter is repeated 2000 times and assigned a label from 0 to 5. The dataset is shuffled and split into 10,000 training and 2000 test samples. Letter images are flattened into time-series signals (Supplementary Fig. S13 ) and fed into an input layer with randomized signed weights, converting grayscale values into bipolar features. These features are then mapped to voltage inputs for a reservoir layer implemented as a 16 × 16 transistor matrix (Fig. 2a, b ). Sparse connectivity is emulated by activating only 10% of the synaptic connections. Two key characteristics distinguish transistor-based synaptic multiplication. First, ambipolar transistors exhibit two linear regions under positive and negative gate biases, enabling a wider and more accurate multiplicative range compared to unipolar (n- or p -type) transistors, which operate in only one linear region. Second, ambipolar devices achieve balanced channel conduction through complementary ZnO and P3HT layers, mitigating the inherent conductivity imbalance of unipolar transistors. In a comparative matrix multiplication experiment (Supplementary Fig. S14 ), synaptic arrays encoded matrix elements as V ds (first matrix) and V g (second matrix). Ambipolar synaptic arrays produced results closest to digital computation, outperforming unipolar counterparts (Supplementary Fig. S15 ). Fig. 2. An ESN constructed with transistors for letter recognition. Open in a new tab a The transformation process of an English letter at the output layer. b Schematic diagram of the reservoir layer constructed with transistors. c Reservoir state output of an ESN. d Confusion matrix of ESN recognizing English letters. When transistors act as neurons, their conductance characteristics and saturation behavior critically influence the reservoir. Ambipolar transistors show symmetric current saturation under both gate polarities beyond the pinch-off voltage, providing an inherent nonlinear activation function analogous to tanh. Under ±30 V time-series inputs (Supplementary Fig. S16 ), n-type neurons output only positive currents, p-type only negative currents, while ambipolar neurons generate balanced ambipolar responses. Consequently, reservoir states differ markedly: n-type ESNs produce outputs from 0 to 0.75 µA, p -type from 0 to – 0.1 µA, and ambipolar ESNs from – 0.1 to 0.1 µA (Fig. 2c ). Despite these differences, all ESN variants achieve 100% accuracy in the letter recognition task (Fig. 2d and Supplementary Fig. S17 ). The complete system operation is further illustrated (Supplementary Fig. S18 ). The MNIST datasets are used for classification 39 – 41 . These images are all 28 × 28 size and are fed into the physical ESN in the same way as above. 60,000 images are used as the training set, while the remaining 10,000 images serve as the test set. Thanks to the inherently excellent nonlinearity of ambipolar transistors, even when performing direct image straightening, the data can be mapped to a high-dimensional space. This allows the readout layer to easily interpret these reservoir states, enabling a single-step solution with high accuracy. The ESN achieves an accuracy of 95.27% on the grayscale MNIST handwritten digit dataset and 86.10% on the MNIST fashion dataset(Supplementary Fig. S19 – S23 ). This is not the ultimate limit of the architecture. To achieve higher accuracy, scaling up the sparse matrix size is required. Given its already high accuracy, further improvements would necessitate additional devices, introducing inherent limitations. Another method involves prepossessing the datasets with the dynamic response of the device and then using the prepossessing datasets to complete the recognition task. Since the dynamic response can provide features for the datasets, effectively adding dynamic nonlinearity, it can also yield favorable results. We further binarize the MNIST dataset and use 4-bit encoding to map the images, followed by processing with a physical ESN for recognition. Different pulse intervals in the 4-bit encoding are employed to impart more features (Supplementary Figs. S24 and S25 ). Information encoding is first established through a four-bit scheme (Fig. 3a ). Different types of nonlinearity are achieved by adjusting the interval pulses. For example, when the interval is increased, the responses produced by the “0,0,0,0” and “1,0,1,0” encodings are different (Fig. 3b ). This gives us further opportunities for exploration. When images are processed using encodings with different pulse intervals, different image mappings are obtained (Fig. 3c ). When these mappings are processed by the physical ESN, they produce different reservoir space states (Fig. 3d ). Both the digital dataset and the fashion dataset exhibit discrete reservoir states (Supplementary Fig. S26 ). However, their time attributes are incorporated due to preprocessing with different pulse intervals. This results in the potential for creating better performance opportunities. When the mappings converted from six pulse intervals are read out by the physical ESN, the accuracy reaches 96.98% (Fig. 3e ). Similarly, on the fashion dataset, after dynamic processing, the accuracy also increases relatively to 86.67% (Supplementary Fig. S27 ), higher than 86.10%. Therefore, introducing nonlinearity and dynamic mechanisms through physical methods in MNIST dataset processing may serve as an effective approach to enhance performance while reducing computational costs. Fig. 3. Dynamic-mechanism-based physical ESN for MNIST handwritten-digit recognition. Open in a new tab a Four-bit encoding built on the dynamic response. b Realization of “0,0,0,0” and “1,0,1,0” codes under different pulse intervals. c Mapping of the MNIST dataset achieved by varying pulse intervals. d Corresponding reservoir-state space generated by the physical ESN for these mappings and e Confusion matrix obtained with five combined pulses, showing the final classification performance. For neuromorphic computing systems, the capability to forecast temporal and spatial sequences is essential for ESNs 42 – 44 . To evaluate this, we apply our ESN to predict the Lorenz attractor, using data at time step t as input and at t + 2 as target. This task represents a second key challenge in time‑series prediction, an area where hardware‑based architectures are actively studied. We explore two training strategies. The first adopts the forgetting‑factor method commonly used in ESNs (Supplementary Fig. S28 ), where the factor α substitutes for explicit dynamic mechanisms. By tuning α, the physical ESN achieves normalized root‑mean‑square errors (NRMSE) of 0.165, 0.28, and 0.206% along the x‑, y‑, and z‑axes, respectively. The forgetting factor ensures continuity in reservoir states across time (Supplementary Figs. S29 – S31 ). The second approach directly embeds temporal attributes by preprocessing the data through the dynamic response of the physical device before training (Fig. 4a ). Here, training proceeds with a forgetting factor of 1, equivalent to a static mapping setup. The device’s dynamic response performs a nonlinear transformation on the Lorenz attractor, and the resulting time‑varying Ids serves as input to the ESN (Fig. 4b–d ). By varying the sampling intervals during preprocessing, different dynamic features can be extracted for prediction. Though the reservoir states under this method appear discrete, they retain encoded temporal information (Supplementary Fig. S32 ). This preprocessing strategy yields NRMSE values of 0.165%, 0.28%, and 0.206% on the x‑, y‑, and z‑axes, respectively (Fig. 4e )—performance that closely matches, and in some dimensions slightly exceeds, that of the forgetting‑factor method. Consequently, the system achieves low prediction error and generates smooth Lorenz attractor trajectories (Fig. 4f ). Fig. 4. Lorenz-attractor prediction task realized by dynamic-response-based physical reservoir computing. Open in a new tab a Schematic of the approach: device dynamics encode and forecast the Lorenz attractor. b–d Nonlinear mapping of Lorenz-attractor trajectories along x, y and z axes using the intrinsic dynamics of the single device (V ds = 1 V, the Lorentz attractor curve is normalized and then converted to a 0 to + 10 V V g signal using DAC). e Prediction NRMSE versus pulse-interval variation. f Forecasted Lorenz-attractor time traces under different operating modes. The system is further applied to predicting typhoon-center imagery and trajectories. Using grayscale image sequences and longitude–latitude data from Typhoon “Kanu” (2023), the ambipolar ESN achieved the lowest image-prediction NRMSE of 10.88% and maintained balanced dynamic activity between positive and negative states even when the reservoir was scaled to 2000 × 2000 (Supplementary Figs. S33 – S35 ). For trajectory prediction, the ambipolar ESN with a 500 × 500 reservoir best fitted the test data, notably attaining a stable latitude‑prediction error in the range of 8.34–8.46% (Supplementary Fig. S36 ). These results collectively demonstrate that the ambipolar ESN excels in multi‑dimensional spatiotemporal forecasting, owing to its rich state diversity and the synergistic interplay between positive and negative activations, which together enhance dynamic‑system modeling. To further validate the multimodal classification capability of our system, we employ electrocardiogram (ECG) signals. The signals are converted into time–frequency images via short-time Fourier transform (STFT) 45 , 46 to form the dataset (Fig. 5a ). This process yields two complementary data modalities: the original 1D waveforms and the derived 2D spectrograms. Both modalities are processed independently by the reservoir, and the resulting states are classified through a shared readout layer. Recognition can be performed on either modality alone (signal mode or image mode) or on their combined representations (dual mode) (Fig. 5b ). When processed by the physical ESN, the signal and image modalities produce distinct reservoir state patterns (Fig. 5c ). Combining these states enhances overall discriminability. We systematically evaluate ESN performance across reservoir scales. The ambipolar ESN consistently achieves the highest accuracy in all modes: signal mode (82.80–88.38%), image mode (84.06–87.46%), and dual mode (87.62–89.77%), with accuracy generally improving as the reservoir size increases until saturation beyond 500 × 500 (Fig. 5d ). The p ‑type ESN delivers stable, scale‑insensitive performance, particularly in dual mode where accuracy remains nearly constant (87.39–88.35%). In contrast, the n‑type ESN peaks in dual mode at 87.81% with a 200 × 200 reservoir, beyond which performance degrades, revealing limited scalability. These results underscore the ambipolar ESN as the most robust and scalable architecture for multimodal processing, achieving superior accuracy while maintaining consistent performance scaling. This advantage stems from its more balanced reservoir state distribution across positive and negative activations (Supplementary Figs. S37 – 39 ). Fig. 5. Multimodal identification task by our ESN. Open in a new tab a ECG Signal, Short-time Fourier transform image of Channel 1, Short-Time Fourier Transform image of Channel 2. b Schematic diagram of multimodal ECG recognition. c Reservoir state of the ECG image and the ECG signal. d Accuracy of multimodal recognition of ESNs with reservoir size = 100 × 100 to 1000 × 1000. Finally, a task utilizing the Mosei dataset for multimodal recognition was conducted. The audio, text, and visual features from Mosei were processed using a pooling mechanism and then classified with SVM. For tri-modal classification (Supplementary Fig. S40 ), the accuracy without the ambipolar ESN is 71.33%, whereas with the ambipolar ESN, it improves to 76.07%. This demonstrates that the ambipolar ESN significantly enhances the accuracy of multimodal recognition. For dual-modal classification, the accuracies for visual/audio, visual/text, and text/audio combinations are 70.86, 76.46, and 76.01%, respectively. For monomodal classification, the accuracies for text, audio, and visual modalities are 76.48, 71.01, and 71.03%, respectively. Using a 10 nm alumina dielectric layer allows the transistor array to operate at lower voltages, yet introduces process-related challenges in uniformity and endurance (Supplementary Fig. S41 ). When the pulse interval is increased sufficiently, the dynamic response of the array gradually transitions toward static behavior (Supplementary Fig. S42 ). In image recognition tasks, relying solely on the final current value after dynamic mapping proves inadequate. For instance, in 5‑bit English letter recognition using only the steady‑state response as the training feature, we observed misclassification patterns (Supplementary Figs. S43 , S44 ), attributable to the loss of transient information. Compared to binarized inputs, conventional 4‑bit digital encoding yields lower accuracy due to feature compression (Supplementary Fig. S45 ). However, by fully utilizing the temporal 4‑bit response data, recognition accuracy improves from 84.92% to 87.94% (Supplementary Fig. S46 ), confirming that dynamic nonlinearity enhances binarized image classification. To further boost accuracy, we introduced sparse projection and activation‑based feature extraction, combining these derived features with raw data for training. This strategy led to additional gains in recognition performance (Supplementary Figs. S47 – S55 ). The array architecture can physically implement such operations through its tunable linear regions and inherent nonlinearity, supporting efficient and accurate image recognition. Nevertheless, while the system exhibits strong recognition capability, further accuracy improvements demand increased computational resources. Moreover, time‑series prediction currently relies on a forgetting factor, implying that the computational model must support online iteration. These limitations can be mitigated by leveraging the device’s intrinsic dynamic response. The additional features generated through dynamics enhance physical ESN performance without digital overhead. In time‑series prediction, preprocessing data through the device’s dynamic equation mapping enables single‑pass inference, eliminating the need for online iterative updates. In summary, ambipolar transistors, with their balanced linear characteristics and bounded output response, are particularly suitable for emulating both the linear weighting of synapses and the nonlinear activation of neurons in a reservoir layer. Since the weights in the reservoir layer are fixed and require no training, such devices represent a promising candidate for constructing recurrent neural networks with inherent dynamic constraints, highlighting considerable research potential. In contrast, p -type and n -type transistors exhibit limitations in building high-performance reservoirs due to their weaker capability to bound signal amplitudes. Under static operation, the system has demonstrated strong potential in tasks such as image recognition, time-series prediction, and multimodal classification. In dynamic mode, by leveraging the intrinsic transient response of the device, temporal correlations can be embedded into the data, leading to further improvements in image recognition performance. Particularly for time-series forecasting, the dynamic mechanism can replace the forgetting factor used in conventional ESNs, transforming iterative online computation into single-step forward processing, thereby enhancing overall computational efficiency. Methods Fabrication of the device and array The bottom gate electrodes were fabricated via physical vapor deposition. Aluminum oxide was subsequently deposited as the gate dielectric using atomic layer deposition. For the semiconductor layers, ZnO was prepared by dissolving the precursor in ammonia to form a uniform solution, while P3HT was dissolved in toluene and thoroughly mixed by stirring and heating to ensure homogeneity. The substrate was placed on a spin coater, and the ZnO precursor solution was dispensed onto it. Spin-coating was performed sequentially at 1500 rpm, 2000 rpm, and 2500 rpm, each for 30 s. After spin-coating, the substrate was annealed at 150 °C for 40 min to remove the solvent and crystallize the ZnO thin film. The P3HT solution was then applied onto the ZnO layer, with the volume carefully controlled to ensure film uniformity, followed by spin-coating at 1500 rpm for 30 s. The resulting film was annealed at 100 °C for 20 min. Device and array structures were defined by covering the thin film with a mask and depositing electrodes via physical vapor deposition. All fabrication steps were conducted in a nitrogen atmosphere. The completed array was encapsulated by covering it with a quartz glass sheet and sealing the edges with UV-curable adhesive. Control of the array Specific rows are selected by applying a feature signal that connects the input voltage to the drain terminals of all transistors in the chosen row via a digital-to-analog converter (DAC). Simultaneously, specific columns are addressed by applying a column selection signal to the gate terminals of all transistors in the selected column. This addressing scheme enables individual control of each unit within the array, facilitating the implementation of matrix multiplication and activation functions. Measurement of the array A custom program was used to sequentially access and test each node in the array, thereby evaluating the operational performance and stability of the array. To measure the output, a resistor with a known value (1 MΩ) was connected in series at the output terminal of the transistor array, allowing the output current to be converted into a voltage signal for further analysis. Device characterization The electrical properties of the device were tested using a semiconductor analyzer (Keysight B2902B). The device was first cutting using Focused Ion Beam Scanning Electron Microscopy (Helios 5 UC 03040707), and then the cross-sectional morphology was taken by Field Emission Transmission Electron Microscopy (JEOL-F200). The surface morphology of the device was characterized using an Atomic Force Microscope (Cypher S/Oxford Instruments Asylum Research). Equation The formula for the status of the reservoir is given as follows: U t = U t − 1 ( 1 − α ) + α tanh ( X t W in + U t − 1 W res ) 1 Where U (t) , U (t−1) , α (alpha), X t , W in , W res are represent for reservoir state at time step t, reservoir state at the previous time step at t−1, forgetting rate, input at time step, input weight matrix and reservoir weight matrix. The solution method for the reading layer uses the linear regression formula: W out = ( U T U ) − 1 ( U T y label ) 2 W out is the readout layer matrix, U is the reservoir state, and y label is the label value. The formula for the Lorentz attractor is given as follows: d x d t = σ ( y − x ) d y d t = x ( ρ − z ) − y d z d t = x y − β z 3 σ, ρ and β are respectively the Prandtl number, the Rayleigh number and the geometric factor. In addition, the gradient descent algorithm is used in the prediction of the typhoon center, with Adam as the optimizer. The loss function used is the Root Mean Square Error (RMSE), and its formula is as follows: RMSE = 1 n ∑ i = 1 n ( y label − y predict ) 2 4 y predict is the predicted result. The formula for NRMSE is given as follows. NRMSE = 1 n ∑ 1 n ( y label − y predict ) 2 y max − y min 5 y max , y min are max value of y label , min value of y label , respectively. Dataset The MNIST handwritten digits and fashion datasets are provided by the National Institute of Standards and Technology. Each dataset consists of 60,000 images for training and 10,000 images for testing. The Lorenz attractor curve comprises 10,000 points, with the first 5000 points used as the training set and the remaining 5000 points as the test set. In this task, the input consists of points 1 to 4998, while the target includes points 3 to 5000. The images of the typhoon center and its movement trajectory are derived from Typhoon Khanun in 2023. The ECG dataset is sourced from the Massachusetts Institute of Technology - Beth Israel Hospital, and includes 17 categories based on their annotations. To ensure dataset validity, ECG signals with more than 100 samples are selected, resulting in a total of 11,811 samples. The first 10,000 samples are used for training, and the remainder are used for testing. The CMU-MOSEI dataset, developed by the Multimodal Communication and Language Analysis Lab at Carnegie Mellon University (CMU), uses binary classification labels: 0 for values less than 0, and 1 for values greater than or equal to 0. In addition, the support vector machine employed in the final section of the study uses a linear kernel function, with the C parameter set to 0.0001. Networks The English alphabet recognition task and Lorenz attractor prediction are achieved through a 16 × 16 ambipolar transistor array, with read/write operations on the array controlled by DACs and an ADC via STM32. MNIST handwritten digits, the Fashion dataset, Lorenz attractor prediction, ECG data classification, and multimodal recognition with the MOSEI dataset are implemented using Python + PyTorch simulations. The simulation is performed by an operational model of the device. The specific physical ESN network structure is provided in the Supplementary Table S2 . Supplementary information Supplementary Information (6MB, pdf) Transparent Peer Review file (6.2MB, pdf) Acknowledgements We acknowledge grants from RSC Sustainable Laboratories Grant (L24-8215098370, to Y.Z.), Guangdong Basic and Applied Basic Research Foundation (2024B1515040002, to S.T.H.), Guangdong Science and Technology Department (Grant No. MS202500156, to Y.Z.), Department of Education of Guangdong Province (Grant No. 2025ZDZX3024, to Y.Z.), Science, Technology and Innovation Commission of Shenzhen Municipality (Grant No. JCYJ20250604181254072, to Y.Z.), State Key Laboratory of Radio Frequency Heterogeneous Integration (Independent Scientific Research Program No. 2024010, to Y.Z.), the Hong Kong Research Grants Council, Young Collaborative Research Grant (C5001-24, to S.T.H.), Research Institute for Smart Energy (U-CDC9, to S.T.H.) and NTUT-SZU Joint Research Program (Grant No. 2026002, to Y.Z.). Author contributions W.-M. Zhong: Conceptualization, Methodology, Writing - Original Draft. W. Zhang: Investigation. Y.-X. Zeng: Data Curation. J.-Y. Zhao: Investigation. Z.-Q. Jia: Investigation. G. Ding: Data Curation. S.-T. Han: Supervision, Data Curation. V.A.L. Roy: Writing - Review & Editing. Y. Zhou: Conceptualization, Project administration, Funding acquisition, Writing - Review & Editing. Peer review Peer review information Nature Communications thanks Yao Guo, Hongseok Oh, and Jianhua Yang for their contribution to the peer review of this work. A peer review file is available. Data availability The physical data generated in this study have been deposited in the Zenodo database [10.5281/zenodo.18475506]. Code availability The code is available at Zenodo [10.5281/zenodo.18459291]. 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. Supplementary information The online version contains supplementary material available at 10.1038/s41467-026-70171-2. References 1. Kumar, S., Williams, R. S. & Wang, Z. Third-order nanocircuit elements for neuromorphic engineering. Nature 585 , 518–523 (2020). [ DOI ] [ PubMed ] [ Google Scholar ] 2. Yu, J. R. et al. Bioinspired mechano-photonic artificial synapse based on graphene/MoS 2 heterostructure. Sci. Adv. 7 , eabd9117 (2021). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 3. Torrejon, J. et al. Neuromorphic computing with nanoscale spintronic oscillators. Nature 547 , 428–431 (2017). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 4. Merolla, P. A. et al. A million spiking-neuron integrated circuit with a scalable communication network and interface. Science 345 , 668–673 (2014). [ DOI ] [ PubMed ] [ Google Scholar ] 5. Roy, K., Jaiswal, A. & Panda, P. Towards spike-based machine intelligence with neuromorphic computing. Nature 575 , 607–617 (2019). [ DOI ] [ PubMed ] [ Google Scholar ] 6. Fuller, E. J. et al. Parallel programming of an ionic floating-gate memory array for scalable neuromorphic computing. Science 364 , 570–574 (2019). [ DOI ] [ PubMed ] [ Google Scholar ] 7. Xu, M. Y. et al. Reconfigurable Neuromorphic Computing: Materials, Devices, and Integration. Adv. Mater. 35 , 2301063 (2023). [ DOI ] [ PubMed ] [ Google Scholar ] 8. Choi, S. et al. Emerging memristive artificial synapses and neurons for energy-efficient neuromorphic computing. Adv. Mater. 32 , 2004659 (2020). [ DOI ] [ PubMed ] [ Google Scholar ] 9. Ding, G. L. et al. Porous crystalline materials for memories and neuromorphic computing systems. Chem. Soc. Rev. 52 , 7071–7136 (2023). [ DOI ] [ PubMed ] [ Google Scholar ] 10. Zhao, J. H. et al. Memristors based on NdNiO 3 nanocrystals film as sensory neurons for neuromorphic computing. Mater. Horiz. 10 , 4521–4531 (2023). [ DOI ] [ PubMed ] [ Google Scholar ] 11. Cao, Z. L. et al. Memristor-based neural networks: a bridge from device to artificial intelligence. Nanoscale Horiz. 8 , 716–745 (2023). [ DOI ] [ PubMed ] [ Google Scholar ] 12. Meng, J. L. et al. Ionic diffusive nanomemristors with dendritic competition and cooperation functions for ultralow voltage neuromorphic computing. ACS Nano 18 , 9150–9159 (2024). [ DOI ] [ PubMed ] [ Google Scholar ] 13. Wang, W. S. et al. Biodegradable oxide neuromorphic transistors for neuromorphic computing and anxiety disorder emulation. ACS Appl. Mater. Interfaces 15 , 47670–47648 (2023). [ DOI ] [ PubMed ] [ Google Scholar ] 14. Lv, Z. Y. et al. Semiconductor quantum dots for memories and neuromorphic computing systems. Chem. Rev. 120 , 3941–4006 (2020). [ DOI ] [ PubMed ] [ Google Scholar ] 15. Qiu, E. R. et al. Reconfigurable cascaded thermal neuristors for neuromorphic computing. Adv. Mater. 36 , 2306818 (2023). [ DOI ] [ PubMed ] [ Google Scholar ] 16. Chen, S. E., Giridharagopal, R. & Ginger, D. S. Artificial neuron transmits chemical signals. Nat. Mater. 22 , 416–418 (2023). [ DOI ] [ PubMed ] [ Google Scholar ] 17. Chen, K. et al. Organic optoelectronic synapse based on photon-modulated electrochemical doping. Nat. Photon. 17 , 629–637 (2023). [ Google Scholar ] 18. van de Burgt, Y. et al. A non-volatile organic electrochemical device as a low-voltage artificial synapse for neuromorphic computing. Nat. Mater. 16 , 414–418 (2017). [ DOI ] [ PubMed ] [ Google Scholar ] 19. Hadiyal, K. et al. Bio-inspired artificial synapse for neuromorphic computing based on NiO nanoparticle thin film. Sci. Rep. 13 , 7481 (2023). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 20. Dai, M. et al. Author correction: realization of tunable artificial synapse and memory based on amorphous oxide semiconductor transistor. Sci. Rep. 11 , 12657 (2021). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 21. Islam, M. M. et al. Optoelectronic synapse using monolayer MoS 2 field effect transistors. Sci. Rep. 10 , 21870 (2020). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 22. Kim, S. et al. Neurorobotic approaches to emulate human motor control with the integration of artificial synapse. Sci. Adv. 8 , eabo3326 (2022). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 23. Qi, Z. Y. et al. Physical reservoir computing based on nanoscale materials and devices. Adv. Func. Mater. 33 , 2306149 (2023). [ Google Scholar ] 24. Schranghamer, T. F., Oberoi, A. & Das, S. Graphene memristive synapses for high precision neuromorphic computing. Nat. Commun. 11 , 5474 (2020). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 25. Liu, K. et al. An optoelectronic synapse based on α-In2Se3 with controllable temporal dynamics for multimode and multiscale reservoir computing. Nat. Electron 5 , 761–773 (2022). [ Google Scholar ] 26. Gao, C. et al. Toward grouped-reservoir computing: organic neuromorphic vertical transistor with distributed reservoir states for efficient recognition and prediction. Nat. Commun. 15 , 740 (2024). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 27. Namiki, W. et al. Experimental demonstration of high-performance physical reservoir computing with nonlinear interfered spin wave multidetection. Adv. Intell. Syst. 5 , 2300228 (2023). [ Google Scholar ] 28. Yamazaki, Y. & Kinoshita, K. Photonic physical reservoir computing with tunable relaxation time constant. Adv. Sci. 11 , 2304804 (2024). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 29. Benfenati, V. et al. A transparent organic transistor structure for bidirectional stimulation and recording of primary neurons. Nat. Mater. 12 , 672–680 (2013). [ DOI ] [ PubMed ] [ Google Scholar ] 30. Verzelli, P., Alippi, C. & Livi, L. Echo state networks with self-normalizing activations on the hyper-sphere. Sci. Rep. 9 , 13887 (2019). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 31. Ebato, Y. et al. Impact of time-history terms on reservoir dynamics and prediction accuracy in echo state networks. Sci. Rep. 14 , 8631 (2024). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 32. Zhang, B. L. et al. Novel gray orthogonal echo state network integrating the process mechanism for dynamic soft sensor development. Ind. Eng. Chem. Res. 60 , 14955–14967 (2021). [ Google Scholar ] 33. Hasegawa, H. et al. Parallel and deep reservoir computing using semiconductor lasers with optical feedback. Nanophotonics 12 , 869–881 (2023). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 34. Tanaka, G. et al. Recent advances in physical reservoir computing: A review. Neural Netw. 115 , 100–123 (2019). [ DOI ] [ PubMed ] [ Google Scholar ] 35. Yan, M. et al. Emerging opportunities and challenges for the future of reservoir computing. Nat. Commun. 15 , 2056 (2024). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 36. Li, X. et al. The fabrication and optimization of OTFT formaldehyde sensors based on Poly(3-hexythiophene)/ZnO composite films. Sci. China Technol. Sci. 56 , 1877–1882 (2013). [ Google Scholar ] 37. Han, S. J. et al. Band alignment effects at the metal electrode interface of poly(3-hexylthiophene):zinc oxide hybrid photovoltaics. Appl. Phys. Lett. 102 , 081607 (2013). [ Google Scholar ] 38. Wang, J. et al. Hierarchical ZnO nanosheet-nanorod architectures for fabrication of poly(3-hexylthiophene)/ZnO hybrid NO 2 sensor. ACS Appl. Mater. & Interfaces 8 , 8600–8607 (2016). [ DOI ] [ PubMed ] [ Google Scholar ] 39. Niazi, S. et al. Training deep Boltzmann networks with sparse Ising machines. Nat. Electron 7 , 610–619 (2024). [ Google Scholar ] 40. Valleti, M. et al. Physics and chemistry from parsimonious representations: image analysis via invariant variational autoencoders. Npj Comput. Mater. 10 , 183 (2024). [ Google Scholar ] 41. Bergomi, M. G. et al. Towards a topological–geometrical theory of group equivariant non-expansive operators for data analysis and machine learning. Nat. Mach. Intell. 1 , 423–433 (2019). [ Google Scholar ] 42. Stewart, I. The Lorenz attractor exists. Nature 406 , 948–949 (2000). [ DOI ] [ PubMed ] [ Google Scholar ] 43. Ghaffari, A. Image compression-encryption method based on two-dimensional sparse recovery and chaotic system. Sci. Rep. 11 , 369 (2021). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 44. Yan, M., Jie, J. & Zhang, P. Chaotic systems with variable indexs for image encryption application. Sci. Rep. 12 , 19585 (2022). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 45. Bi, K. et al. Accurate medium-range global weather forecasting with 3D neural networks. Nature 619 , 533–538 (2023). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 46. Hirokawa, J. et al. The influence of electrocardiogram (ECG) filters on the heights of R and T waves in children. Sci. Rep. 12 , 13279 (2022). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Associated Data This section collects any data citations, data availability statements, or supplementary materials included in this article. Supplementary Materials Supplementary Information (6MB, pdf) Transparent Peer Review file (6.2MB, pdf) Data Availability Statement The physical data generated in this study have been deposited in the Zenodo database [10.5281/zenodo.18475506]. The code is available at Zenodo [10.5281/zenodo.18459291]. Articles from Nature Communications are provided here courtesy of Nature Publishing Group ACTIONS View on publisher site PDF (2.9 MB) Cite Collections Permalink PERMALINK Copy RESOURCES Similar articles Cited by other articles Links to NCBI Databases Cite Copy Download .nbib .nbib Format: AMA APA MLA NLM Add to Collections Create a new collection Add to an existing collection Name your collection * Choose a collection Unable to load your collection due to an error Please try again Add Cancel Follow NCBI NCBI on X (formerly known as Twitter) NCBI on Facebook NCBI on LinkedIn NCBI on GitHub NCBI RSS feed Connect with NLM NLM on X (formerly known as Twitter) NLM on Facebook NLM on YouTube National Library of Medicine 8600 Rockville Pike Bethesda, MD 20894 Web Policies FOIA HHS Vulnerability Disclosure Help Accessibility Careers NLM NIH HHS USA.gov Back to Top