On-Skin Artificial Intelligence via Supramolecular Polymer Memtransistors

On-Skin Artificial Intelligence via Supramolecular Polymer Memtransistors | 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 On-Skin Artificial Intelligence via Supramolecular Polymer Memtransistors Jin Young Oh, Ngoc Thanh Phuong Vo, Eun Joo Yoo, Kyu Ho Jung, and 5 more This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-7993782/v1 This work is licensed under a CC BY 4.0 License Status: Posted Version 1 posted You are reading this latest preprint version Abstract On-skin artificial intelligence (AI) demands hardware that couples skin-like mechanics with efficient, real-time computation on noisy, spatiotemporal biosignals. We introduce a supramolecular polymer memtransistor that unifies intrinsic stretchability, autonomous self-healing and low-power neuromorphic dynamics in a single material platform. The device integrates a supramolecular elastomer matrix with a p-n heterojunction semiconductor to realize charge trapping-driven short-term plasticity, high on/off ratios (>103) and tight device-to-device uniformity (σ/μ = 5.25%). Operating energies span 0.29 fJ-1.8 nJ per event, approaching the lower bound of biological synapses while retaining reliable control of synaptic weights. Arrays (7 × 7) serve as a physical reservoir for on-device reservoir computing, achieving >99% accuracy in spoken-digit recognition and robust emotion recognition (74% under 30% biaxial strain; 71% after self-healing), all maintained during 30% biaxial deformation and after autonomous recovery from deliberate damage. Beyond classification, recursive multi-step forecasting with online learning stably models chaotic dynamics with normalized RMSE ≲ 0.02, sustaining accurate long-horizon predictions. These results establish supramolecular polymer memtransistors as a materials-driven route to elastic, damage-tolerant and energy-efficient neuromorphic electronics that perform AI inference and prediction directly on the body, enabling bio-integrated systems for speech, affect and complex physiological time-series analysis. Physical sciences/Materials science/Materials for devices/Electronic devices Physical sciences/Mathematics and computing/Computer science Figures Figure 1 Figure 2 Figure 3 Figure 4 Introduction The future of skin electronics hinges on their capability to seamlessly integrate with human skin, not only by mimicking its mechanical properties for comfort, durability, and long-term functionality, but also by enabling autonomous, real-time artificial intelligence (AI) computing directly on our bodies. This advanced form of skin electronics could revolutionize bio-integrated applications by autonomously analyzing physiological data, recognizing complex biological signals such as speech patterns and emotional states, and facilitating interactive human–machine interfaces and adaptive soft robotic systems. Recent developments in polymer electronic materials have significantly enhanced mechanical compliance, enabling superior integration onto the dynamically deformable surfaces of biological tissues. However, despite these advancements, the electrical performance of polymer-based organic devices remains insufficient compared to conventional inorganic silicon-based electronics, thereby restricting their potential to perform real-time, high-efficiency AI processing tasks on-skin. Alongside the material challenge, the current computational architecture also represents a fundamental barrier to achieving effective on-skin AI. The von Neumann architecture, which underpins modern electronics, relies on physically separate memory and processing units, leading to frequent and energy-intensive data transfer between components. This results in significant inefficiencies, latency, and high energy consumption, particularly problematic for wearable applications requiring continuous, real-time data processing and low power consumption. To overcome these intrinsic constraints, recent attention has focused on biologically inspired neuromorphic computing architectures, which provide efficient alternatives capable of handling dynamic, chaotic, and spatiotemporal signal analysis with reduced computational overhead. [1,2] Among various neuromorphic architectures, reservoir computing (RC), [3] a specialized form of recurrent neural network (RNN), [4,5] has emerged as particularly advantageous for real-time, on-device AI applications. Unlike conventional deep learning architectures that typically require extensive training resources, RC leverages a dynamic reservoir to intrinsically encode complex temporal signals into linearly separable, high-dimensional states. This intrinsic capability significantly reduces computational overhead, making RC highly suitable for efficient, adaptive processing of nonlinear, temporal, and chaotic signals in skin electronics. To effectively realize RC-based neuromorphic computing, memtransistors, which simultaneously integrate memory and transistor functionalities, have emerged as ideal hardware components. [6-8] Memtransistors inherently exhibit short-term plasticity (STP), a crucial synaptic property allowing transient storage and dynamic modulation of conductance states in response to temporal signals, mimicking biological synaptic behavior. [9-12] Depending on design and conditions, they can also support other forms of plasticity, offering flexibility for neuromorphic computing. In addition, the memtransistors also provide enhanced controllability in active-matrix array configurations due to the presence of additional terminals, making them particularly well-suited for scalable, real-time, and adaptive AI processing in on-skin electronic systems. [13,14] Despite these clear advantages, current memtransistor implementations still predominantly suffer from limited on/off current ratios, high power consumption, and substantial variability, hindering their ability to achieve highly accurate and energy-efficient reservoir computing (RC). Furthermore, these devices are typically composed of rigid and brittle inorganic materials, severely limiting their compatibility with skin-integrated systems. To realize practical on-skin AI electronics, it is essential to develop memtransistors that synergistically combine high performance neuromorphic computing with skin-like stretchability and autonomous self-healing ability. [15] Recently, supramolecular polymer materials have emerged as promising candidates for skin-integrated electronics, leveraging intrinsic self-healing and exceptional stretchability enabled by dynamic, reversible supramolecular interactions. [16-18] Nevertheless, despite significant advances in supramolecular polymer-based electronics, previous studies have exclusively demonstrated conventional transistor configurations without integrated neuromorphic functions. Thus, the development of memtransistors explicitly engineered for neuromorphic computing remains unexplored, yet constitutes a critical step toward adaptive, mechanically robust, damage-resistant, and intelligent on-skin electronic systems. In this study, we introduce a supramolecular polymer memtransistor specifically engineered to enable neuromorphic AI computing directly on the skin. Leveraging intrinsic short-term plasticity (STP) and dynamic supramolecular interactions, our memtransistor simultaneously achieves highly energy efficient neuromorphic processing (0.29 fJ - 1.8 nJ) with high on/off ratio exceeding 10 3 on/off ratio and narrow variability (σ/μ = 5.25%). By integrating our device within a reservoir computing framework, we demonstrate its exceptional capability to efficiently process nonlinear and chaotic spatiotemporal signals. Moreover, this device consistently maintains stable electrical performance even under a 30% biaxial strain after experiencing mechanical damage and subsequent autonomous self-healing. We further demonstrate a stable 7×7 active-matrix memtransistor array, confirming the feasibility of scalable integration for neuromorphic systems. To the best of knowledge, we present the first demonstration of chaotic signal prediction applied to human sensory information processing using a supramolecular polymer-based neuromorphic memtransistor, highlighting its potential in advanced tasks related with time series information processing such as spoken digit classification (Audio MNIST), [19] emotion recognition (RAVDESS), [20] and chaotic time-series forecasting. This integrative approach represents an emerging demonstration of a supramolecular polymer memtransistor tailored explicitly for neuromorphic computing in bio-integrated systems, highlighting a new paradigm that combines mechanical resilience, low-power consumption, and adaptive AI processing capabilities. Design and Characterization of Supramolecular Polymer Memtransistor The schematic of a supramolecular polymer memtransistor active-matrix array (7 × 7) designed for seamless integration with human skin, alongside the detailed structure of an individual device, is illustrated in Fig. 1a . Each memtransistor consists entirely of supramolecular polymer composites that encompass semiconductor, dielectric, and electrode layers. These devices function via a charge-trapping mechanism, facilitated by precise energy-level alignment engineered at the p-n heterojunction ( Fig. 1b , Supplementary Figs. 1-3 ). [21-23] We employed poly(2,5-bis(2-octyldodecyl)-3,6-di(thiophen-2-yl) diketopyrrolo [3,4-c] pyrrole-1,4-dione-alt-thieno [3,2-b]thiophen) (DPPT-TT) as the p-type semiconductor, and poly[[1,2,3,6,7,8-hexahydro-2,7-bis(2-octyldodecyl)-1,3,6,8-dioxobenzo[lMn][3,8]phenanthroline-4,9-diyl][2,2'-bithiophene]-5,5'-diyl] (N2200) as the n-type semiconductor for p-n junction based semiconducting layer. A supramolecular polymeric elastomer (SPE) composed of poly(dimethylsiloxane)-4,4′-methylenebis(phenyl urea)-isophorone bisurea (PDMS–MPU 6 –IU 4 ) served as a stretchable and self-healing elastomer matrix for the memtransistor components, owing to its excellent elasticity, autonomous self-healing capability, high compatibility with all components (semiconductor, electrode, and dielectric) ( Supplementary Figs. 4-8 ). [18] The supramolecular p-n junction semiconductor film was prepared by transfer-printing of the p-type and n-type semiconductor layers, each individually blended at the molecular level with the SPE matrix at optimized mass ratios (DPPT-TT:SPE = 3:7, N2200:SPE = 5:5). These mass ratios were systematically balanced to achieve desirable electrical characteristics, including effective threshold voltage (V th ) shifts indicative of trap capacity, high on-current ( Fig. 1c ), and robust mechanical resilience characterized by significant stretchability and autonomous self-healing ( Fig. 1d , Supplementary Figs. 9,10 ). Both supramolecular semiconductor blends exhibited distinctive nanoweb-like morphologies with clear phase separation, enabling efficient dissipation of mechanical stress through reversible supramolecular interactions and sustaining conductive pathways under biaxial strain and after autonomous healing ( Fig. 1e,f , Supplementary Fig. 11-14 , and Supplementary Note 5). The memtransistor array demonstrated highly consistent and uniform transfer characteristics across all 49 devices, with an average on-current of approximately 5 × 10 -7 A, low leakage currents (~10 -10 A), high on/off current ratios (~10⁴) without kink effect ( Supplementary Figs. 15,16 , and Supplementary Note 1). The devices exhibited anticlockwise hysteresis with wide V th shifts (from 2.3 to 37.5 V) during gate voltage (V G ) sweeps ( Fig. 1g ), maintaining reliability at low drain voltages (V D = -1 V) and a stable V th shift (~37 V) ( Supplementary Figs. 17,18 ). To establish the origin of reservoir characteristics, control devices lacking the n-type semiconductor layer were fabricated and evaluated. Unlike p-n junction-based devices, these control devices exhibited no change in hysteresis and trap capacity characteristics, confirming that reservoir functionality specifically originates from energy-level alignment within the p-n junction interface ( Fig. 1b and Supplementary Figs. 19-21 ). Electron-trapping activation energy (E A ), quantified to be approximately 27.93 meV, confirmed electron trapping at the heterojunction as the primary mechanism for observed synaptic behavior ( Fig. 1h, bottom and Supplementary Figs. 22,23 and Supplementary Note 2). Capacitance-voltage (C-V) analysis of metal-insulator-semiconductor (MIS) capacitors further validated enhanced charge trapping at the semiconductor-dielectric interface, corroborating charge trapping at the p-n junction ( Supplementary Fig. 24 ). Additionally, the effective Schottky barrier height ( ϕ S ), derived from temperature-dependent output current measurements, elucidated charge injection and transport dynamics modulated by V G (Supplementary Note 3, 4). Increasing V G effectively reduced ϕ S , enhancing carrier injection into the active channel ( Fig. 1h, top ). Application of V G induced hole accumulation at the semiconductor-insulator interface, resulting in energy band bending and reducing the source-semiconductor barrier, subsequently elevating output current. Electron trapping by the n-type layer further enhanced hole concentration, contributing to reservoir retention. After gate-pulse termination, trapped electrons recombined with channel holes, gradually diminishing output current and realizing short-term plasticity (STP) functionality. Different gate biases injected varied charge concentrations into the active layer, enriching carrier dynamics essential for spatiotemporal mapping in reservoir computing. The intrinsic stretchability and self-healing capabilities of the supramolecular polymer memtransistors were systematically characterized through morphological and electrical analyses. Each component of the devices (semiconductor, electrode, and dielectric layers) exhibited intrinsic stretchability, maintaining structural integrity and electrical performance under mechanical strains of up to 30% (uniaxial and biaxial) as well as autonomous self-healing capabilities ( Fig. 1i,j ). Leveraging these insights, we assessed the overall mechanical resilience of the complete memtransistor devices. The devices consistently maintained their initial transfer characteristics, such as on-current (~10 -7 A) and ΔV th shift (~36.5 V), even under 30% biaxial strains, demonstrating compatibility with the mechanical compliance requirements for human skin ( Figs. 1k,l and Supplementary Figs. 25,26 , Table S1 ). The autonomous self-healing capability of the memtransistors was evaluated by deliberately cutting the device channel using a surgical blade, creating microscale damage (a defined 3 µm-wide gap). The damaged device regained functionality after approximately 12 hours, and its transfer characteristics fully recovered to the original pristine state within three days of autonomous healing ( Figs. 1m,n and Supplementary Fig. 27 ). Neuromorphic Synaptic Behaviors of Supramolecular Polymer Memtransistors The synaptic characteristics of the supramolecular polymer memtransistors were systematically evaluated to demonstrate their suitability for reservoir computing (RC)-based neuromorphic systems ( Fig. 2 and Supplementary Fig. 28 ). The excitatory postsynaptic current (EPSC) responses induced by single electrical pulses were measured by applying presynaptic voltage pulses (+40 V) with pulse widths ranging from 0.75 to 9 s to the gate electrode, while maintaining a constant drain voltage (V D = -1 V). As shown in Fig. 2a , the EPSC amplitude increased progressively with pulse width due to enhanced electron trapping at the p-n junction interface. The dependence of the EPSC on/off ratio and decay time (τ) on pulse width was investigated under various device states, including pristine condition, 30% uniaxial and biaxial strains, and after autonomous self-healing ( Fig. 2b and Supplementary Figs. 29-31 ). Moreover, the EPSC amplitude increased proportionally with the presynaptic pulse amplitude, varied from 10 V to 90 V, and remained stable with narrow variation even under mechanical deformation (up to 30% strain) and following autonomous self-healing ( Fig. 2c,d and Supplementary Fig. 32-34 ). Short-term synaptic plasticity was quantified using the paired-pulse facilitation (PPF) index, defined as the amplitude ratio (A 2 /A 1 ) of two successive EPSC peaks. Two consecutive presynaptic pulses (+40 V) with pulse intervals (Δt pre ) ranging from 0.375 to 72 s were applied at a fixed drain voltage (V D = -1 V) ( Fig. 2e and Supplementary Fig. 35-38 ). The PPF index gradually decreased from 335% to 135% as Δt pre increased from 0.375 to 72 s. Both the magnitude and overall trend of the PPF index remained consistent even under substantial mechanical deformation (30% uniaxial and biaxial strains) and after autonomous self-healing ( Fig. 2f ). The EPSC responses were further characterized under trains of presynaptic pulses (1 to 50 successive pulses, +40 V, 3 s each), demonstrating consistent increases in EPSC amplitude and on/off ratio up to 3000, with minimal variation under pristine, strained (30% uniaxial and biaxial), and autonomously healed conditions ( Fig. 2g,h and Supplementary Fig. 39-47 ). To confirm synaptic stability and robustness, repeated endurance cycling was conducted through long-term potentiation (LTP) and depression (LTD) measurements. Fig. 2i and Supplementary Fig. 48 presents stable and repeatable transitions between LTP and LTD states during continuous alternating presynaptic spike cycles (250 cycles total), each comprising 20 potentiation spikes (+40 V, 3 s) followed by 20 depression spikes (-0.5 V, 3 s). The current transitions remained highly stable, exhibiting negligible variation over a total of 10,000 pulses, indicating that the device possesses reliable short-term memory characteristics under various interval times at -1 V D ( Fig. 2j ). These results indicate that this memristor exhibits robust electrical durability a high on/off ratio, and reliable controllability in synaptic weight modulation, combined with intrinsic stretchability and self-healing ability not previously demonstrated. Furthermore, the energy consumption of these artificial synapses spans a broad range (from 0.29 fJ to 1.8 nJ), with the lowest achievable values approaching those of biological synapses in the human brain (1–100 fJ), significantly surpassing previously reported transistor-based artificial synaptic devices ( Fig. 2k , Supplementary Fig. 49 , and Table S2 ). [24-31] Reservoir Computing of Supramolecular Polymer Memtransistors To demonstrate the applicability of our supramolecular polymer memtransistors for efficient temporal neuromorphic processing, we implemented a physical reservoir computing (PRC) framework that leverages their inherent STP. This hardware-based approach is highly suitable for spatiotemporal signal processing in wearable electronics platforms, enabling on-skin classification tasks that rely on sequential pattern recognition, such as speech and emotion recognition. We fabricated a 7 × 7 active-matrix array of memtransistors to construct the hardware reservoir ( Fig. 3a ). Each memtransistor (M₁, M₂, ..., Mₙ) functions as a dynamic reservoir node, receiving spike train inputs and generating time-varying current outputs that collectively form the reservoir state. Figure 3b presents the schematic overview of the PRC architecture. In this system, a time-dependent input signal u(t) is projected into a high-dimensional dynamic state space x(t) via fixed input weights W in . Prior to entering the reservoir layer, the sequential input signals are processed through a masking step, enabling richer spatiotemporal representations within the reservoir. It has been widely adopted in PRC to enhance the separability of input patterns and improve classification accuracy. [32] The nonlinear reservoir, realized through a network of memtransistors, transforms these signals into high-dimensional dynamics. This transformation leverages internal fading memory effects, which are critical for temporal signal processing. Only the output weights W out are trained using supervised learning to generate the final output y(t) that matches the desired target y target (t) , simplifying training complexity and computational demands. To validate the dynamic encoding ability of individual memtransistors, we applied temporally modulated binary input sequences ("0000" to "1111") using 40 V for "1" and 1 V for "0", as shown in Fig. 3c . The devices exhibited characteristic current increases during "1" pulses and gradual decay during "0" intervals, clearly demonstrating fading memory and STP essential for reservoir computing. These dynamic responses remained stable under mechanical strain (30% uniaxial and biaxial) and even after mechanical damage followed by autonomous self-healing ( Supplementary Fig. 50 ), confirming both the functional stability and mechanical resilience. Additionally, these characteristics were well preserved across the array, with the data exhibiting excellent uniformity across 49 devices, with low device-to-device variation (σ/μ = 5.25%), ensuring scalable and reliable neuromorphic performance ( Fig. 3d and Supplementary Fig. 51 ). To benchmark the reservoir’s capability for sequential classification tasks involving speech, we employed two representative datasets: AudioMNIST for digit recognition and RAVDESS for emotion recognition. Figure 3e illustrates preprocessing of raw speech into time-frequency representations, utilizing cochleagram transformations for digit recognition tasks (emphasizing temporal precision) [33] and mel-spectrogram transformations for emotion recognition tasks (capturing perceptual acoustic features). [34] These spectrograms were converted into binary spike trains, effectively encoding phonetic and emotional nuances for reservoir-based classification (Supplementary Note 6, 7). Figure 3f shows that classification accuracy on the AudioMNIST dataset increased with mask length, reaching over 99% accuracy using all 49 devices. This trend is attributed to enhanced feature extraction and frequency resolution enabled by longer masks, allowing for more refined recognition of pitch modulation and syllabic structures ( Supplementary Fig. 52 ). The classification accuracy remained stable (~99%) even under severe mechanical deformations (30% uniaxial and biaxial) and even post self-healing, confirming robust functionality under mechanical stress ( Fig. 3g and Supplementary Fig. 53 ). Evaluating more complex tasks, emotion recognition achieved high accuracy both under mechanical strain (73.6%) and after healing (70.83%), exhibiting stable misclassification patterns ( Fig. 3h,i and Supplementary Fig. 54 ). Such performance compares favorably with previously reported studies (typically ~65–75%), [35,36] affirming the mechanical and functional integrity of our PRC system. We also tested the system on the MNIST dataset to validate the spatial computation capability, achieving >97% classification accuracy ( Supplementary Fig. 55 ), further confirming the versatility of our memtransistor-based reservoir architecture. Recursive Chaotic Time-Series Prediction for On-Skin Artificial Intelligence While classification tasks such as speech and emotion recognition effectively demonstrate our system's capability in temporal pattern recognition, these problems remain relatively structured and limited in complexity. In contrast, biological signals frequently encountered in wearable applications including neural, cardiac, and respiratory activities often exhibit strong nonlinearity, high noise levels, and even chaotic behavior under certain physiological states, thereby posing a significantly greater challenge for predictive modeling. [37] To evaluate the applicability of our system to such complex and unstructured temporal data, we adopted a canonical benchmark for chaotic dynamics: the Lorenz attractor. Originally developed to model atmospheric convection, the Lorenz system exhibits strong nonlinearity and extreme sensitivity to initial conditions, making it an ideal testbed for assessing the temporal memory and dynamic processing capabilities of PRC systems ( Supplementary Fig. 56 ). Prediction tasks employing the Lorenz system typically follow two approaches: one-step-ahead and multi-step-ahead forecasting. The one-step-ahead method yields accurate single-step predictions by relying on ground truth data at each inference step yet inherently limits autonomous forecasting beyond a single time step ( Supplementary Fig. 57 ). To overcome this limitation, we implemented recursive multi-step-ahead prediction coupled with online training, aiming to suppress errors accumulation and enhance long-term prediction stability. Figure 4a schematically illustrates the recursive forecasting paradigm employed in our system. At each cycle, the reservoir predicts the system state at time t+1 using historical data from steps 1 to t. This prediction is then recursively fed back into the reservoir to forecast future states autonomously, without access to ground truth inputs. This feedback-based architecture facilitates extended, self-sustained prediction over long time horizons (Supplementary Note 8). Utilizing the pristine memtransistor array as the reservoir, we trained the system on Lorenz trajectories up to time step 32,000. Beyond this point, the system executed autonomous recursive forecasting cycles comprising 100 steps of prediction followed by 50 steps of online retraining. Due to the inherently chaotic nature of the Lorenz system, even minor deviations in prediction rapidly amplify beyond acceptable limits ( Supplementary Fig. 58 ). Nevertheless, our reservoir reliably preserved the attractor’s characteristic butterfly-shaped orbit throughout the recursive process, indicating that prediction errors remained within a controllable range ( Fig. 4b ). A magnified view of the z-axis prediction emphasizes the close alignment between predicted and true values, particularly within complex oscillatory regions. Figure 4c presents the reconstructed 3D trajectory, demonstrating that predicted values (purple dotted line) accurately replicate the characteristic butterfly-shaped orbit of the Lorenz attractor compared with the ground truth (gray solid line). This highlights the reservoir’s proficiency in modeling nonlinear chaotic dynamics in phase space. This further substantiates that online learning effectively mitigates prediction drift, maintaining high fidelity during extended forecasts ( Supplementary Fig. 59 ). To assess mechanical robustness, we repeated the Lorenz prediction experiments under conditions of uniaxial strain, biaxial strain, and after autonomous self-healing ( Supplementary Fig. 60 ). Figure 4d summarizes the normalized root mean square error (NRMSE) across these varying mechanical conditions, consistently demonstrating low error values (~0.02 or below). This quantitative evaluation further validates the mechanical resilience and long-term predictive stability of our stretchable PRC platform for on-skin chaotic time-series forecasting. To further demonstrate the robustness of our reservoir across diverse dynamical environments, we extended trajectory forecasting experiments to other chaotic attractors. These tasks were performed using the same online training structure as for the Lorenz attractor. As shown in Fig. 4e,f , the predicted trajectories closely match the ground truth for both the Dadras-Momeni attractor and the Halvorsen attractor, thereby confirming the universality of our device as a physical reservoir ( Supplementary Figs. 61,62 ). In this case, NRMSE values of approximately 0.01374 and 0.03814 were obtained, confirming that accurate predictions were achieved in all cases, which demonstrates the ability to sustain accurate modeling of highly nonlinear dynamics over extended horizons, highlighting a materials-based pathway toward bio-integrated AI systems capable of both precise pattern recognition and long-term adaptive prediction. [38, 39] Methods Materials: Bis(3-aminopropyl) terminated poly(dimethylsiloxane) (H 2 N-PDMS-NH 2 , M n = 5000) was purchased from Gelest Inc.. Poly(2,5-bis(2-octyldodecyl)−3,6-di(thiophen-2-yl)diketopyrrolo[3,4-c]pyrrole-1,4-dione-alt-thieno[3,2-b]thiophen) (DPPT-TT) (M w /M n = 100k/42k, PDI 100k, PDI < 3.0) semiconducting polymers were purchased from Derthon. Poly(dimethylsiloxane) (PDMS, Sylgard 184) and its cross-linker were purchased by Dow Corning. The PDMS was cured with a ratio of 8:1 (base/cross-linker, w/w) at 65 °C overnight for the transfer printing stamp. SEBS H1062 (S/EB weight ratio of 18/82) was supported by Asahi Kasei company. Ag (silver, 99.99%, 3–5 mm granule) was purchased from SY SCIENCE. The Trichloro(octadecyl)silane (OTS) solution and diiodomethane were purchased by Sigma Aldrich. The Ethyl-amine (Et 3 N), anhydrous chloroform, and toluene were purchased by Sigma Aldrich. Methanol (MeOH) solution was purchased by SAMCHUN. All chemicals and materials were used without any purification. Synthesis of supramolecular polymeric elastomer (SPE) PDMS-MPU 6 -IU 4 PDMS-MPU 6 -IU 4 polymer was synthesized according to previously reported methods. [18] H 2 N-PDMS-NH 2 (30 g, M n = 5000, 1 eq) was dissolved in 120 mL Chloroform at 0 °C under nitrogen atmosphere. Et 3 N (3 mL) was added to the solution of H 2 N-PDMS-NH 2 and stirred for 1 h. After that, a mixture solution of 4,4’– Methylenebis(phenyl isocyanate) (MPU) (0.9 g, 6 eq) and Isophorone diisocyanate (IU) (0.54 g, 4 eq) in CHCl 3 was added dropwise. The solution was then allowed to warm to room temperature and stirred for 4 days. After reaction, MeOH (5 mL) was added and stirred for 30 min to remove the remaining isocyanate. Then, the solution was concentrated to ½ of its volume. 18 mL MeOH was poured into the mixture solution to give a viscous liquid. After settling for 30 min, the upper clear solution was then decanted. 30 mL CHCl 3 was added to dissolve the product. The dissolution-precipitation-decantation process was repeated three times for purifying and the final product was subjected in ambient condition to remove the solvent and trace of Et 3 N. Device fabrications and characterizations Fully stretchable and self-healing memtransistor preparation: The solution for the p-type semiconductor was prepared by dissolving DPPT-TT (0.21 wt%) and SPE (0.49 wt%) with a total of 0.7 wt%, in anhydrous chloroform. The solution for the n-type semiconductor was prepared by dissolving N2200 (0.0075 wt%) and SPE (0.0075 wt%) with a total of 0.015 wt%, in anhydrous chloroform. Both solutions were stirred at 50 °C for 4 hours. The solutions were spun on an OTS-treated SiO 2 /Si wafer at 1000 rpm for p-type semiconductor (film thickness: 100 nm) and 1500 rpm for n-type semiconductor in 1 min after filtration with a PTFE-D (0.2 µm) filter (film thickness: 20 nm). Both p-type and n-type semiconducting films were then annealed at 80 o C for 30 min. All above processes were carried out under an N 2 atmosphere in a glove box with extremely low levels of moisture (H 2 O < 0.01 parts per million (ppm)) and oxygen (O 2 < 0.01 ppm). The SPE substrate solution was spin-coated onto an OTS-treated SiO 2 /Si wafer using the SPE solution (50 mg/mL in chloroform) at 2000 rpm for 1 minute without any heat treatment. The resulting SPE substrate was transferred onto the SEBS substrate to be stretchable and elastic. 100 nm thick Ag gate electrode was thermally evaporated onto the SPE substrate at a speed of 0.2 nm/s under high vacuum conditions (below 5.0×10 -6 torr). Then, the SPE dielectric (60 mg/ml, 1000 rpm in 1 min on OTS-treated SiO 2 , 1.5 µm thick), n-type semiconducting film (in a 5:5 ratio of N2200 and SPE), and p-type semiconducting film (in a 3:7 ratio of DPPT-TT and SPE) were sequentially transferred onto the gate electrode. Finally, 100 nm thick Ag source/drain electrodes were thermally evaporated onto the top p-type semiconducting film. Characterization The electrical, hysteresis, and synaptic characteristics of the devices were measured under ambient conditions by a probe station, which is connected to KEITHLEY 4200. The strain-stress curve was obtained by a force tester (AND, MCT-2150; strain rate: 100 mm/min). UV-Vis-Nir spectra were measured using a spectrophotometer (JASCO, V-770). For the cutting process, a surgical blade (Feather, No. 25) was used. Surface structures and current images were obtained with atomic force microscopy (AFM; Bruker MultiMode 8-HR) under ambient conditions. DMT modulus mappings were measured using PeakForce quantitative nanoscale mechanical (QNM) AFM. Optical images were obtained with an optical microscope (OM; Leica DM4 M). The thicknesses of the film were obtained with an ellipsometer (WS WONWOO S-TRC2000). X-ray photoelectron spectra depth profiling was obtained with XPS equipment (Thermo Electron, K-Alpha). Surface energy and contact angle were obtained by PHOENIX-MT(T). Work function of p-type and n-type semiconductor was measured using Riken Keiki (Japan)/AC-2 with UV intensity 50 nW and 200 nW. Declarations Acknowledgments This research was supported by the Basic Science Research Program through the National Research Foundation of Korea (NRF) funded by the Ministry of Science and ICT (MSIT) (RS-2021-NR061555, RS-2020-NR049601), GRRC program of Gyeonggi province (GRRCKYUNGHEE2023-B03) and the Korea Institute for the Advancement of Technology (KIAT) and the Ministry of Trade, Industry & Energy (MOTIE) of the Republic of Korea (RS-2024-00434908, RS-2021-KI002513, RS-2025-25435993, RS-2024-00406182, and RS-2025-02305609). Author contributions N.T.P.V., E.J.Y., S.H.L. and J.Y.O. conceived the study and N.T.P.V., E.J.Y., S.H.L. and J.Y.O. designed the experiments. N.T.P.V., E.J.Y. conducted all experiments. N.T.P.V., E.J.Y., K.H.J., M.W.J., S.H.P., T.A.N., H.R.C., S.H.L., J.Y.O. analyzed and discussed the data. N.T.P.V., E.J.Y., S.H.L. and J.Y.O. wrote the manuscript. Competing interests Authors declare that they have no competing interests. Data availability Data are available on request. Correspondence and requests for materials should be addressed to S.H.L and J.Y.O. Supplementary information The online version contains supplementary material available at https://doi.org/##.#### References Zidan, M. A., Strachan, J. P. & Lu, W. D. The future of electronics based on memristive systems. Nat. Electron. 1 , 22–29 (2018). Horowitz, M. in 2014 IEEE international solid-state circuits conference digest of technical papers ( ISSCC ). ( IEEE , 2014), pp. 10–14. Lukoševičius, M. & Jaeger, H. Reservoir computing approaches to recurrent neural network training. Comput. Sci. Rev. 3 , 127–149 (2009). Hopfield, J. J. Neural networks and physical systems with emergent collective computational abilities. Proceedings of the national academy of sciences 79 , 2554–2558 (1982). Werbos, P. J. Backpropagation through time: what it does and how to do it. Proceedings of the IEEE 78 , 1550–1560 (2002). Liang, X. et al. Physical reservoir computing with emerging electronics. Nat. Electron. 7 , 193–206 (2024). Tanaka, G. et al. Recent advances in physical reservoir computing: A review. Neural Networks 115 , 100–123 (2019). Yan, M. et al. Emerging opportunities and challenges for the future of reservoir computing. Nat. Commun. 15 , 2056 (2024). Du, C., Cai, F. et al. Reservoir computing using dynamic memristors for temporal information processing. Nat. Commun. 8 , 2204 (2017). Liu, Z., Zhang, Q., Xie, D. et al. Interface-type tunable oxygen ion dynamics for physical reservoir computing. Nat. Commun. 14 , 7176 (2023). Moon, J., Ma, W. et al. Temporal data classification and forecasting using a memristor-based reservoir computing system. Nat. Electron. 2 , 480–487 (2019). Wu, X. et al. Ultralow-power optoelectronic synaptic transistors based on polyzwitterion dielectrics for in-sensor reservoir computing. Sci. Adv. 10 , eadn4524 (2024). Choi, S. et al. 3D-integrated multilayered physical reservoir array for learning and forecasting time-series information. Nat. Commun. 15 , 2044 (2024). Park, S. O., Jeong, H., Park, J., Bae, J. & Choi, S. Experimental demonstration of highly reliable dynamic memristor for artificial neuron and neuromorphic computing. Nat. Commun. 13 , 2888 (2022). Patil, C. S. et al. Neuromorphic devices for electronic skin applications. Mater. Horiz. 12 , 2045-2088 (2025). Oh, J. Y., Lee, Y., Lee, T. W. Skin‐Mountable Functional Electronic Materials for Bio‐Integrated Devices. Adv. Healthc. Mater. 13 , 2303797 (2024). Wang, S. & Urban, M. W. Self-healing polymers. Nat. Rev. Mater. 5 , 562-583, (2020). Vo, N. T. P., Nam, T. U. et al. Autonomous self-healing supramolecular polymer transistors for skin electronics. Nat Commun. 15 , 3433 (2024). Becker, S., Vielhaben, J. et al. Audiomnist: Exploring explainable artificial intelligence for audio analysis on a simple benchmark. J. Frankl. Inst. 361 , 418–428 (2024). Livingstone, S. R. & Russo, F. A. The Ryerson Audio-Visual Database of Emotional Speech and Song (RAVDESS): A dynamic, multimodal set of facial and vocal expressions in North American English. PLoS One 13 , e0196391 (2018). Kim, M. H., Jeong, M. W. et al. Mechanically robust stretchable semiconductor metallization for skin-inspired organic transistors. Sci. Adv. 8 , eade2988 (2022). Kim, J. S., Jeong, M. W. et al. Intrinsically Stretchable Subthreshold Organic Transistors for Highly Sensitive Low‐Power Skin‐Like Active‐Matrix Temperature Sensors. Adv. Funct. Mat. 34 , 2305252 (2024). Jung, K. H., Hyun, J. et al. A biocompatible elastomeric organic transistor for implantable electronics. Nat. Electron. 8 , 1-13, (2025). Wu, X., Wang, S. et al. Wearable in-sensor reservoir computing using optoelectronic polymers with through-space charge-transport characteristics for multi-task learning. Nat. Comm. 14 , 468 (2023). Gao, C. et al. Toward grouped-reservoir computing: organic neuromorphic vertical transistor with distributed reservoir states for efficient recognition and prediction. Nat. Comm. 15 , 740 (2024). Zhong, Y. et al. Dynamic memristor-based reservoir computing for high-efficiency temporal signal processing. Nat. Comm. 12 , 408, (2021). Liu, K. et al. An optoelectronic synapse based on α-In 2 Se 3 with controllable temporal dynamics for multimode and multiscale reservoir computing. Nat. Electron. 5 , 761-773 (2022). Kim, J., Park, E. C., Shin, W. et al. Analog reservoir computing via ferroelectric mixed phase boundary transistors. Nat. Comm. 15 , 9147 (2024). Liu, Z., Zhang, Q., Xie, D. et al. Interface-type tunable oxygen ion dynamics for physical reservoir computing. Nat. Commun. 14 , 7176 (2023). Appeltant, L. et al. Information processing using a single dynamical node as complex system. Nat. Commun. 2 , 468 (2011). Wang, S., Chen, X. et al. An organic electrochemical transistor for multi-modal sensing, memory and processing. Nat. Electron. 6 , 281-291 (2023). Appeltant, L. et al. Information processing using a single dynamical node as complex system. Nat. Commun. 2 , 468 (2011). Yoo, S., Chae, S. et al. Efficient data processing using tunable entropy-stabilized oxide memristors. Nat. Electron. 7 , 426–436 (2024). Zhang, T., Feng, G., Liang, J. & An, T. Acoustic scene classification based on Mel spectrogram decomposition and model merging. Appl. Acoust. 182 , 108258 (2021). Luna-Jiménez, C. et al. Multimodal Emotion Recognition on RAVDESS Dataset Using Transfer Learning. Sensors 21 , 7665 (2021). Issa, D., Demirci, M. F. & Yazici, A. Speech emotion recognition with deep convolutional neural networks. Biomed. Signal Process 59 , 101894 (2020). Wu, G. Q. et al. Chaotic Signatures of Heart Rate Variability and Its Power Spectrum in Health, Aging and Heart Failure. PLoS One 4 , e4323 (2009). Dadras, S. & Momeni, H. R. A novel three-dimensional autonomous chaotic system generating two, three and four-scroll attractors. Phys. Lett. A 373 , 3637–3642 (2009). Sprott, J. C. Elegant Chaos: Algebraically Simple Chaotic Flows ( World Scientific , Singapore, 2010). Additional Declarations There is NO Competing Interest. Supplementary Files 20251031Supplementarymaterialsjy.pdf Supplementary information Cite Share Download PDF Status: Posted 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-7993782","acceptedTermsAndConditions":true,"allowDirectSubmit":true,"archivedVersions":[],"articleType":"Article","associatedPublications":[],"authors":[{"id":557687533,"identity":"0bba89ad-ee76-4d4e-9d3b-343bb47f4230","order_by":0,"name":"Jin Young Oh","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAAq0lEQVRIiWNgGAWjYBACxgYGNoYPBw7AOERqYZxBkhYgYGPmIUkL84zkY49tztxJbGA//IBx5h5iHDYjLd0458azxAaeNAPGDc+I0pJjJp3z4XBiA0MOA+ODA8RqsQBp4X9DihaGG0AtEkBbNhClpedZmmTPmcPGbRLPDA7OIEaLYXvyMYkfxw7L9vMnP3zYQ5SWCQkQBhsQE6OBgUGenzh1o2AUjIJRMJIBAChPPmnww5LmAAAAAElFTkSuQmCC","orcid":"https://orcid.org/0000-0003-2260-9960","institution":"Kyung Hee University","correspondingAuthor":true,"prefix":"","firstName":"Jin","middleName":"Young","lastName":"Oh","suffix":""},{"id":557687534,"identity":"5056941b-f2d6-43c9-9ffe-71e6bff2dabe","order_by":1,"name":"Ngoc Thanh Phuong Vo","email":"","orcid":"","institution":"Kyung Hee University","correspondingAuthor":false,"prefix":"","firstName":"Ngoc","middleName":"Thanh Phuong","lastName":"Vo","suffix":""},{"id":557687535,"identity":"6223954f-5197-4ff6-a4a3-e066302a8ab2","order_by":2,"name":"Eun Joo Yoo","email":"","orcid":"","institution":"Kyung Hee University","correspondingAuthor":false,"prefix":"","firstName":"Eun","middleName":"Joo","lastName":"Yoo","suffix":""},{"id":557687536,"identity":"e503389d-ba08-4a0e-8761-a9dbaf0d49da","order_by":3,"name":"Kyu Ho Jung","email":"","orcid":"","institution":"Kyung Hee University","correspondingAuthor":false,"prefix":"","firstName":"Kyu","middleName":"Ho","lastName":"Jung","suffix":""},{"id":557687537,"identity":"ad47a53e-3999-4656-8a07-0a47c5658115","order_by":4,"name":"Min Woo Jeong","email":"","orcid":"","institution":"Kyung Hee University","correspondingAuthor":false,"prefix":"","firstName":"Min","middleName":"Woo","lastName":"Jeong","suffix":""},{"id":557687538,"identity":"019178d0-ddcb-4e5b-9b31-b9a899b5d064","order_by":5,"name":"Seon Hoo Park","email":"","orcid":"","institution":"Kyung Hee University","correspondingAuthor":false,"prefix":"","firstName":"Seon","middleName":"Hoo","lastName":"Park","suffix":""},{"id":557687539,"identity":"52e2b594-fc51-4d12-a06e-6d658cdb4bd3","order_by":6,"name":"Thuy An Nguyen","email":"","orcid":"","institution":"Kyung Hee University","correspondingAuthor":false,"prefix":"","firstName":"Thuy","middleName":"An","lastName":"Nguyen","suffix":""},{"id":557687540,"identity":"c3666b9b-94a4-42dc-a3cc-da91d406cdcc","order_by":7,"name":"Hye Rin Chang","email":"","orcid":"","institution":"Kyung Hee University","correspondingAuthor":false,"prefix":"","firstName":"Hye","middleName":"Rin","lastName":"Chang","suffix":""},{"id":557687541,"identity":"548f755c-8831-405a-b3ca-de9c1e55e252","order_by":8,"name":"Seung Hwan Lee","email":"","orcid":"","institution":"Kyung Hee University","correspondingAuthor":false,"prefix":"","firstName":"Seung","middleName":"Hwan","lastName":"Lee","suffix":""}],"badges":[],"createdAt":"2025-10-31 02:56:15","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-7993782/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-7993782/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":97863101,"identity":"9c129e6f-7bbf-4ed5-947f-860d0ff5a641","added_by":"auto","created_at":"2025-12-10 09:10:14","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":2061779,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eDesign and characterization of supramolecular polymer memtransistor. a, \u003c/strong\u003ePhotograph of memtransistor device (left), inset optical microscope (OM) image, electrical symbol (top right), and schematic illustration of the device structure (bottom right). \u003cstrong\u003eb,\u003c/strong\u003e Chemical structures of active materials: p-type semiconductor (DPPT-TT, top left), n-type semiconductor (N2200, top right), and supramolecular polymeric elastomer (SPE, PDMS-MPU\u003csub\u003e6\u003c/sub\u003e-IU\u003csub\u003e4\u003c/sub\u003e, bottom).\u0026nbsp; \u003cstrong\u003ec,\u003c/strong\u003e Threshold voltage shift (ΔV\u003csub\u003eth\u003c/sub\u003e) and on-current (I\u003csub\u003eon\u003c/sub\u003e, measured at V\u003csub\u003eG\u003c/sub\u003e = -50 V, V\u003csub\u003eD\u003c/sub\u003e = -10 V) with various blending ratios of p-type and n-type semiconductors within SPE. \u003cstrong\u003ed,\u003c/strong\u003e Variations in self-healing efficiency and crack-onset strain depending on the blending ratio of p-type and n-type semiconductors with SPE. Atomic force microscopy (AFM) height images (top) and X-ray photoelectron spectroscopy (XPS) depth profiles (bottom) for \u003cstrong\u003ee,\u003c/strong\u003e p-type semiconductor layer and \u003cstrong\u003ef,\u003c/strong\u003e n-type trapping layer. \u003cstrong\u003eg,\u003c/strong\u003e Electrical hysteresis characteristics in transfer curve of the memtransistor devices under varying gate voltage (V\u003csub\u003eG\u003c/sub\u003e) sweep ranges at constant drain voltage (V\u003csub\u003eD\u003c/sub\u003e = -10 V). \u003cstrong\u003eh,\u003c/strong\u003e Arrhenius correlation plots at different temperatures, extracted activation energies and delay time constants for transistor devices (bottom), and corresponding potential barrier structure for charge trapping at the p-n heterojunction (top). OM and AFM height images of \u003cstrong\u003ei,\u003c/strong\u003e p-type semiconductor layers and \u003cstrong\u003ej,\u003c/strong\u003e n-type trapping layers under stretched (left) and self-healed (right) conditions. Scale bars: OM images, 25 µm; AFM images, 1 µm. \u003cstrong\u003ek,\u003c/strong\u003e Transfer curves of memtransistor in pristine, 30% uniaxial stretched, and 30% biaxial stretched states (V\u003csub\u003eD\u003c/sub\u003e = -10 V). \u003cstrong\u003el,\u003c/strong\u003e Corresponding ΔV\u003csub\u003eth\u003c/sub\u003e and drain current (ΔI\u003csub\u003eD\u003c/sub\u003e) changes in pristine, 30% uniaxially stretched, and 30% biaxially stretched states. \u003cstrong\u003em,\u003c/strong\u003e Transfer curves of the memtransistor measured as a function of healing time (V\u003csub\u003eD\u003c/sub\u003e = -10 V). \u003cstrong\u003en,\u003c/strong\u003e Corresponding ΔVth shifts and ΔI\u003csub\u003eD\u003c/sub\u003e with increasing healing time.\u003c/p\u003e","description":"","filename":"1.png","url":"https://assets-eu.researchsquare.com/files/rs-7993782/v1/e5c5ff6d0b6bbd2d8a45c1d2.png"},{"id":97863102,"identity":"07864fd0-9889-4af9-a678-9cd1708055d6","added_by":"auto","created_at":"2025-12-10 09:10:15","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":699245,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eSynaptic characteristics of supramolecular memtransistor. a,\u003c/strong\u003e Single-pulse-induced excitatory postsynaptic current (EPSC) of a pristine device measured at various pulse widths (gate voltage, V\u003csub\u003eG\u003c/sub\u003e =40 V). \u003cstrong\u003eb,\u003c/strong\u003e Extracted relaxation time constant (τ, top) and on/off ratio (bottom) as functions of pulse width for devices under pristine conditions, 30% uniaxial and biaxial strains, and post self-healing. \u003cstrong\u003ec,\u003c/strong\u003e Single-pulse-induced EPSC responses of a pristine device at varying pulse voltages, t\u003csub\u003epulse\u003c/sub\u003e = 3 s. \u003cstrong\u003ed,\u003c/strong\u003e τ (top) and on/off ratio (bottom) as functions of pulse voltage for devices under pristine conditions, 30% uniaxial strain, 30% biaxial strain, and after self-healing. \u003cstrong\u003ee,\u003c/strong\u003e Representative EPSC triggered by paired pulses (each pulse: 40 V, 3 s). A\u003csub\u003e1\u003c/sub\u003e and A\u003csub\u003e2\u003c/sub\u003e are EPSC values of the first and second pulses, respectively, separated by 3 s. \u003cstrong\u003ef,\u003c/strong\u003e PPF index change with pristine, 30% uniaxial and biaxial stretched, and after healing devices with different ∆t\u003csub\u003epre\u003c/sub\u003e values (t\u003csub\u003epre\u003c/sub\u003e = t\u003csub\u003epulse\u003c/sub\u003e + t\u003csub\u003eread\u003c/sub\u003e).\u0026nbsp; \u003cstrong\u003eg,\u003c/strong\u003e EPSC triggered by various number of pulses (each pulse: 40 V, 3 s). \u003cstrong\u003eh,\u003c/strong\u003e τ value (top) and on/off ratio (bottom) with respect to the number of pulses for the devices without strain, with 30% uniaxial and biaxial strain, and after self-healing.\u003cstrong\u003e i,\u003c/strong\u003e Endurance characteristics of the memtransistor during consecutive long-term potentiation (LTP, 20 pulses at 40 V, 3 s each) and long-term depression (LTD, 20 pulses at -0.5 V, 3 s each) cycles over 10,000 total pulses. \u003cstrong\u003ej,\u003c/strong\u003e Short-term plasticity characteristics exhibited by the transistor device as a function of varying pulse interval times. All measurements in Figure 2 are conducted at V\u003csub\u003eD\u003c/sub\u003e = -1 V. \u003cstrong\u003ek,\u003c/strong\u003e Comparison plot of on/off ratio versus energy consumption, benchmarked against previously reported synaptic transistor devices.\u003c/p\u003e","description":"","filename":"2.png","url":"https://assets-eu.researchsquare.com/files/rs-7993782/v1/48a38f0e73c99b9d2a3c1f40.png"},{"id":97863100,"identity":"e78e9c4b-297f-4466-9b21-0c12b749db5a","added_by":"auto","created_at":"2025-12-10 09:10:14","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":1053487,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eReservoir Computing for Temporal Data Processing: Simulation and Experimental Validation. a,\u003c/strong\u003ePhotograph of the 7 × 7 active-matrix memtransistor array on a flexible substrate. \u003cstrong\u003eb,\u003c/strong\u003e Schematic of the physical reservoir computing (PRC) architecture. The input layer converts temporal signals into spike trains, which are then projected into a high-dimensional dynamic space via the memtransistor-based reservoir layer. The readout layer linearly maps these dynamic states to the desired output using trained weights. \u003cstrong\u003ec,\u003c/strong\u003e Output current responses from the memtransistor reservoir to different 4-bit input pulse sequences (“0000”– “1111”), representing distinct reservoir states. \u003cstrong\u003ed,\u003c/strong\u003eCorresponding 3D contour plot of device-to-device variation for selected 4-bit input codes (“0000,” “1000,” “1100,” “1110,” and “1111,”), showing minimal variability and high uniformity across the array. \u003cstrong\u003ee,\u003c/strong\u003e Preprocessing steps for speech recognition: raw audio signals from AudioMNIST and RAVDESS datasets are converted into 2D spectrograms using cochleagrams (AudioMNIST) or mel-spectrograms (RAVDESS) and subsequently digitized into binary spike sequences for input into the reservoir. \u003cstrong\u003ef,\u003c/strong\u003e Classification accuracy on the AudioMNIST dataset as a function of reservoir mask length (M = 9, 16, 25, and 49), showing improved performance with longer masks due to enhanced spatiotemporal representation. \u003cstrong\u003eg,\u003c/strong\u003e AudioMNIST classification accuracy under various mechanical states of the device: pristine, 30% uniaxial strain, 30% biaxial strain, and after healing. \u003cstrong\u003eh,\u003c/strong\u003e Confusion matrix for RAVDESS classification using a device subjected to 30% biaxial strain. \u003cstrong\u003ei,\u003c/strong\u003e Confusion matrix for RAVDESS classification using a device after tearing and subsequent self-healing.\u003c/p\u003e","description":"","filename":"3.png","url":"https://assets-eu.researchsquare.com/files/rs-7993782/v1/2bd2c292c320d17ea94e4f60.png"},{"id":97863103,"identity":"d1b8e1c2-d16b-48e8-b348-1832d64b4da0","added_by":"auto","created_at":"2025-12-10 09:10:15","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":1224563,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eMulti-step Forecasting of Lorenz Attractor via Reservoir Computing with Supramolecular Polymer Memtransistors. a,\u003c/strong\u003e Schematic illustrating the recursive input scheme for multi-step-ahead forecasting. The reservoir is sequentially fed with previous time step values (\u003cem\u003ex(t)\u003c/em\u003e, \u003cem\u003ey(t)\u003c/em\u003e, \u003cem\u003ez(t)\u003c/em\u003e) to predict (\u003cem\u003ex(t+1)\u003c/em\u003e, \u003cem\u003ey(t+1)\u003c/em\u003e, \u003cem\u003ez(t+1)\u003c/em\u003e), which are recursively used as inputs for subsequent predictions. \u003cstrong\u003eb,\u003c/strong\u003e One-dimensional time series forecasting of the x, y, and z-axes via online training. The purple line represents the training data, while the blue line indicates the predicted values. Gray-shaded regions denote intervals where weights are updated via online training, and unshaded regions correspond to inference-only periods. Forecasting was performed in cycles of 50-step training and 100-step prediction. In the enlarged view of the z-axis prediction, the solid line indicates ground truth, and the dotted line shows the predicted output. \u003cstrong\u003ec,\u003c/strong\u003eThree-dimensional reconstructed trajectory using the predicted values from (b), showing close agreement with the ground truth and preserving the characteristic shape of the Lorenz attractor. \u003cstrong\u003ed,\u003c/strong\u003e Normalized root mean square error (NRMSE) of forecasts under various mechanical states of the memtransistor reservoir (pristine, 30% uniaxial strain, 30% biaxial strain, and self-healed). \u003cstrong\u003ee,\u003c/strong\u003e Predicted trajectories closely match the ground truth for both the Dadras-Momeni attractor and \u003cstrong\u003ef,\u003c/strong\u003e the Halvorsen attractor, thereby confirming the universality of our device as a physical reservoir.\u003c/p\u003e","description":"","filename":"4.png","url":"https://assets-eu.researchsquare.com/files/rs-7993782/v1/29b8e7fc175c2852636d3d95.png"},{"id":105563004,"identity":"647fc11f-f793-4e5d-aeb5-95cb6e14da11","added_by":"auto","created_at":"2026-03-27 12:45:36","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":6848858,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-7993782/v1/541958f6-a602-40de-b7f8-64c4f8a45678.pdf"},{"id":97899177,"identity":"94a9c3be-0689-4033-8616-65ec996043b2","added_by":"auto","created_at":"2025-12-10 15:41:55","extension":"pdf","order_by":1,"title":"","display":"","copyAsset":false,"role":"supplement","size":5945234,"visible":true,"origin":"","legend":"Supplementary information","description":"","filename":"20251031Supplementarymaterialsjy.pdf","url":"https://assets-eu.researchsquare.com/files/rs-7993782/v1/12b9f3ea757e3e12f73de2ee.pdf"}],"financialInterests":"There is \u003cb\u003eNO\u003c/b\u003e Competing Interest.","formattedTitle":"On-Skin Artificial Intelligence via Supramolecular Polymer Memtransistors","fulltext":[{"header":"Introduction","content":"\u003cp\u003eThe future of skin electronics hinges on their capability to seamlessly integrate with human skin, not only by mimicking its mechanical properties for comfort, durability, and long-term functionality, but also by enabling autonomous, real-time artificial intelligence (AI) computing directly on our bodies. This advanced form of skin electronics could revolutionize bio-integrated applications by autonomously analyzing physiological data, recognizing complex biological signals such as speech patterns and emotional states, and facilitating interactive human\u0026ndash;machine interfaces and adaptive soft robotic systems. Recent developments in polymer electronic materials have significantly enhanced mechanical compliance, enabling superior integration onto the dynamically deformable surfaces of biological tissues. However, despite these advancements, the electrical performance of polymer-based organic devices remains insufficient compared to conventional inorganic silicon-based electronics, thereby restricting their potential to perform real-time, high-efficiency AI processing tasks on-skin.\u003c/p\u003e\n\u003cp\u003eAlongside the material challenge, the current computational architecture also represents a fundamental barrier to achieving effective on-skin AI. The von Neumann architecture, which underpins modern electronics, relies on physically separate memory and processing units, leading to frequent and energy-intensive data transfer between components. This results in significant inefficiencies, latency, and high energy consumption, particularly problematic for wearable applications requiring continuous, real-time data processing and low power consumption. To overcome these intrinsic constraints, recent attention has focused on biologically inspired neuromorphic computing architectures, which provide efficient alternatives capable of handling dynamic, chaotic, and spatiotemporal signal analysis with reduced computational overhead.\u003csup\u003e[1,2]\u003c/sup\u003e\u003c/p\u003e\n\u003cp\u003eAmong various neuromorphic architectures, reservoir computing (RC),\u003csup\u003e[3]\u003c/sup\u003e a specialized form of recurrent neural network (RNN),\u003csup\u003e[4,5]\u003c/sup\u003e has emerged as particularly advantageous for real-time, on-device AI applications. Unlike conventional deep learning architectures that typically require extensive training resources, RC leverages a dynamic reservoir to intrinsically encode complex temporal signals into linearly separable, high-dimensional states. This intrinsic capability significantly reduces computational overhead, making RC highly suitable for efficient, adaptive processing of nonlinear, temporal, and chaotic signals in skin electronics. To effectively realize RC-based neuromorphic computing, memtransistors, which simultaneously integrate memory and transistor functionalities, have emerged as ideal hardware components.\u003csup\u003e[6-8]\u003c/sup\u003e\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eMemtransistors inherently exhibit short-term plasticity (STP), a crucial synaptic property allowing transient storage and dynamic modulation of conductance states in response to temporal signals, mimicking biological synaptic behavior.\u003csup\u003e[9-12]\u003c/sup\u003e Depending on design and conditions, they can also support other forms of plasticity, offering flexibility for neuromorphic computing. In addition, the memtransistors also provide enhanced controllability in active-matrix array configurations due to the presence of additional terminals, making them particularly well-suited for scalable, real-time, and adaptive AI processing in on-skin electronic systems.\u003csup\u003e[13,14]\u003c/sup\u003e Despite these clear advantages, current memtransistor implementations still predominantly suffer from limited on/off current ratios, high power consumption, and substantial variability, hindering their ability to achieve highly accurate and energy-efficient reservoir computing (RC). Furthermore, these devices are typically composed of rigid and brittle inorganic materials, severely limiting their compatibility with skin-integrated systems.\u003c/p\u003e\n\u003cp\u003eTo realize practical on-skin AI electronics, it is essential to develop memtransistors that synergistically combine high performance neuromorphic computing with skin-like stretchability and autonomous self-healing ability.\u003csup\u003e[15]\u003c/sup\u003e Recently, supramolecular polymer materials have emerged as promising candidates for skin-integrated electronics, leveraging intrinsic self-healing and exceptional stretchability enabled by dynamic, reversible supramolecular interactions.\u003csup\u003e[16-18]\u003c/sup\u003e Nevertheless, despite significant advances in supramolecular polymer-based electronics, previous studies have exclusively demonstrated conventional transistor configurations without integrated neuromorphic functions. Thus, the development of memtransistors explicitly engineered for neuromorphic computing remains unexplored, yet constitutes a critical step toward adaptive, mechanically robust, damage-resistant, and intelligent on-skin electronic systems.\u003c/p\u003e\n\u003cp\u003eIn this study, we introduce a supramolecular polymer memtransistor specifically engineered to enable neuromorphic AI computing directly on the skin. Leveraging intrinsic short-term plasticity (STP) and dynamic supramolecular interactions, our memtransistor simultaneously achieves highly energy efficient neuromorphic processing (0.29 fJ - 1.8 nJ) with high on/off ratio exceeding 10\u003csup\u003e3\u003c/sup\u003e on/off ratio and narrow variability (\u0026sigma;/\u0026mu; = 5.25%). By integrating our device within a reservoir computing framework, we demonstrate its exceptional capability to efficiently process nonlinear and chaotic spatiotemporal signals. Moreover, this device consistently maintains stable electrical performance even under a 30% biaxial strain after experiencing mechanical damage and subsequent autonomous self-healing. We further demonstrate a stable 7\u0026times;7 active-matrix memtransistor array, confirming the feasibility of scalable integration for neuromorphic systems. To the best of knowledge, we present the first demonstration of chaotic signal prediction applied to human sensory information processing using a supramolecular polymer-based neuromorphic memtransistor, highlighting its potential in advanced tasks related with time series information processing such as spoken digit classification (Audio MNIST),\u003csup\u003e[19]\u003c/sup\u003e emotion recognition (RAVDESS),\u003csup\u003e[20]\u003c/sup\u003e and chaotic time-series forecasting. \u0026nbsp;This integrative approach represents an emerging demonstration of a supramolecular polymer memtransistor tailored explicitly for neuromorphic computing in bio-integrated systems, highlighting a new paradigm that combines mechanical resilience, low-power consumption, and adaptive AI processing capabilities.\u003c/p\u003e\u003cp\u003e\u003cstrong\u003eDesign and Characterization of Supramolecular Polymer Memtransistor\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe schematic of a supramolecular polymer memtransistor active-matrix array (7 \u0026times; 7) designed for seamless integration with human skin, alongside the detailed structure of an individual device, is illustrated in\u0026nbsp;\u003cstrong\u003eFig. 1a\u003c/strong\u003e. Each memtransistor consists entirely of supramolecular polymer composites that encompass semiconductor, dielectric, and electrode layers. These devices function via a charge-trapping mechanism, facilitated by precise energy-level alignment engineered at the p-n heterojunction (\u003cstrong\u003eFig. 1b\u003c/strong\u003e,\u0026nbsp;\u003cstrong\u003eSupplementary\u003c/strong\u003e \u003cstrong\u003eFigs. 1-3\u003c/strong\u003e).\u003csup\u003e[21-23]\u003c/sup\u003e We employed poly(2,5-bis(2-octyldodecyl)-3,6-di(thiophen-2-yl) diketopyrrolo [3,4-c] pyrrole-1,4-dione-alt-thieno [3,2-b]thiophen) (DPPT-TT) as the p-type semiconductor, and poly[[1,2,3,6,7,8-hexahydro-2,7-bis(2-octyldodecyl)-1,3,6,8-dioxobenzo[lMn][3,8]phenanthroline-4,9-diyl][2,2\u0026apos;-bithiophene]-5,5\u0026apos;-diyl] (N2200) as the n-type semiconductor for p-n junction based semiconducting layer. A supramolecular polymeric elastomer (SPE) composed of poly(dimethylsiloxane)-4,4\u0026prime;-methylenebis(phenyl urea)-isophorone bisurea (PDMS\u0026ndash;MPU\u003csub\u003e6\u003c/sub\u003e\u0026ndash;IU\u003csub\u003e4\u003c/sub\u003e) served as a stretchable and self-healing elastomer matrix for the memtransistor components, owing to its excellent elasticity, autonomous self-healing capability, high compatibility with all components (semiconductor, electrode, and dielectric) (\u003cstrong\u003eSupplementary Figs. 4-8\u003c/strong\u003e).\u003csup\u003e[18]\u003c/sup\u003e The supramolecular p-n junction semiconductor film was prepared by transfer-printing of the p-type and n-type semiconductor layers, each individually blended at the molecular level with the SPE matrix at optimized mass ratios (DPPT-TT:SPE = 3:7, N2200:SPE = 5:5). These mass ratios were systematically balanced to achieve desirable electrical characteristics, including effective threshold voltage (V\u003csub\u003eth\u003c/sub\u003e) shifts indicative of trap capacity, high on-current (\u003cstrong\u003eFig. 1c\u003c/strong\u003e), and robust mechanical resilience characterized by significant stretchability and autonomous self-healing (\u003cstrong\u003eFig. 1d\u003c/strong\u003e,\u0026nbsp;\u003cstrong\u003eSupplementary Figs. 9,10\u003c/strong\u003e). Both supramolecular semiconductor blends exhibited distinctive nanoweb-like morphologies with clear phase separation, enabling efficient dissipation of mechanical stress through reversible supramolecular interactions and sustaining conductive pathways under biaxial strain and after autonomous healing (\u003cstrong\u003eFig. 1e,f\u003c/strong\u003e,\u0026nbsp;\u003cstrong\u003eSupplementary Fig. 11-14\u003c/strong\u003e, and Supplementary Note 5). The memtransistor array demonstrated highly consistent and uniform transfer characteristics across all 49 devices, with an average on-current of approximately 5\u0026nbsp;\u0026times;\u0026nbsp;10\u003csup\u003e-7\u003c/sup\u003e A, low leakage currents (~10\u003csup\u003e-10\u003c/sup\u003e A), high on/off current ratios (~10⁴) without kink effect (\u003cstrong\u003eSupplementary Figs. 15,16\u003c/strong\u003e, and Supplementary Note 1). The devices exhibited anticlockwise hysteresis with wide V\u003csub\u003eth\u003c/sub\u003e shifts (from 2.3 to 37.5 V) during gate voltage (V\u003csub\u003eG\u003c/sub\u003e) sweeps (\u003cstrong\u003eFig. 1g\u003c/strong\u003e), maintaining reliability at low drain voltages (V\u003csub\u003eD\u003c/sub\u003e = -1 V) and a stable V\u003csub\u003eth\u003c/sub\u003e shift (~37 V) (\u003cstrong\u003eSupplementary Figs. 17,18\u003c/strong\u003e).\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eTo establish the origin of reservoir characteristics, control devices lacking the n-type semiconductor layer were fabricated and evaluated. Unlike p-n junction-based devices, these control devices exhibited no change in hysteresis and trap capacity characteristics, confirming that reservoir functionality specifically originates from energy-level alignment within the p-n junction interface (\u003cstrong\u003eFig. 1b\u003c/strong\u003e and \u003cstrong\u003eSupplementary Figs. 19-21\u003c/strong\u003e). Electron-trapping activation energy (E\u003csub\u003eA\u003c/sub\u003e), quantified to be approximately 27.93 meV, confirmed electron trapping at the heterojunction as the primary mechanism for observed synaptic behavior (\u003cstrong\u003eFig. 1h, bottom\u003c/strong\u003e and \u003cstrong\u003eSupplementary Figs. 22,23\u003c/strong\u003e and Supplementary Note 2). Capacitance-voltage (C-V) analysis of metal-insulator-semiconductor (MIS) capacitors further validated enhanced charge trapping at the semiconductor-dielectric interface, corroborating charge trapping at the p-n junction (\u003cstrong\u003eSupplementary Fig. 24\u003c/strong\u003e). Additionally, the effective Schottky barrier height (\u003cem\u003eϕ\u003csub\u003eS\u003c/sub\u003e\u003c/em\u003e), derived from temperature-dependent output current measurements, elucidated charge injection and transport dynamics modulated by V\u003csub\u003eG\u003c/sub\u003e (Supplementary Note 3, 4). Increasing V\u003csub\u003eG\u003c/sub\u003e effectively reduced \u003cem\u003eϕ\u003csub\u003eS\u003c/sub\u003e\u003c/em\u003e, enhancing carrier injection into the active channel (\u003cstrong\u003eFig. 1h, top\u003c/strong\u003e). Application of V\u003csub\u003eG\u003c/sub\u003e induced hole accumulation at the semiconductor-insulator interface, resulting in energy band bending and reducing the source-semiconductor barrier, subsequently elevating output current. Electron trapping by the n-type layer further enhanced hole concentration, contributing to reservoir retention. After gate-pulse termination, trapped electrons recombined with channel holes, gradually diminishing output current and realizing short-term plasticity (STP) functionality. Different gate biases injected varied charge concentrations into the active layer, enriching carrier dynamics essential for spatiotemporal mapping in reservoir computing.\u003c/p\u003e\n\u003cp\u003eThe intrinsic stretchability and self-healing capabilities of the supramolecular polymer memtransistors were systematically characterized through morphological and electrical analyses. Each component of the devices (semiconductor, electrode, and dielectric layers) exhibited intrinsic stretchability, maintaining structural integrity and electrical performance under mechanical strains of up to 30% (uniaxial and biaxial) as well as autonomous self-healing capabilities (\u003cstrong\u003eFig. 1i,j\u003c/strong\u003e). Leveraging these insights, we assessed the overall mechanical resilience of the complete memtransistor devices. The devices consistently maintained their initial transfer characteristics, such as on-current (~10\u003csup\u003e-7\u003c/sup\u003e A) and \u0026Delta;V\u003csub\u003eth\u003c/sub\u003e shift (~36.5 V), even under 30% biaxial strains, demonstrating compatibility with the mechanical compliance requirements for human skin (\u003cstrong\u003eFigs. 1k,l\u003c/strong\u003e and \u003cstrong\u003eSupplementary Figs. 25,26\u003c/strong\u003e, \u003cstrong\u003eTable S1\u003c/strong\u003e). The autonomous self-healing capability of the memtransistors was evaluated by deliberately cutting the device channel using a surgical blade, creating microscale damage (a defined 3 \u0026micro;m-wide gap). The damaged device regained functionality after approximately 12 hours, and its transfer characteristics fully recovered to the original pristine state within three days of autonomous healing (\u003cstrong\u003eFigs. 1m,n\u003c/strong\u003e and \u003cstrong\u003eSupplementary Fig. 27\u003c/strong\u003e).\u003cstrong\u003e\u0026nbsp;\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eNeuromorphic Synaptic Behaviors of Supramolecular Polymer Memtransistors\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe synaptic characteristics of the supramolecular polymer memtransistors were systematically evaluated to demonstrate their suitability for reservoir computing (RC)-based neuromorphic systems (\u003cstrong\u003eFig. 2\u003c/strong\u003e and \u003cstrong\u003eSupplementary\u0026nbsp;Fig. 28\u003c/strong\u003e). The excitatory postsynaptic current (EPSC) responses induced by single electrical pulses were measured by applying presynaptic voltage pulses (+40 V) with pulse widths ranging from 0.75 to 9 s to the gate electrode, while maintaining a constant drain voltage (V\u003csub\u003eD\u003c/sub\u003e = -1 V). As shown in \u003cstrong\u003eFig. 2a\u003c/strong\u003e, the EPSC amplitude increased progressively with pulse width due to enhanced electron trapping at the p-n junction interface. The dependence of the EPSC on/off ratio and decay time (\u0026tau;) on pulse width was investigated under various device states, including pristine condition, 30% uniaxial and biaxial strains, and after autonomous self-healing (\u003cstrong\u003eFig. 2b\u003c/strong\u003e and \u003cstrong\u003eSupplementary\u0026nbsp;Figs. 29-31\u003c/strong\u003e). Moreover, the EPSC amplitude increased proportionally with the presynaptic pulse amplitude, varied from 10 V to 90 V, and remained stable with narrow variation even under mechanical deformation (up to 30% strain) and following autonomous self-healing (\u003cstrong\u003eFig. 2c,d\u003c/strong\u003e and \u003cstrong\u003eSupplementary\u0026nbsp;Fig. 32-34\u003c/strong\u003e).\u003c/p\u003e\n\u003cp\u003eShort-term synaptic plasticity was quantified using the paired-pulse facilitation (PPF) index, defined as the amplitude ratio (A\u003csub\u003e2\u003c/sub\u003e/A\u003csub\u003e1\u003c/sub\u003e) of two successive EPSC peaks. Two consecutive presynaptic pulses (+40 V) with pulse intervals (\u0026Delta;t\u003csub\u003epre\u003c/sub\u003e) ranging from 0.375 to 72 s were applied at a fixed drain voltage (V\u003csub\u003eD\u003c/sub\u003e = -1 V) (\u003cstrong\u003eFig. 2e\u003c/strong\u003e and \u003cstrong\u003eSupplementary\u0026nbsp;Fig. 35-38\u003c/strong\u003e). The PPF index gradually decreased from 335% to 135% as \u0026Delta;t\u003csub\u003epre\u003c/sub\u003e increased from 0.375 to 72 s. Both the magnitude and overall trend of the PPF index remained consistent even under substantial mechanical deformation (30% uniaxial and biaxial strains) and after autonomous self-healing (\u003cstrong\u003eFig. 2f\u003c/strong\u003e). The EPSC responses were further characterized under trains of presynaptic pulses (1 to 50 successive pulses, +40 V, 3 s each), demonstrating consistent increases in EPSC amplitude and on/off ratio up to 3000, with minimal variation under pristine, strained (30% uniaxial and biaxial), and autonomously healed conditions (\u003cstrong\u003eFig. 2g,h\u003c/strong\u003e and \u003cstrong\u003eSupplementary\u0026nbsp;Fig. 39-47\u003c/strong\u003e).\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eTo confirm synaptic stability and robustness, repeated endurance cycling was conducted through long-term potentiation (LTP) and depression (LTD) measurements. \u003cstrong\u003eFig. 2i\u003c/strong\u003e and \u003cstrong\u003eSupplementary\u0026nbsp;Fig. 48\u003c/strong\u003e presents stable and repeatable transitions between LTP and LTD states during continuous alternating presynaptic spike cycles (250 cycles total), each comprising 20 potentiation spikes (+40 V, 3 s) followed by 20 depression spikes (-0.5 V, 3 s). The current transitions remained highly stable, exhibiting negligible variation over a total of 10,000 pulses, indicating that the device possesses reliable short-term memory characteristics under various interval times at -1 V\u003csub\u003eD\u003c/sub\u003e (\u003cstrong\u003eFig. 2j\u003c/strong\u003e). These results indicate that this memristor exhibits robust electrical durability a high on/off ratio, and reliable controllability in synaptic weight modulation, combined with intrinsic stretchability and self-healing ability not previously demonstrated. Furthermore, the energy consumption of these artificial synapses spans a broad range (from 0.29 fJ to 1.8 nJ), with the lowest achievable values approaching those of biological synapses in the human brain (1\u0026ndash;100 fJ), significantly surpassing previously reported transistor-based artificial synaptic devices (\u003cstrong\u003eFig. 2k\u003c/strong\u003e, \u003cstrong\u003eSupplementary\u0026nbsp;Fig. 49\u003c/strong\u003e, and \u003cstrong\u003eTable S2\u003c/strong\u003e).\u003csup\u003e[24-31]\u003c/sup\u003e\u003cstrong\u003e\u0026nbsp;\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eReservoir Computing of Supramolecular Polymer Memtransistors\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eTo demonstrate the applicability of our supramolecular polymer memtransistors for efficient temporal neuromorphic processing, we implemented a physical reservoir computing (PRC) framework that leverages their inherent STP. This hardware-based approach is highly suitable for spatiotemporal signal processing in wearable electronics platforms, enabling on-skin classification tasks that rely on sequential pattern recognition, such as speech and emotion recognition. We fabricated a 7 \u0026times; 7 active-matrix array of memtransistors to construct the hardware reservoir (\u003cstrong\u003eFig. 3a\u003c/strong\u003e). Each memtransistor (M₁, M₂, ..., Mₙ) functions as a dynamic reservoir node, receiving spike train inputs and generating time-varying current outputs that collectively form the reservoir state. \u0026nbsp;\u003cstrong\u003eFigure 3b\u003c/strong\u003e presents the schematic overview of the PRC architecture. In this system, a time-dependent input signal \u003cem\u003eu(t)\u003c/em\u003e is projected into a high-dimensional dynamic state space \u003cem\u003ex(t)\u003c/em\u003e via fixed input weights \u003cem\u003eW\u003csub\u003ein\u003c/sub\u003e\u003c/em\u003e. Prior to entering the reservoir layer, the sequential input signals are processed through a masking step, enabling richer spatiotemporal representations within the reservoir. It has been widely adopted in PRC to enhance the separability of input patterns and improve classification accuracy.\u003csup\u003e[32]\u003c/sup\u003e The nonlinear reservoir, realized through a network of memtransistors, transforms these signals into high-dimensional dynamics. This transformation leverages internal fading memory effects, which are critical for temporal signal processing. Only the output weights \u003cem\u003eW\u003csub\u003eout\u003c/sub\u003e\u003c/em\u003e are trained using supervised learning to generate the final output \u003cem\u003ey(t)\u003c/em\u003e that matches the desired target \u003cem\u003ey\u003csub\u003etarget\u003c/sub\u003e(t)\u003c/em\u003e, simplifying training complexity and computational demands.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eTo validate the dynamic encoding ability of individual memtransistors, we applied temporally modulated binary input sequences (\u0026quot;0000\u0026quot; to \u0026quot;1111\u0026quot;) using 40 V for \u0026quot;1\u0026quot; and 1 V for \u0026quot;0\u0026quot;, as shown in \u003cstrong\u003eFig. 3c\u003c/strong\u003e. The devices exhibited characteristic current increases during \u0026quot;1\u0026quot; pulses and gradual decay during \u0026quot;0\u0026quot; intervals, clearly demonstrating fading memory and STP essential for reservoir computing. These dynamic responses remained stable under mechanical strain (30% uniaxial and biaxial) and even after mechanical damage followed by autonomous self-healing (\u003cstrong\u003eSupplementary\u0026nbsp;Fig. 50\u003c/strong\u003e), confirming both the functional stability and mechanical resilience. Additionally, these characteristics were well preserved across the array, with the data exhibiting excellent uniformity across 49 devices, with low device-to-device variation (\u0026sigma;/\u0026mu; = 5.25%), ensuring scalable and reliable neuromorphic performance (\u003cstrong\u003eFig. 3d\u003c/strong\u003e and \u003cstrong\u003eSupplementary\u0026nbsp;Fig. 51\u003c/strong\u003e).\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eTo benchmark the reservoir\u0026rsquo;s capability for sequential classification tasks involving speech, we employed two representative datasets: AudioMNIST for digit recognition and RAVDESS for emotion recognition. \u003cstrong\u003eFigure 3e\u003c/strong\u003e illustrates preprocessing of raw speech into time-frequency representations, utilizing cochleagram transformations for digit recognition tasks (emphasizing temporal precision)\u003csup\u003e[33]\u003c/sup\u003e and mel-spectrogram transformations for emotion recognition tasks (capturing perceptual acoustic features).\u003csup\u003e[34]\u003c/sup\u003e These spectrograms were converted into binary spike trains, effectively encoding phonetic and emotional nuances for reservoir-based classification (Supplementary Note 6, 7). \u003cstrong\u003eFigure 3f\u003c/strong\u003e shows that classification accuracy on the AudioMNIST dataset increased with mask length, reaching over 99% accuracy using all 49 devices. This trend is attributed to enhanced feature extraction and frequency resolution enabled by longer masks, allowing for more refined recognition of pitch modulation and syllabic structures (\u003cstrong\u003eSupplementary\u0026nbsp;Fig. 52\u003c/strong\u003e). The classification accuracy remained stable (~99%) even under severe mechanical deformations (30% uniaxial and biaxial) and even post self-healing, confirming robust functionality under mechanical stress (\u003cstrong\u003eFig. 3g\u003c/strong\u003e and \u003cstrong\u003eSupplementary\u0026nbsp;Fig. 53\u003c/strong\u003e).\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eEvaluating more complex tasks, emotion recognition achieved high accuracy both under mechanical strain (73.6%) and after healing (70.83%), exhibiting stable misclassification patterns (\u003cstrong\u003eFig. 3h,i\u003c/strong\u003e and \u003cstrong\u003eSupplementary\u0026nbsp;Fig. 54\u003c/strong\u003e). Such performance compares favorably with previously reported studies (typically ~65\u0026ndash;75%),\u003csup\u003e[35,36]\u003c/sup\u003e affirming the mechanical and functional integrity of our PRC system. We also tested the system on the MNIST dataset to validate the spatial computation capability, achieving \u0026gt;97% classification accuracy (\u003cstrong\u003eSupplementary\u0026nbsp;Fig. 55\u003c/strong\u003e), further confirming the versatility of our memtransistor-based reservoir architecture.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eRecursive Chaotic Time-Series Prediction for On-Skin Artificial Intelligence\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eWhile classification tasks such as speech and emotion recognition effectively demonstrate our system\u0026apos;s capability in temporal pattern recognition, these problems remain relatively structured and limited in complexity. In contrast, biological signals frequently encountered in wearable applications including neural, cardiac, and respiratory activities often exhibit strong nonlinearity, high noise levels, and even chaotic behavior under certain physiological states, thereby posing a significantly greater challenge for predictive modeling.\u003csup\u003e[37]\u003c/sup\u003e To evaluate the applicability of our system to such complex and unstructured temporal data, we adopted a canonical benchmark for chaotic dynamics: the Lorenz attractor. Originally developed to model atmospheric convection, the Lorenz system exhibits strong nonlinearity and extreme sensitivity to initial conditions, making it an ideal testbed for assessing the temporal memory and dynamic processing capabilities of PRC systems (\u003cstrong\u003eSupplementary\u0026nbsp;Fig. 56\u003c/strong\u003e).\u003c/p\u003e\n\u003cp\u003ePrediction tasks employing the Lorenz system typically follow two approaches: one-step-ahead and multi-step-ahead forecasting. The one-step-ahead method yields accurate single-step predictions by relying on ground truth data at each inference step yet inherently limits autonomous forecasting beyond a single time step (\u003cstrong\u003eSupplementary\u0026nbsp;Fig. 57\u003c/strong\u003e). To overcome this limitation, we implemented recursive multi-step-ahead prediction coupled with online training, aiming to suppress errors accumulation and enhance long-term prediction stability. \u003cstrong\u003eFigure 4a\u003c/strong\u003e schematically illustrates the recursive forecasting paradigm employed in our system. At each cycle, the reservoir predicts the system state at time t+1 using historical data from steps 1 to t. This prediction is then recursively fed back into the reservoir to forecast future states autonomously, without access to ground truth inputs. This feedback-based architecture facilitates extended, self-sustained prediction over long time horizons (Supplementary Note 8). Utilizing the pristine memtransistor array as the reservoir, we trained the system on Lorenz trajectories up to time step 32,000. Beyond this point, the system executed autonomous recursive forecasting cycles comprising 100 steps of prediction followed by 50 steps of online retraining. Due to the inherently chaotic nature of the Lorenz system, even minor deviations in prediction rapidly amplify beyond acceptable limits (\u003cstrong\u003eSupplementary\u0026nbsp;Fig. 58\u003c/strong\u003e). Nevertheless, our reservoir reliably preserved the attractor\u0026rsquo;s characteristic butterfly-shaped orbit throughout the recursive process, indicating that prediction errors remained within a controllable range (\u003cstrong\u003eFig. 4b\u003c/strong\u003e). A magnified view of the z-axis prediction emphasizes the close alignment between predicted and true values, particularly within complex oscillatory regions. \u003cstrong\u003eFigure 4c\u003c/strong\u003e presents the reconstructed 3D trajectory, demonstrating that predicted values (purple dotted line) accurately replicate the characteristic butterfly-shaped orbit of the Lorenz attractor compared with the ground truth (gray solid line). This highlights the reservoir\u0026rsquo;s proficiency in modeling nonlinear chaotic dynamics in phase space. This further substantiates that online learning effectively mitigates prediction drift, maintaining high fidelity during extended forecasts (\u003cstrong\u003eSupplementary\u0026nbsp;Fig. 59\u003c/strong\u003e). To assess mechanical robustness, we repeated the Lorenz prediction experiments under conditions of uniaxial strain, biaxial strain, and after autonomous self-healing (\u003cstrong\u003eSupplementary\u0026nbsp;Fig. 60\u003c/strong\u003e). \u003cstrong\u003eFigure 4d\u003c/strong\u003e summarizes the normalized root mean square error (NRMSE) across these varying mechanical conditions, consistently demonstrating low error values (~0.02 or below). This quantitative evaluation further validates the mechanical resilience and long-term predictive stability of our stretchable PRC platform for on-skin chaotic time-series forecasting.\u003c/p\u003e\n\u003cp\u003eTo further demonstrate the robustness of our reservoir across diverse dynamical environments, we extended trajectory forecasting experiments to other chaotic attractors. These tasks were performed using the same online training structure as for the Lorenz attractor. As shown in \u003cstrong\u003eFig. 4e,f\u003c/strong\u003e, the predicted trajectories closely match the ground truth for both the Dadras-Momeni attractor and the Halvorsen attractor, thereby confirming the universality of our device as a physical reservoir (\u003cstrong\u003eSupplementary\u0026nbsp;Figs. 61,62\u003c/strong\u003e). In this case, NRMSE values of approximately 0.01374 and 0.03814 were obtained, confirming that accurate predictions were achieved in all cases, which demonstrates the ability to sustain accurate modeling of highly nonlinear dynamics over extended horizons, highlighting a materials-based pathway toward bio-integrated AI systems capable of both precise pattern recognition and long-term adaptive prediction.\u003csup\u003e[38, 39]\u003c/sup\u003e\u003c/p\u003e"},{"header":"Methods","content":"\u003cp\u003e\u003cstrong\u003eMaterials:\u0026nbsp;\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eBis(3-aminopropyl) terminated poly(dimethylsiloxane) (H\u003csub\u003e2\u003c/sub\u003eN-PDMS-NH\u003csub\u003e2\u003c/sub\u003e, M\u003csub\u003en\u003c/sub\u003e\u0026thinsp;=\u0026thinsp;5000) was purchased from Gelest Inc.. Poly(2,5-bis(2-octyldodecyl)\u0026minus;3,6-di(thiophen-2-yl)diketopyrrolo[3,4-c]pyrrole-1,4-dione-alt-thieno[3,2-b]thiophen) (DPPT-TT) (M\u003csub\u003ew\u003c/sub\u003e/M\u003csub\u003en\u003c/sub\u003e\u0026thinsp;=\u0026thinsp;100k/42k, PDI\u0026thinsp;\u0026lt;\u0026thinsp;3.0) and poly[[1,2,3,6,7,8-hexahydro-2,7-bis(2-octyldodecyl)-1,3,6,8-dioxobenzo[lMn][3,8]phenanthroline-4,9-diyl][2,2\u0026apos;-bithiophene]-5,5\u0026apos;-diyl] (N2200) (M\u003csub\u003ew\u003c/sub\u003e \u0026gt; 100k, PDI \u0026lt; 3.0) semiconducting polymers were purchased from Derthon. Poly(dimethylsiloxane) (PDMS, Sylgard 184) and its cross-linker were purchased by Dow Corning. The PDMS was cured with a ratio of 8:1 (base/cross-linker, w/w) at 65 \u0026deg;C overnight for the transfer printing stamp. SEBS H1062 (S/EB weight ratio of 18/82) was supported by Asahi Kasei company. Ag (silver, 99.99%, 3\u0026ndash;5\u0026thinsp;mm granule) was purchased from SY SCIENCE. The Trichloro(octadecyl)silane (OTS) solution and diiodomethane were purchased by Sigma Aldrich. The Ethyl-amine (Et\u003csub\u003e3\u003c/sub\u003eN), anhydrous chloroform, and toluene were purchased by Sigma Aldrich. Methanol (MeOH) solution was purchased by SAMCHUN. All chemicals and materials were used without any purification.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eSynthesis of supramolecular polymeric elastomer (SPE) PDMS-MPU\u003csub\u003e6\u003c/sub\u003e-IU\u003csub\u003e4\u003c/sub\u003e\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003ePDMS-MPU\u003csub\u003e6\u003c/sub\u003e-IU\u003csub\u003e4\u003c/sub\u003e polymer was synthesized according to previously reported methods.\u003csup\u003e[18]\u003c/sup\u003e H\u003csub\u003e2\u003c/sub\u003eN-PDMS-NH\u003csub\u003e2\u003c/sub\u003e (30\u0026thinsp;g, M\u003csub\u003en\u003c/sub\u003e\u0026thinsp;=\u0026thinsp;5000, 1 eq) was dissolved in 120\u0026thinsp;mL Chloroform at 0\u0026thinsp;\u0026deg;C under nitrogen atmosphere. Et\u003csub\u003e3\u003c/sub\u003eN (3\u0026thinsp;mL) was added to the solution of H\u003csub\u003e2\u003c/sub\u003eN-PDMS-NH\u003csub\u003e2\u003c/sub\u003e and stirred for 1\u0026thinsp;h. After that, a mixture solution of 4,4\u0026rsquo;\u0026ndash; Methylenebis(phenyl isocyanate) (MPU) (0.9\u0026thinsp;g, 6 eq) and Isophorone diisocyanate (IU) (0.54\u0026thinsp;g, 4 eq) in CHCl\u003csub\u003e3\u003c/sub\u003e was added dropwise. The solution was then allowed to warm to room temperature and stirred for 4 days.\u003c/p\u003e\n\u003cp\u003eAfter reaction, MeOH (5\u0026thinsp;mL) was added and stirred for 30\u0026thinsp;min to remove the remaining isocyanate. Then, the solution was concentrated to \u0026frac12; of its volume. 18\u0026thinsp;mL MeOH was poured into the mixture solution to give a viscous liquid. After settling for 30\u0026thinsp;min, the upper clear solution was then decanted. 30\u0026thinsp;mL CHCl\u003csub\u003e3\u003c/sub\u003e was added to dissolve the product. The dissolution-precipitation-decantation process was repeated three times for purifying and the final product was subjected in ambient condition to remove the solvent and trace of Et\u003csub\u003e3\u003c/sub\u003eN.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eDevice fabrications and characterizations\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eFully stretchable and self-healing memtransistor preparation:\u0026nbsp;\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe solution for the p-type semiconductor was prepared by dissolving DPPT-TT (0.21 wt%) and SPE (0.49 wt%) with a total of 0.7 wt%, in anhydrous chloroform. The solution for the n-type semiconductor was prepared by dissolving N2200 (0.0075 wt%) and SPE (0.0075 wt%) with a total of 0.015 wt%, in anhydrous chloroform. Both solutions were stirred at 50 \u0026deg;C for 4 hours. The solutions were spun on an OTS-treated SiO\u003csub\u003e2\u003c/sub\u003e/Si wafer at 1000 rpm for p-type semiconductor (film thickness: 100 nm) and 1500 rpm for n-type semiconductor in 1 min after filtration with a PTFE-D (0.2 \u0026micro;m) filter (film thickness: 20 nm). Both p-type and n-type semiconducting films were then annealed at 80 \u003csup\u003eo\u003c/sup\u003eC for 30 min. All above processes were carried out under an N\u003csub\u003e2\u003c/sub\u003e atmosphere in a glove box with extremely low levels of moisture (H\u003csub\u003e2\u003c/sub\u003eO \u0026lt; 0.01 parts per million (ppm)) and oxygen (O\u003csub\u003e2\u003c/sub\u003e \u0026lt; 0.01 ppm).\u003c/p\u003e\n\u003cp\u003eThe SPE substrate solution was spin-coated onto an OTS-treated SiO\u003csub\u003e2\u003c/sub\u003e/Si wafer using the SPE solution (50 mg/mL in chloroform) at 2000 rpm for 1 minute without any heat treatment. The resulting SPE substrate was transferred onto the SEBS substrate to be stretchable and elastic. 100 nm thick Ag gate electrode was thermally evaporated onto the SPE substrate at a speed of 0.2 nm/s under high vacuum conditions (below 5.0\u0026times;10\u003csup\u003e-6\u003c/sup\u003e torr). Then, the SPE dielectric (60 mg/ml, 1000 rpm in 1 min on OTS-treated SiO\u003csub\u003e2\u003c/sub\u003e, 1.5 \u0026micro;m thick), n-type semiconducting film (in a 5:5 ratio of N2200 and SPE), and p-type semiconducting film (in a 3:7 ratio of DPPT-TT and SPE) were sequentially transferred onto the gate electrode. Finally, 100 nm thick Ag source/drain electrodes were thermally evaporated onto the top p-type semiconducting film. \u0026nbsp;\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eCharacterization\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe electrical, hysteresis, and synaptic characteristics of the devices were measured under ambient conditions by a probe station, which is connected to KEITHLEY 4200. The strain-stress curve was obtained by a force tester (AND, MCT-2150; strain rate: 100 mm/min). UV-Vis-Nir spectra were measured using a spectrophotometer (JASCO, V-770). For the cutting process, a surgical blade (Feather, No. 25) was used. Surface structures and current images were obtained with atomic force microscopy (AFM; Bruker MultiMode 8-HR) under ambient conditions. DMT modulus mappings were measured using PeakForce quantitative nanoscale mechanical (QNM) AFM. Optical images were obtained with an optical microscope (OM; Leica DM4 M). The thicknesses of the film were obtained with an ellipsometer (WS WONWOO S-TRC2000). X-ray photoelectron spectra depth profiling was obtained with XPS equipment (Thermo Electron, K-Alpha). Surface energy and contact angle were obtained by PHOENIX-MT(T). Work function of p-type and n-type semiconductor was measured using Riken Keiki (Japan)/AC-2 with UV intensity 50 nW and 200 nW.\u003c/p\u003e\n"},{"header":"Declarations","content":"\u003cp\u003e\u003cstrong\u003eAcknowledgments\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThis research was supported by the Basic Science Research Program through the National Research Foundation of Korea (NRF) funded by the Ministry of Science and ICT (MSIT) (RS-2021-NR061555, RS-2020-NR049601), GRRC program of Gyeonggi province (GRRCKYUNGHEE2023-B03) and the Korea Institute for the Advancement of Technology (KIAT) and the Ministry of Trade, Industry \u0026amp; Energy (MOTIE) of the Republic of Korea (RS-2024-00434908, RS-2021-KI002513, RS-2025-25435993, RS-2024-00406182, and RS-2025-02305609).\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAuthor contributions\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eN.T.P.V., E.J.Y., S.H.L. and J.Y.O. conceived the study and N.T.P.V., E.J.Y., S.H.L. and J.Y.O. designed the experiments. N.T.P.V., E.J.Y. conducted all experiments. N.T.P.V., E.J.Y., K.H.J., M.W.J., S.H.P., T.A.N., H.R.C., S.H.L., J.Y.O. analyzed and discussed the data. N.T.P.V., E.J.Y., S.H.L. and J.Y.O. wrote the manuscript.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eCompeting interests\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eAuthors declare that they have no competing interests.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eData availability\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eData are available on request. Correspondence and requests for materials should be addressed to S.H.L and J.Y.O.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eSupplementary information\u003c/strong\u003e\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eThe online version contains supplementary material available at https://doi.org/##.####\u003c/p\u003e\n"},{"header":"References","content":"\u003col\u003e\n\u003cli\u003eZidan, M. A., Strachan, J. P. \u0026amp; Lu, W. D. The future of electronics based on memristive systems. \u003cem\u003eNat. Electron.\u003c/em\u003e\u003cstrong\u003e1\u003c/strong\u003e, 22\u0026ndash;29 (2018).\u003c/li\u003e\n\u003cli\u003eHorowitz, M. in 2014 \u003cem\u003eIEEE international solid-state circuits conference digest of technical papers\u003c/em\u003e (\u003cem\u003eISSCC\u003c/em\u003e). (\u003cem\u003eIEEE\u003c/em\u003e, 2014), pp. 10\u0026ndash;14.\u003c/li\u003e\n\u003cli\u003eLuko\u0026scaron;evičius, M. \u0026amp; Jaeger, H. Reservoir computing approaches to recurrent neural network training. \u003cem\u003eComput. Sci. Rev.\u003c/em\u003e\u003cstrong\u003e3\u003c/strong\u003e, 127\u0026ndash;149 (2009).\u003c/li\u003e\n\u003cli\u003eHopfield, J. J. Neural networks and physical systems with emergent collective computational abilities. \u003cem\u003eProceedings of the national academy of sciences\u003c/em\u003e\u003cstrong\u003e79\u003c/strong\u003e, 2554\u0026ndash;2558 (1982).\u003c/li\u003e\n\u003cli\u003eWerbos, P. J. Backpropagation through time: what it does and how to do it. \u003cem\u003eProceedings of the IEEE\u003c/em\u003e\u003cstrong\u003e78\u003c/strong\u003e, 1550\u0026ndash;1560 (2002).\u003c/li\u003e\n\u003cli\u003eLiang, X. et al. Physical reservoir computing with emerging electronics. \u003cem\u003eNat. Electron.\u003c/em\u003e\u003cstrong\u003e7\u003c/strong\u003e, 193\u0026ndash;206 (2024).\u003c/li\u003e\n\u003cli\u003eTanaka, G. et al. Recent advances in physical reservoir computing: A review. \u003cem\u003eNeural Networks\u003c/em\u003e\u003cstrong\u003e115\u003c/strong\u003e, 100\u0026ndash;123 (2019).\u003c/li\u003e\n\u003cli\u003eYan, M. et al. Emerging opportunities and challenges for the future of reservoir computing. \u003cem\u003eNat. Commun.\u003c/em\u003e\u003cstrong\u003e15\u003c/strong\u003e, 2056 (2024).\u003c/li\u003e\n\u003cli\u003eDu, C., Cai, F. et al. Reservoir computing using dynamic memristors for temporal information processing. \u003cem\u003eNat. Commun.\u003c/em\u003e\u003cstrong\u003e8\u003c/strong\u003e, 2204 (2017).\u003c/li\u003e\n\u003cli\u003eLiu, Z., Zhang, Q., Xie, D. et al. Interface-type tunable oxygen ion dynamics for physical reservoir computing. \u003cem\u003eNat. Commun.\u003c/em\u003e\u003cstrong\u003e14\u003c/strong\u003e, 7176 (2023).\u003c/li\u003e\n\u003cli\u003eMoon, J., Ma, W. et al. Temporal data classification and forecasting using a memristor-based reservoir computing system. \u003cem\u003eNat. Electron.\u003c/em\u003e\u003cstrong\u003e2\u003c/strong\u003e, 480\u0026ndash;487 (2019).\u003c/li\u003e\n\u003cli\u003eWu, X. et al. Ultralow-power optoelectronic synaptic transistors based on polyzwitterion dielectrics for in-sensor reservoir computing. \u003cem\u003eSci. Adv.\u003c/em\u003e\u003cstrong\u003e10\u003c/strong\u003e, eadn4524 (2024).\u003c/li\u003e\n\u003cli\u003eChoi, S. et al. 3D-integrated multilayered physical reservoir array for learning and forecasting time-series information. \u003cem\u003eNat. Commun.\u003c/em\u003e\u003cstrong\u003e15\u003c/strong\u003e, 2044 (2024).\u003c/li\u003e\n\u003cli\u003ePark, S. O., Jeong, H., Park, J., Bae, J. \u0026amp; Choi, S. Experimental demonstration of highly reliable dynamic memristor for artificial neuron and neuromorphic computing. \u003cem\u003eNat. Commun.\u003c/em\u003e\u003cstrong\u003e13\u003c/strong\u003e, 2888 (2022).\u003c/li\u003e\n\u003cli\u003ePatil, C. S. et al. Neuromorphic devices for electronic skin applications. \u003cem\u003eMater. Horiz.\u003c/em\u003e\u003cstrong\u003e12\u003c/strong\u003e, 2045-2088 (2025). \u003c/li\u003e\n\u003cli\u003eOh, J. Y., Lee, Y., Lee, T. W. Skin‐Mountable Functional Electronic Materials for Bio‐Integrated Devices. \u003cem\u003eAdv. Healthc. Mater.\u003c/em\u003e\u003cstrong\u003e13\u003c/strong\u003e, 2303797 (2024).\u003c/li\u003e\n\u003cli\u003eWang, S. \u0026amp; Urban, M. W. Self-healing polymers. \u003cem\u003eNat. Rev. Mater.\u003c/em\u003e\u003cstrong\u003e5\u003c/strong\u003e, 562-583, (2020).\u003c/li\u003e\n\u003cli\u003eVo, N. T. P., Nam, T. U. et al. Autonomous self-healing supramolecular polymer transistors for skin electronics. \u003cem\u003eNat Commun.\u003c/em\u003e\u003cstrong\u003e15\u003c/strong\u003e, 3433 (2024).\u003c/li\u003e\n\u003cli\u003eBecker, S., Vielhaben, J. et al. Audiomnist: Exploring explainable artificial intelligence for audio analysis on a simple benchmark. \u003cem\u003eJ. Frankl. Inst.\u003c/em\u003e\u003cstrong\u003e361\u003c/strong\u003e, 418\u0026ndash;428 (2024).\u003c/li\u003e\n\u003cli\u003eLivingstone, S. R. \u0026amp; Russo, F. A. The Ryerson Audio-Visual Database of Emotional Speech and Song (RAVDESS): A dynamic, multimodal set of facial and vocal expressions in North American English. \u003cem\u003ePLoS One\u003c/em\u003e\u003cstrong\u003e13\u003c/strong\u003e, e0196391 (2018).\u003c/li\u003e\n\u003cli\u003eKim, M. H., Jeong, M. W. et al. Mechanically robust stretchable semiconductor metallization for skin-inspired organic transistors. \u003cem\u003eSci. Adv.\u003c/em\u003e\u003cstrong\u003e8\u003c/strong\u003e, eade2988 (2022).\u003c/li\u003e\n\u003cli\u003eKim, J. S., Jeong, M. W. et al. Intrinsically Stretchable Subthreshold Organic Transistors for Highly Sensitive Low‐Power Skin‐Like Active‐Matrix Temperature Sensors. \u003cem\u003eAdv. Funct. Mat.\u003c/em\u003e\u003cstrong\u003e34\u003c/strong\u003e, 2305252 (2024).\u003c/li\u003e\n\u003cli\u003eJung, K. H., Hyun, J. et al. A biocompatible elastomeric organic transistor for implantable electronics. \u003cem\u003eNat. Electron.\u003c/em\u003e\u003cstrong\u003e8\u003c/strong\u003e, 1-13, (2025).\u003c/li\u003e\n\u003cli\u003eWu, X., Wang, S. et al. Wearable in-sensor reservoir computing using optoelectronic polymers with through-space charge-transport characteristics for multi-task learning. \u003cem\u003eNat. Comm.\u003c/em\u003e\u003cstrong\u003e14\u003c/strong\u003e, 468 (2023).\u003c/li\u003e\n\u003cli\u003eGao, C. et al. Toward grouped-reservoir computing: organic neuromorphic vertical transistor with distributed reservoir states for efficient recognition and prediction. \u003cem\u003eNat. Comm.\u003c/em\u003e\u003cstrong\u003e15\u003c/strong\u003e, 740 (2024).\u003c/li\u003e\n\u003cli\u003eZhong, Y. et al. Dynamic memristor-based reservoir computing for high-efficiency temporal signal processing. \u003cem\u003eNat. Comm.\u003c/em\u003e\u003cstrong\u003e12\u003c/strong\u003e, 408, (2021).\u003c/li\u003e\n\u003cli\u003eLiu, K. et al. An optoelectronic synapse based on \u0026alpha;-In\u003csub\u003e2\u003c/sub\u003eSe\u003csub\u003e3\u003c/sub\u003e with controllable temporal dynamics for multimode and multiscale reservoir computing. \u003cem\u003eNat. Electron.\u003c/em\u003e\u003cstrong\u003e5\u003c/strong\u003e, 761-773 (2022).\u003c/li\u003e\n\u003cli\u003eKim, J., Park, E. C., Shin, W. et al. Analog reservoir computing via ferroelectric mixed phase boundary transistors. \u003cem\u003eNat. Comm.\u003c/em\u003e\u003cstrong\u003e15\u003c/strong\u003e, 9147 (2024).\u003c/li\u003e\n\u003cli\u003eLiu, Z., Zhang, Q., Xie, D. et al. Interface-type tunable oxygen ion dynamics for physical reservoir computing. \u003cem\u003eNat. Commun.\u003c/em\u003e\u003cstrong\u003e14\u003c/strong\u003e, 7176 (2023).\u003c/li\u003e\n\u003cli\u003eAppeltant, L. et al. Information processing using a single dynamical node as complex system. \u003cem\u003eNat. Commun.\u003c/em\u003e\u003cstrong\u003e2\u003c/strong\u003e, 468 (2011).\u003c/li\u003e\n\u003cli\u003eWang, S., Chen, X. et al. An organic electrochemical transistor for multi-modal sensing, memory and processing. \u003cem\u003eNat. Electron.\u003c/em\u003e\u003cstrong\u003e6\u003c/strong\u003e, 281-291 (2023).\u003c/li\u003e\n\u003cli\u003eAppeltant, L. et al. Information processing using a single dynamical node as complex system. \u003cem\u003eNat. Commun.\u003c/em\u003e\u003cstrong\u003e2\u003c/strong\u003e, 468 (2011).\u003c/li\u003e\n\u003cli\u003eYoo, S., Chae, S. et al. Efficient data processing using tunable entropy-stabilized oxide memristors. \u003cem\u003eNat. Electron.\u003c/em\u003e\u003cstrong\u003e7\u003c/strong\u003e, 426\u0026ndash;436 (2024).\u003c/li\u003e\n\u003cli\u003eZhang, T., Feng, G., Liang, J. \u0026amp; An, T. Acoustic scene classification based on Mel spectrogram decomposition and model merging. \u003cem\u003eAppl. Acoust.\u003c/em\u003e\u003cstrong\u003e182\u003c/strong\u003e, 108258 (2021).\u003c/li\u003e\n\u003cli\u003eLuna-Jim\u0026eacute;nez, C. et al. Multimodal Emotion Recognition on RAVDESS Dataset Using Transfer Learning. \u003cem\u003eSensors\u003c/em\u003e\u003cstrong\u003e21\u003c/strong\u003e, 7665 (2021).\u003c/li\u003e\n\u003cli\u003eIssa, D., Demirci, M. F. \u0026amp; Yazici, A. Speech emotion recognition with deep convolutional neural networks. \u003cem\u003eBiomed. Signal Process\u003c/em\u003e\u003cstrong\u003e59\u003c/strong\u003e, 101894 (2020).\u003c/li\u003e\n\u003cli\u003eWu, G. Q. et al. Chaotic Signatures of Heart Rate Variability and Its Power Spectrum in Health, Aging and Heart Failure. \u003cem\u003ePLoS One\u003c/em\u003e\u003cstrong\u003e4\u003c/strong\u003e, e4323 (2009).\u003c/li\u003e\n\u003cli\u003eDadras, S. \u0026amp; Momeni, H. R. A novel three-dimensional autonomous chaotic system generating two, three and four-scroll attractors. \u003cem\u003ePhys. Lett. A\u003c/em\u003e\u003cstrong\u003e373\u003c/strong\u003e, 3637\u0026ndash;3642 (2009).\u003c/li\u003e\n\u003cli\u003eSprott, J. C. Elegant Chaos: Algebraically Simple Chaotic Flows (\u003cem\u003eWorld Scientific\u003c/em\u003e, Singapore, 2010).\u003c/li\u003e\n\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":true,"hideJournal":true,"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":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true},"keywords":"","lastPublishedDoi":"10.21203/rs.3.rs-7993782/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-7993782/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"On-skin artificial intelligence (AI) demands hardware that couples skin-like mechanics with efficient, real-time computation on noisy, spatiotemporal biosignals. We introduce a supramolecular polymer memtransistor that unifies intrinsic stretchability, autonomous self-healing and low-power neuromorphic dynamics in a single material platform. The device integrates a supramolecular elastomer matrix with a p-n heterojunction semiconductor to realize charge trapping-driven short-term plasticity, high on/off ratios (\u003e103) and tight device-to-device uniformity (σ/μ = 5.25%). Operating energies span 0.29 fJ-1.8 nJ per event, approaching the lower bound of biological synapses while retaining reliable control of synaptic weights. Arrays (7 × 7) serve as a physical reservoir for on-device reservoir computing, achieving \u003e99% accuracy in spoken-digit recognition and robust emotion recognition (74% under 30% biaxial strain; 71% after self-healing), all maintained during 30% biaxial deformation and after autonomous recovery from deliberate damage. Beyond classification, recursive multi-step forecasting with online learning stably models chaotic dynamics with normalized RMSE ≲ 0.02, sustaining accurate long-horizon predictions. These results establish supramolecular polymer memtransistors as a materials-driven route to elastic, damage-tolerant and energy-efficient neuromorphic electronics that perform AI inference and prediction directly on the body, enabling bio-integrated systems for speech, affect and complex physiological time-series analysis.","manuscriptTitle":"On-Skin Artificial Intelligence via Supramolecular Polymer Memtransistors","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2025-12-10 09:09:59","doi":"10.21203/rs.3.rs-7993782/v1","editorialEvents":[{"type":"communityComments","content":0}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"227208f5-d1ca-4a8b-a116-b5e0268ec6bd","owner":[],"postedDate":"December 10th, 2025","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"posted","subjectAreas":[{"id":59356752,"name":"Physical sciences/Materials science/Materials for devices/Electronic devices"},{"id":59356753,"name":"Physical sciences/Mathematics and computing/Computer science"}],"tags":[],"updatedAt":"2026-03-20T17:46:08+00:00","versionOfRecord":[],"versionCreatedAt":"2025-12-10 09:09:59","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-7993782","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-7993782","identity":"rs-7993782","version":["v1"]},"buildId":"8U1c8b4HqxoKbykW_rLl7","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}