Machine learning-assisted design of carbon nanotube edge computing circuits for monolithic epidermal systems Authors: Zebang Luo, Li Xiang, Xiongfeng Zou, Jie Liu, Hui Wang, Huaidong Ye, Yulong Yuan, Haitao Zhang, Xingge Yu, Youfan Hu, Anlian Pan Published: 2026 ABSTRACT AbstractThe rapid development of multimodal epidermal sensing requires scalable, energy-efficient data processing architectures capable of processing large volumes of raw data. Conventional systems suffer from high energy consumption and transmission latency due to the physical separation of sensors and processors. Here, we present an ultrathin flexible edge computing circuit based on carbon nanotube thin-film transistors (CNT-TFTs) and machine learning (ML)-assisted design. By incorporating substrate engineering, ML-derived device modeling, and industry-compatible design methodologies, we establish a complete toolchain from device to system. The ML model achieves 91.2% prediction accuracy, enabling simulation-guided optimization of logic gates. A CNT-based standard cell library enables the construction of flexible circuits with 361 transistors and 160 logic gates. Monolithic integration with an 8-channel tilt sensor achieves 62.5% data compression while maintaining functionality after undergoing 360° deformation. This work establishes an ML-assisted CNT circuit design framework for fully integrated flexible edge computing, enabling scalable wearable applications.Subject terms: Electrical and electronic engineering, Carbon nanotubes and fullerenes, Electronic devices Introduction The emergence of epidermal sensing and intelligent data processing in ultrathin wearable systems represents a critical frontier in next-generation healthcare diagnostics1–3, interactive robotics4,5, and wearable technologies6–8. While recent advances in epidermal electronics on ultrathin substrates, including the discrete sensors9, transistors10–12, and small-scale integrated circuits13,14, have achieved substantial progress in biotic-abiotic interface engineering, these electronics now confront a fundamental scalability challenge: increasing sensor-channel counts in multimodal monitoring systems generate exponentially growing raw data that overwhelm conventional signal processing architectures15–17. Such a challenge lies in the physical separation between the sensing modules and computational units, in which large volumes of unprocessed data are collected locally by sensors and must be transmitted to computational units or cloud-based systems, leading to significant concerns of high energy cost for data transmission18,19. Edge computing represents a distributed architecture paradigm that relocates processing resources from remote centralized units to the physical proximity of local sensing nodes20,21. This approach fundamentally reengineers data processing workflows by performing in situ feature extraction and data compression near the source of signal generation, which could significantly reduce data processing latency and data transmission burdens22. For ultrathin epidermal systems and other flexible integrated platforms, the predominant reliance on a hybrid integration strategy, where conventional rigid silicon chips are mounted on flexible substrates23–25, faces severe limitations. This approach demands either costly wafer thinning processes or area-consuming serpentine interconnects, which escalate manufacturing expenses, compromise mechanical robustness through strain-localization at the soft-rigid interfaces26, and fail to deliver fully flexible platforms, making hybrid integration not a viable long-term solution for scalable and cost-sensitive applications. Alternatively, monolithic co-integration of sensors and signal processing elements on ultrathin substrates, which were built exclusively with flexible materials and corresponding fabrication techniques, would be mass-producible, durable, and economically viable toward the Internet of Things (IoT) era15,27,28. Carbon nanotubes (CNTs) emerge as a promising semiconductor that combines high carrier mobility, intrinsic flexibility, and low-cost processability for large-area flexible electronics29–31. Significant progress has been made with CNT-based systems-on-chip on rigid silicon substrates, achieving integrations exceeding 1000 transistors in functional computation circuits, such as microprocessors, tensor processing units, and three-dimensional integrated sensing systems32. Critically, achieving comparable integration and functional complexity with CNTs on flexible ultrathin substrates remains a critical, unmet challenge33,34. So far, most flexible CNT-based circuits are limited to a few tens of transistors in integration scale35–37. Even the limited demonstrations of flexible CNT-based circuits incorporating hundreds of transistors are also confined to repeating the same stereotypes of inverters into simple ring oscillators33,38. These demonstrations serve as valuable proof-of-concept integration prototypes. However, they inherently lack computational capabilities to support sophisticated signal processing tasks, underscoring the urgent need for functionally complex CNT signal-processing circuits in practical, flexible integrated systems. In this article, we report edge computing circuits built on ultrathin flexible substrates (~2 μm) using CNT-based thin-film transistors (TFTs), addressing a critical challenge in monolithic flexible system integration. This natively flexible circuit achieves in situ monolithic integration with flexible sensors on the same substrate, enabling proximate data compression at the sensor node by encoding 8-channel sensory inputs into 3-bit digital outputs. Unlike previous ultrathin flexible electronic systems, our platform can both sense and process multichannel stimuli without relying on external rigid silicon chips or discrete signal-conversion circuits. The circuit integrates 160 logic gates (361 TFTs), representing a CNT-based flexible integrated system developed using industry-compatible design methodologies. This advance is enabled by two key strategies: substrate-engineered device fabrication and machine learning-assisted circuit design. To address the yield challenge in ultrathin (≤10 μm) electronics, we developed a self-delaminating substrate that enables both robust adhesion to temporary rigid carriers during fabrication and damage-free release afterward. Conventional separation methods (e.g., laser ablation39,40, mechanical peeling41–43, electrochemical etching37,44,45) often introduce mechanical stress, thermal damage, or chemical contamination, which severely reduce device yield. This issue is especially critical for integrated circuits, where yield decreases rapidly as device count increases. Here, the self-delaminating substrate enabled 100% successful release of the ultrathin device substrates without macroscopic fracture, and transistors selected from different regions of the delaminated ultrathin films exhibited normal electrical characteristics, thereby mitigating yield losses that are endemic to ultrathin device fabrication. In parallel, the lack of design-automation tools for CNT-based flexible electronics remains a major bottleneck, restricting circuit development to manual transistor-level optimization. To address this, we establish an industry-compatible framework that incorporates a machine learning-generated CNT transistor PDK with compact models and a standard cell library, thereby overcoming the design-automation bottleneck. This infrastructure enables automated design and rapid development of CNT-based edge computing circuits. Results Hardware concept of flexible edge computing The hardware concept of our ultrathin edge computing systems is illustrated in Fig. 1a. By monolithically co-integrating sensing units and signal processing circuits in physical proximity, our design enables proximate edge computation. Figure 1b contrasts conventional architectures, characterized by physical separation between sensors and processors (top panel). In such systems, sensor arrays generate massive signals requiring large-volume data transmission to centralized processors, in contrast to the edge computation paradigm. Crucially, the edge computation approach bypasses the inherent limitations through in situ signal processing and compression before transmission (Fig. 1b, bottom). This architecture establishes a fully integrated pathway for physical signal acquisition and preprocessing (detailed in Supplementary Fig. 1), representing a hardware strategy for scalable, flexible electronics. Flexible near-sensor computing system for data compression. a Illustration showing a flexible near-sensor computing system conformally laminated on the skin. b Comparison between conventional sensing and data-processing systems and the flexible near-sensor computing system. The upper panel illustrates the sensing and recognition process using a conventional sensor and external data-processing system, whereas the lower panel illustrates near-sensor data processing using the flexible near-sensor computing system. c Upper panel: Spatial layout of the computing circuits. Lower panel: Schematic illustration of a carbon nanotube (CNT)-based inverter. Scale bar, 1 μm. d Illustration showing the overall flow and flexible electronics of the near sensor computing system implementation. Figure 1c presents a three-dimensional schematic of the physical layout of our ultrathin CNT-based edge computing circuit. Fabricated exclusively with CNT TFTs on ultrathin plastic substrates (~2 μm thickness). The CNT transistors utilize a local bottom gate structure (fabrication details in Methods). Different from the standard silicon process, the metal layers for routing are co-designed with the transistors during circuit design to simplify processing and improve fabrication yield. For example, as depicted in Fig. 1c, the metal 1 layer (M1), which routes signals between functional units, is designed in the same plane as the transistor gate metal, while the metal 2 layer (M2) for intercell routing is positioned in the same plane as the source/drain metals. This configuration reduces routing congestion through vertical spatial decoupling of signal paths, thereby enhancing the integration density. Figure 1d illustrates the system architecture of our ultrathin electronics platform. Here, we use a custom-made tilt sensor as a representative example (also applicable to other sensor modalities). The tilt sensor generates distinct electrical signals across eight channels corresponding to angular orientation, which are processed in situ by the flexible monolithic edge computing circuits. This integrated framework performs on-chip signal conditioning, including amplification and clipping, to generate digital waveforms in real time, followed by encoding the eight-channel data into compressed three-bit outputs via a monolithically integrated digital unit. Critically, this approach achieves real-time data compression at the sensing interface, eliminating the need for high-volume raw data transmission in conventional sensing systems. Substrate engineering for ultrathin devices Substrate engineering forms an important basis for ultrathin flexible electronics, where thickness optimization must reconcile mechanical robustness, chemical stability and compatibility with micro-fabrication processes. Although polyimide (PI) thin films offer industrially standardized temperature and chemical stability compared to alternative plastic polymers, current approaches in industrial foundries, where PI is cured at high temperatures (~350°C) on temporary rigid carriers (e.g., glass/silicon wafers) to enable strong adhesion and withstand harsh semiconductor manufacturing process that need multiple thermal/chemical exposures, create an intrinsic limitation in post-fabrication separation. The present separation strategies face significant limitations: the dry process based on laser release39,40 requires sophisticated processing parameter optimization to circumvent laser-induced damage, thereby introducing additional process complexity; meanwhile, the wet process of chemical etching37,44,45 raises concerns of chemical contamination, which potentially compromises device reliability and yield. Here, we employed a rational molecular design strategy aimed at developing a polyimide (PI)-based substrate with simultaneous robust adhesion and self-delamination capability from the temporary silicon wafer carriers. By incorporating a proportion of siloxane (Si–O–Si) segments into the PI backbone through copolymerization (details in Supplementary Note 1), we achieved PI thin films that robustly adhere to silicon wafers without high-temperature curing (detailed in Methods). The introduced siloxane segments provide sufficient interactions with silicon wafers possessing dangling -Si bonds on their surfaces, demonstrating adhesion strength comparable to conventional, high-temperature-cured PI substrates under harsh semiconductor fabrication conditions. Crucially, as demonstrated in Fig. 2a, the PI-SiO foil can easily and spontaneously delaminate from silicon wafers when immersed in water. Atomic force microscopy (AFM, shown in Supplementary Fig. 4) reveals an interface with inherent roughness between the plastic foil and the silicon wafers, creating gaps at the interface that trigger capillary forces and water penetration. Initially, a slight trimming of the edges is required to ensure that the tightly sealed wafer boundary does not block water penetration. As immersion time increases, water progressively penetrates the interface (Fig. 2b), leading to complete self-delamination. Supplementary Video 1 confirms the very gentle delamination process, devoid of significant mechanical stress or damage. Meanwhile, the use of deionized water, rather than the harsh etchants or ionic solutions, ensures the delamination process is free from chemical contamination, eliminating risks to device yield and robustness. This combination of sufficient process-compatible adhesion and easy damage-free delamination enabled 100% device fabrication yield on ultrathin flexible substrates. Compared to conventional PI substrates of the same thickness, where devices are prone to cracking or spalling damage during separation (Supplementary Fig. 5), our substrate engineering technique offers a fundamentally simpler, non-destructive, and contamination-free pathway. The fabricated devices and circuits on the ultrathin substrates could be deformed (e.g., crumpling, Fig. 2c) and enable intimate conformal contact to diverse biological surfaces, such as the plant leaf (Fig. 2d) and human palms (Fig. 2e). Self-delamination film and characteristics of carbon nanotube thin-film transistors. a Schematic illustration of the self-delamination process. b Photograph showing the self-delamination process of the flexible near-sensor computing system. c Photograph of the device after crumpling without damage. d Photograph of the device attached to a leaf surface. e Photograph of the device attached to a human palm. f Optical microscope image and transfer curves of 100 carbon nanotube thin-film transistors with W/L = 10 measured at Vds = −0.1 V. Scale bar, 100 μm. g–j Corresponding statistical distributions of the current on/off ratio Ion/Ioff, width-normalized off-state current Ioff/W, threshold voltage Vth, and subthreshold swing, respectively. k Output characteristics of a representative carbon nanotube thin-film transistor with W/L = 10. Uniformity of flexible transistors The flexible edge computing circuits presented in this article utilize an established wafer-scale fabrication process for CNT TFTs, enabling circuit manufacture on 2-inch substrates through sequential deposition, patterning, and etching cycles (Details present in Methods). Robust circuit design necessitates precise characterization of device performance and uniformity as critical process parameters. To this end, we systematically measured the electrical characteristics of 100 p-type CNT TFTs (W/L = 100/10 μm) distributed across the ultrathin flexible substrate. Figure 2f reveals narrow dispersion in the transfer characteristic curves, indicative of high-performance uniformity across the devices. Statistical analysis of the current on/off ratio (Fig. 2g) shows a close fit to a Gaussian distribution, with a high average value of 5.50 ± 0.17 decimal units, signifying clear discrimination between the on and off states. Further characterization (Fig. 2h) demonstrates very low normalized off-state currents (Ioff/W), with 61% of the TFTs displaying <0.075 pA μm−1 and 98% of the TFTs displaying <0.475 pA μm−1, pointing to small static power dissipation. The threshold voltage (Vth), extracted via maximum transconductance extrapolation (Supplementary Fig. 6), exhibits a narrow distribution at −0.73 ± 0.05 V (Fig. 2i). Critically, the small standard deviation in Vth (50 mV) occupies only 4% of the operational gate voltage window (1.25 V), ensuring precise switching threshold control essential for complex circuits. Further characterization yields a subthreshold swing (SS) of 104.7 ± 7.6 mV dec−1 (Fig. 2j), calculated as ΔVgs/Δlog10(Ids), and a field-effect mobility (Supplementary Fig. 7) of 33.50 ± 3.56 cm2 V−1 s−1 (extraction details shown in Supplementary Fig. 8), validating performance suitable for further integration. Representative output characteristics (Fig. 2k) exhibit well-defined linear and saturation regimes across a gate voltage sweep (Vgs = −0.1 V to −2 V, step −0.1 V; Vds = 0 to −2 V), confirming effective gate modulation and channel pinch-off for implementing robust logic functions in circuit integration. To evaluate the mechanical flexibility of the CNT TFTs on ultrathin PI-SiO substrates, we systematically characterized their electrical performance under progressive bending conditions (Supplementary Fig. 9). Optical images (Supplementary Fig. 9a–d) visually confirm device integrity across diverse bending radii of 1.87, 1.31, 1.01 and 0.77 cm. Remarkably, transfer characteristics acquired during bending (Supplementary Fig. 9e) exhibit negligible deviation, demonstrating stable electrical operation under mechanical strain. The key parameters, including threshold voltage (Vth), subthreshold swing (SS), and current on/off ratio, remained highly stable with negligible changes across all bending radii (Supplementary Fig. 9f). To further assess the mechanical durability of the devices under repeated deformation, we additionally performed cyclic bending tests up to 3000 cycles, with electrical characterization carried out every 300 cycles (Supplementary Fig. 10). The transfer characteristics remained highly consistent throughout the cycling test, while the extracted on/off ratio (lg(Ion/Ioff)) and threshold voltage (Vth) showed only minor variation over the full 3000-cycle measurement, with lg(Ion/Ioff) of 5.09 ± 0.11 and Vth = −0.57 ± 0.007 V. These results further confirm that repeated bending induces minimal degradation in the characteristics of the CNT-TFTs. These results establish the mechanical robustness of our CNT-TFTs on the ultrathin flexible substrate, fulfilling a fundamental requirement for reliable flexible and wearable electronics. Design of flexible computing circuits Figure 3a presents industry tool-compatible hierarchical design flows of our ultrathin edge computing circuit, comprising three co-optimized stages: process design, layout design, and fabrication process (demarcated in blue, yellow, and green frames, respectively). Firstly, the process design stage (blue frame in Fig. 3a) seamlessly adapts the standard silicon-based process design kits (PDKs) to CNT technology, which is developed through CNT TFT device modeling, while maintaining full compatibility with unmodified commercial EDA tools (e.g., parameterized cells, physical verification decks, and I/O libraries). This approach preserves established tools and enables us to leverage decades of existing EDA infrastructures to design and widely adopt in the mainstream semiconductor industry. Subsequently, the layout design stage (yellow frame in Fig. 3a) incorporates an experimentally validated standard logic cell library (comprising 12 cells, including NOT, NOR, NAND gates, etc.), fabricated using our previously established CNT TFT manufacturing processes, to serve as fundamental circuit building blocks. Automated routing and placement within industrial EDA environments optimized signal integrity and area efficiency based on pre-verified electrical/geometric properties of these cells. Finally, wafer-scale fabrication (green frame in Fig. 3a) translated the geometric layouts into practical functional circuits, employing material stacks and low-temperature processes (<150°C) specifically engineered for compatibility with ultrathin polymeric substrates. Design flow and machine learning (ML)-assisted compact modeling for edge computing circuits. a Design and manufacturing flow for flexible near-sensor computing circuits. The content in the gray frame is the ML-assisted carbon nanotube (CNT)-based transistor compact modeling process flow. b Training epoch loss over time, with a smoothed trend showing gradual convergence. c Training epoch R² indicating the improvement in model fitting as training progresses. d Comparison of ML-predicted current values with target sample data. e Comparison of ML-prediction, spice and experiment results for lg Ids–Vgs at different drain voltages and different transistor sizes. Rectangles represent ML outputs, triangles represent SPICE outputs and lines correspond to experimental results (Exp.). Different colours correspond to different W/L and Vds conditions. f Voltage transfer characteristics of inverters with different sizes under SPICE simulation based on ML. g Inverter transient response of SPICE simulation with ML compact model. Precise device modeling of flexible CNT TFTs is fundamental for designing and simulating corresponding integrated circuits. Therefore, it is noteworthy that we established a CNT TFT compact model for circuit simulations based on a machine learning (ML)-assisted approach. Conventional physics-accurate compact modeling strategies, which are reliant on semi-empirical equations derived from idealized device physics, prove inadequate and challenging in capturing the multidimensional correlations between CNT network morphology (density, alignment and purity) and electrical characteristics. This limitation is exacerbated in flexible CNT devices, where polymeric substrates and low-temperature processing would introduce complex mass defects, interface charges, and non-ideal carrier transport mechanisms. To address these limitations, we developed an ML-assisted framework through systematic experimental design: fabricating a library of CNT-TFT variants with different channel geometries, followed by standardized electrical characterization to extract transfer curves (Supplementary Fig. 11) and key parameters (e.g., Vth, SS, Ion/Ioff, Rc, etc.). Following preprocessing of the measured data, we trained a multilayer perceptron (MLP) neural network architecture to map the geometric parameters of the device to its electrical characteristics, such as transfer and output characteristics. Here, to enhance the neural network training efficiency, Min-Max normalization during data preprocessing scaled input features to the [0,1] range, ultimately achieving 91.2% prediction accuracy for drain current. Figure 3b, c demonstrates stable training convergence, with loss decreased from 0.407 to 0.389 while the coefficient of determination (R2) improved from 0.992 to 0.9985, indicating effective nonlinear mapping without overfitting. Subsequently, the trained neural network was converted into a compact model with Verilog-A format compatible with commercial EDA platforms (e.g., Cadence Virtuoso). The simulation program with integrated circuit emphasis (SPICE) results using this model replicated experimental I-V characteristics at a root-mean-square error (RMSE) of 0.057. Validation under operational bias (Fig. 3d and Supplementary Fig. 12) revealed close alignment between ML predictions and experimental data, with SPICE-simulated transconductance deviations below 0.28% for devices (W/L = 15, Vds = −0.1 V). Systematic validation of our ML-derived compact model demonstrates multi-parameter consistency between simulated and experimental electrical characteristics. Figure 3e and Supplementary Fig. 13a reveal closely aligned predictions for both drain current (Ids) and transconductance (gm) across varying W/L ratios and drain voltages (Vds). Comprehensive validation over diverse Vds further confirms model robustness, with Supplementary Fig. 13b, c exhibiting a small deviation in Ids and gm prediction accuracy. Building on this foundation, Supplementary Fig. 14 extends the validated framework, predicting transfer characteristics across different W/L ratios (Supplementary Fig. 14a) and voltage-dependent responses for a device with fixed W/L = 10 (Supplementary Fig. 14b). Beyond validation within a single fabrication environment, we examined the robustness and generalizability of the ML-assisted framework using device datasets fabricated in two additional laboratories (Lab 2 and Lab 3), resulting in three laboratory datasets in total. For direct comparison, 20 device structures were selected across all datasets under the same testing conditions. Supplementary Fig. 15a, b shows the transfer characteristics of these devices measured from Lab 2 and Lab 3, while Supplementary Fig. 15c, d further shows the Vds-dependent transfer characteristics of these devices. These results confirm that the fabricated devices retain physically consistent structure-dependent and bias-dependent electrical behavior across laboratories. We then applied the same ML-assisted modeling workflow to the Lab 2 and Lab 3 datasets, including data preprocessing, model training and prediction, followed by conversion into SPICE-compatible compact models. Based on these compact models, SPICE simulations were performed for direct comparison with the corresponding experimental results. As shown in Supplementary Fig. 16a–f, the transfer characteristics and corresponding transconductance (gm) of devices from Lab 2 and Lab 3 show close agreement between experiment and SPICE simulation, indicating that the framework can capture the relevant electrical characteristics of devices fabricated in different laboratories and batches. To quantitatively assess the predictive stability of the framework across independent fabrication batches, we further analyzed the same matched 20-device set from the three laboratory datasets using extracted key parameters, including subthreshold swing (SS), threshold voltage (Vth), and current on/off ratio (lg(Ion/Ioff)), as shown in Supplementary Fig. 16g–l. Here, the relative deviations for SS and lg(Ion/Ioff), as well as the deviation for Vth, are defined as follows: Where, the SSSPICE, lgIonIoffSPICE and Vth−SPICE represent the simulated results of SPICE models, while the SSExp, lgIonIoffExp, and Vth−Exp are extracted data from the experimental results. The relative deviations in SS and lgIon/Ioff, together with the deviation in Vth, all remain within narrow and comparable ranges across the three laboratory datasets. Specifically, the average relative deviations are ~0.03–0.04 for SS and ~0.05–0.10 for lgIon/Ioff, while the average deviation in Vth is ~0.03–0.04 V. Importantly, deviations between SPICE prediction and experimental results are at nearly a constant level for devices from the different fabrication batches (Lab 2 and Lab 3). Overall, the consistent agreement between SPICE-simulation and experimental results across multiple datasets demonstrates the reliability of the proposed framework. Such agreement is maintained across datasets obtained from different laboratories and fabrication batches, indicating that the framework can effectively capture device behavior in a batch-independent manner and deliver stable predictive performance across data sources. Together, these results validate its predictive accuracy and scalability. Crucially, the model’s circuit-level predictive capability is further demonstrated through inverters, which serve as a fundamental building block of integrated circuits. As shown in Fig. 3f and Supplementary Fig. 17, our ML-assisted SPICE simulations efficiently generate voltage transfer characteristics (VTCs). In these simulations, the driver transistor was fixed at W/L = 2, while the load transistor geometry was systematically varied to enable automated optimization of output swing, noise margins, and voltage gain. Leveraging this capability, we implemented a five-stage cascaded inverter (driver transistor: W/L = 2, load transistor: W/L = 15), which achieves robust digital switching with 95% voltage swing preservation across stages (Fig. 3g), thereby validating the model’s seamless transfer from individual device physics to complex circuit behavior. Statistical uniformity of basic logic inverters As an example for practical circuit optimization, leveraging our ML-assisted design framework, we designed and fabricated inverters with strategically optimized transistor width-to-length (W/L) ratios—notably implementing a significantly reduced input-stage ratio versus the output stage (Fig. 4a). This geometric optimization was iteratively optimized via ML-empowered SPICE simulations, achieving nearly rail-to-rail output voltage switching, reaching close-to-VDD output state (Supplementary Fig. 18). Voltage transfer characteristic (VTC) analysis quantified inverter metrics: full-swing switching from 5 V to near 0 V output under VDD of 5 V (Supplementary Fig. 19), symmetrical noise margins (NMH = 40% VDD, NML = 41% VDD) surpassing prior implementations based on flexible CNT-TFT, and switching threshold centered at approximately half VDD (2.45 V ± 0.07 V). Carbon nanotube (CNT) inverters and ring oscillator characterization. a Optical images of a typical inverter. Scale bar, 50 μm. b Voltage transfer characteristics (VTCs) of 150 inverters (Vdd = 5 V). Inset, Circuit diagram. c Key voltage metrics extracted from the 150 measured inverter logic gates. d Statistical distribution of voltage swing, Swing=(VOUT,MAX−VOUT,MIN)/Vdd. e statistical distribution of inverter voltage gain. f Statistical distribution of switching threshold voltage (Vsth). g, h Static noise margins extracted from the distributions in (c), where SNMH=(VOH−VIH)/Vdd and SNML=(VIL−VOL)/Vdd (g, h). i Optical image of the ring oscillator. Scale bar, 100 μm j Input–output characteristics of the ring oscillator. k Measured oscillation frequency of the ring oscillator as a function of supply voltage. While the optimized W/L configuration shows the targeted electrical behavior in individual devices, the statistical uniformity of logic inverters serves as a critical benchmark for integrated circuit viability. Comprehensive characterization of voltage transfer curves across 150 devices (Fig. 4b) confirms high process uniformity in our CNT technology, evidenced by 95 ± 0.007% voltage swing preservation (Fig. 4d), and narrow distributions at switching points (VIL, VIH, VOL, VOH; Fig. 4c). The compound advantages of high gain (101.5 ± 8.3; Fig. 4e), matched logic levels, and symmetric noise margins (NMH = 39.2% ± 2% VDD, NML = 40.7% ± 2% VDD; Fig. 4g, h) provide the noise tolerance and voltage headroom essential for overcoming logic depth limitations in flexible CNT-based circuits. Based on the demonstrated electrical uniformity of the inverters, we integrated five identical inverter stages into a functional ring oscillator (RO) (Fig. 4i). Operating at 5 V supply voltage, the oscillator generated a stable signal with a frequency of 91 kHz (Fig. 4j, corresponding to a 1.1 μs propagation delay per stage. Frequency tuning across a 3–7 V supply range (Fig. 4k) maintained waveform integrity while exhibiting linear voltage dependence (Supplementary Fig. 20). Given the extreme sensitivity of ring oscillators to variations in inverter parameters, including noise margin, voltage gain, and switching threshold, this predictable, condition-invariant oscillation behavior directly validates the excellent electrical reproducibility and uniformity achieved across our fabricated inverters. Basic cells for constructing edge computing circuits Parameterized cell libraries are foundational elements for integrated circuit design, providing transistor technology-specific layout templates and design rules. As delineated in Fig. 3a (yellow frame), our layout design starts with the strategic adaptation of a silicon CMOS standard cell library for carbon nanotube (CNT) technology. Through rigorous functional pruning, we developed a cell library as listed in Supplementary Table 1, comprising 12 experimentally validated CNT-TFT cells optimized for target applications. The cells were fabricated and characterized to verify functionality and performance metrics, establishing a fully qualified standard cell library for CNT-based circuit design. We then demonstrate the fundamental PMOS logic gates, which serve as the standard cells for edge computing circuits through systematic integration. As shown in Fig. 5 (with corresponding device diagrams provided in Supplementary Fig. 21), all demonstrated logic cells use the CNT TFTs with “Zero-Vgs” load configurations as the pull-down networks, whereas other CNT TFTs form the pull-up network. The fabricated NOT, NOR, NAND, and AND gates, as illustrated in Fig. 5 a, c, e, g, exhibit rail-to-rail switching between 0 V (logic “0”) and 5 V (logic “1”) (Fig. 5 b, d, f, h). This performance is extended to three-input NAND gates (Supplementary Fig. 22). Critically, the observed full-swing switching directly results from geometry optimizations enabled by our developed design tools. Logic gates and computing circuits characterization. a–h Pseudo-colored optical images of NOT, NOR, NAND and AND gates and their corresponding input–output characteristics. Scale bars, 100 μm. i Optical image of the near-sensor computing circuit and corresponding circuit diagram. Scale bar, 500 μm. j Input–output characterization of the computing circuit. Beyond static logic, we designed and fabricated wave-shaping circuits using cascaded inverters to transform low-amplitude sensor signals into standardized digital waveforms. The four-stage cascaded inverters (Supplementary Fig. 23a) progressively amplify and clip input signals within the inverter transition region, converting sinusoidal inputs into rail-to-rail square waves. Input-output characterization (Supplementary Fig. 23b) reveals the device’s ability to amplify input signal amplitudes and its intrinsic signal clipping behavior due to the supply voltage limitation. Owing to the stable operation of the established basic cells, we construct the edge computing circuit by standard placement-and-routing procedures within existing EDA tools, which is shown in yellow frames of Fig. 3a. This circuit on the ultrathin flexible substrate integrates 361 CNT TFTs configured into 160 logic gates (Fig. 5i), occupying 0.168 cm2 and achieving a transistor density of 2150 devices cm−2. The system’s signal processing capability was evaluated through phase-modulated inputs of identical frequency (Supplementary Fig. 24), successfully encoding multi-channel inputs into compressed outputs as per designed truth tables (Supplementary Fig. 25). Furthermore, this circuit implements descending priority ranking from I7 to I0, enabling deterministic identification of the highest-priority active inputs (Fig. 4j). This priority-aware architecture enables deployment in multi-sensor networks requiring concurrent event processing. Integrated sensing and computing system Building upon our CNT-based computing circuits, we demonstrate a flexible edge computing system that monolithically integrates a multi-channel tilt sensor and embedded processing circuits on a single flexible substrate. As illustrated in Fig. 6a, the integrated flexible system combines an 8-channel tilt sensor and the 8-input edge computing core, achieving terminal-to-terminal signal processing from physical motion detection to compressed digital output. Here, the detailed fabrication process of the tilt sensor is presented in Method, and Supplementary Fig. 26. A photograph of the entire system is shown in Fig. 6b, with a compact layout and high integration density. The system exhibits great mechanical flexibility, as demonstrated in Fig. 6c. The upper panel shows the system mounted on a human hand. In tests, the integrated system remained undamaged through various wrist flexion angles and full 360° rolling deformation (lower panel) without interconnect failure (see also Supplementary Fig. 27), confirming its wearability for ultrathin flexible electronics. Flexible near-sensor computing system. a Schematic illustration of the flexible near-sensor computing system architecture, integrating sensing, preprocessing and computation. b Photograph of the fabricated flexible near-sensor computing system. c Demonstration of the mechanical flexibility of the device, including attachment to human skin and bending into a rolled configuration. d Three-stage signal-processing flow implemented in the system. e Real-time signal-processing performance of the integrated flexible near-sensor computing system under different motion directions. The upper panels show the signals after amplification and clipping, whereas the lower panels display the digitally compressed outputs generated by the computing circuits. The three-stage signal-processing pipeline (Fig. 6d) begins with tilt-direction detection using an 8-directional sensor, followed by signal amplification and clipping to standardize the sensor outputs into 8-channel digital inputs. These inputs are subsequently processed by the edge-computing core through priority-aware encoding, compressing eight channels (I7–I0) into three output bits and thereby achieving 62.5% data reduction while preserving the directional information content. Real-time operation (Fig. 6e) demonstrates immediate processing of sensor-generated signals during hand rotation. To clarify the operating mechanism, a schematic illustration of the overall working process is provided in Supplementary Fig. 31. During hand rotation, the liquid-metal droplet in the tilt sensor dynamically changes its position with the motion, and once the motion stops, the final tilt direction determines which sensor channel is activated. For example, when the hand rotates leftward and reaches the state corresponding to channel 6, the liquid-metal droplet comes into contact with channel 6, thereby establishing an electrical conduction path and activating this sensing channel. As a result, channel 6 provides a logic-high input to the circuit, while the other channels remain at logic low. With the circuit inputs defined from top to bottom as I7 to I0, this state corresponds to the input code “01000000”, which is then processed by the digital circuit into the corresponding compressed output. Crucially, this monolithic near-sensor computation architecture eliminates external bulky processors and signal conversion chips, significantly enhancing data processing efficiency while reducing latency and energy consumption. Such capability is particularly advantageous in human–machine interaction scenarios, where immediate and accurate recognition of motion direction is crucial. To quantify the operating power consumption of the integrated circuit, we measured the time-dependent output current under a supply voltage of VDD = 3 V during continuous operation (Supplementary Fig. 32). Supplementary Fig. 32a shows the measured operation current IDD as a function of time, indicating stable circuit operation throughout the measurement period. Supplementary Fig. 32b presents the calculation of the average operating power (Pavg), where I(t) represents the measured current, total represents the total measurement time, and VDD represents the supply voltage. From this analysis, the average operating power of the 361-transistor circuit is ∼68.8 μW. These results provide quantitative support for the stable low-power operation of the proposed near-sensor computing system. Discussion In conclusion, we developed a machine learning-assisted, industry-compatible framework for CNT-TFTs, enabling compact modeling, systematic logic gate optimization, and scalable circuit construction on ultrathin flexible substrates. The demonstrated circuits achieved high electrical uniformity, robust signal processing, and real-time data compression in a monolithically integrated tilt-sensing system, which could significantly help epidermal electronic systems reduce data-transmission burdens in wearable applications. Taken together, these results establish a practical route toward flexible edge-computing platforms that combine device-level uniformity, circuit-level functionality, and system-level information processing in a single ultrathin form factor. Beyond the digital sensing modality demonstrated here, the proposed monolithic platform may also be extendable to analog sensing applications, such as electrophysiological monitoring, provided that additional signal-conditioning functions, including amplification, filtering, and signal conversion, can be integrated. Addressing these requirements will require further advances in device uniformity, noise suppression, operational stability, and power efficiency. At the same time, further advances in model interpretability and sensitive analysis, large-area process uniformity, interconnect/routing strategies, signal integrity, fabrication yield, and more comprehensive standard-cell/design infrastructure will be important for extending this framework toward broader sensing modalities and much larger CNT-TFT circuit systems. Methods Substrate fabrication First, under a nitrogen atmosphere, (1–x) equivalents of 4,4’-diaminobis(2-trifluoromethylphenyl) (TFDB) were added into a reaction flask containing DMAc, and the mixture was stirred for 2 min to dissolve. Then, 1 equivalent of 3,3’,4,4’-biphenyltetracarboxylic dianhydride (BPDA) was added, and the mixture was heated to 50 °C and reacted for 30 min. Once the solution became clear, x equivalents (x = 1%, 3%, 5%) of Si-430 were added, and the mixture was stirred for 6 h, followed by vacuum degassing for 30 min, yielding the PAA precursor solution. Subsequently, a silicon wafer was used as a rigid support, and the PAA precursor solution was spin-coated onto a wafer at 4000 rpm, forming a PAA wet film. The wafer with the PAA film was then placed in an oven for thermal imidization. The temperature was initially set to 100 °C, and then increased stepwise by 100 °C per hour until reaching 250 °C, converting the PAA wet film into a PI-SiO film. Flexible tilt sensor fabrication The PI (30 μm, Kapton HN) films were cleaned with ethanol before processing. Firstly, a computer-designed pattern was written on the cleaned PI thin films using a CO2 laser platform (PLS6.150D, Universal Laser Systems) (wavelength of 10.6 μm, power ranging from 12.5 to 15 W, and writing speed of 230 mm/s). The laser platform was written line-by-line on the PI thin film to convert the PI into LIG. The whole laser writing process was carried out in air. After the laser processing, the PDMS precursor was spin-coated onto the LIG-patterned PI surface at 1000 rpm, followed by thermal curing at 100 °C for 1 h. Once fully cured, the PDMS layer was peeled off, successfully transferring the LIG patterns onto the flexible PDMS substrate. Then, the PDMS surface was further processed using a laser to locally roughen specific areas. The central region was then selectively treated to enhance hydrophilicity. Liquid metal was subsequently dispensed into these hydrophilic zones, completing the fabrication of the multi-channel tilt sensor structure. Transistor fabrication For bottom-gated carbon nanotube (CNT) thin film transistors (TFTs), Ti/Au (5/30 nm) local gates were first defined on the substrate using standard photolithography, followed by electron beam evaporation (EBE) and a lift-off process. Next, a 10 nm HfO2 dielectric layer was deposited by atomic layer deposition (ALD), and patterned using standard photolithography and inductively coupled plasma reactive ion etching (ICP-RIE) in an SF6/O2/Ar atmosphere. Carbon nanotubes were then deposited onto the substrate via a dip-coating method. Subsequently, the CNT film was patterned using photolithography and ICP-RIE. Finally, source and drain electrodes (Ti/Pd/Au, 5/30/10 nm) were fabricated on the CNT film using photolithography and electron beam evaporation. Compact model generation CNT thin-film transistors (TFTs) with various channel width-to-length (W/L) ratios were designed and fabricated using MEMS-based processes. Transfer characteristics of each device were experimentally measured, and key parameters, including threshold voltage, on/off ratio, and subthreshold swing, were extracted. The extracted dataset was preprocessed using Min–Max normalization to scale all features to the range 0–1, facilitating faster convergence and reducing bias during neural network (NN) training. The NN model was trained and evaluated to ensure convergence, as monitored by the decrease in loss and the increase in the coefficient of determination (R2) over training epochs. Following training, the NN model was converted into Verilog-A format for compatibility with the Cadence Virtuoso SPICE simulation platform. Model accuracy was validated by comparing ML predictions, SPICE simulations, and experimental data for selected W/L ratios and drain voltages. The validated compact model was then employed in circuit-level SPICE simulations. Electrical characterization All electrical measurements were carried out in air at room temperature using a semiconductor parameter analyzer (Keysight B1500), digital oscilloscope (Siglent SDS1024X) and signal generators (Siglent SDG1022X). The output analysis of the computing circuits was implemented by a logic analyzer (DSLogic, DreamSourceLab, Co., Ltd.). A better electrical performance is expected if the vacuum level of the metal evaporation chamber can be improved. Human participant statement The on-skin demonstrations shown in this study involved voluntary participation and did not include physiological intervention or personal data collection. According to institutional guidelines, formal ethical approval was not required. Informed consent was obtained from all participants for participation and publication of images. Supplementary information Supplementary Information Description of Additional Supplementary Files Supplementary Video 1 Transparent Peer Review file Acknowledgements The authors acknowledge the support from the Hunan Institute of Optoelectronic Integration, College of Materials Science and Engineering, Hunan University, and the Hunan Institute of Advanced Sensing and Information Technology for device fabrication and characterization, Xiangtan University. Author contributions LX and YH conceived the experiment. LX, YH and AP supervised this project. LX and ZL designed the devices and circuits. ZL fabricated the devices, performed the electrical measurements, and analyzed and interpreted the data with input from XZ, JL, HW, HY, YY, HZ, XY, and LX and ZL. The manuscript was written with contributions from all authors, and all authors approved the final version of the manuscript. Peer review Peer review information Nature Communications thanks the anonymous reviewer(s) for their contribution to the peer review of this work. A peer review file is available. Funding This work was supported in part by the National Key R&D Program of China under Grants 2024YFB4405300, 2022YFA1402501 and 2022YFA1204300; the National Natural Science Foundation of China under Grants 62101181, 52221001, 62090035, 52372146 and U22A20138; the Natural Science Foundation of Hunan Province under Grant 2023JJ20016; and the Key Research and Development Plan of Hunan Province under Grants 2022GK3002 and 2023GK2012. Data availability The source data supporting the findings of this study have been deposited in Zenodo at 10.5281/zenodo.2015129446. Additional data are available from the corresponding authors upon request. Code availability The code supporting the findings of this study has been deposited in Zenodo at 10.5281/zenodo.2015129446. Additional code is available from the corresponding authors upon request. Competing interests The authors declare no competing interests. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. These authors contributed equally: Zebang Luo, Li Xiang, Xiongfeng Zou. Contributor Information Li Xiang, Email: xiangli93@hnu.edu.cn. Youfan Hu, Email: youfanhu@pku.edu.cn. Anlian Pan, Email: anlian.pan@hnu.edu.cn. Supplementary information The online version contains supplementary material available at 10.1038/s41467-026-74053-5. References Associated Data Supplementary Materials Supplementary Information Description of Additional Supplementary Files Supplementary Video 1 Transparent Peer Review file Data Availability Statement The source data supporting the findings of this study have been deposited in Zenodo at 10.5281/zenodo.2015129446. Additional data are available from the corresponding authors upon request. The code supporting the findings of this study has been deposited in Zenodo at 10.5281/zenodo.2015129446. Additional code is available from the corresponding authors upon request.