The fabrication process of the flexible sensor array is illustrated in Figure 2A. Each sensing unit consists of an mPDMS substrate integrated with an MWCNT conductive layer and flexible electrodes. The mPDMS substrate was fabricated via a replica-molding strategy. A structured mold was first prepared using selective laser sintering (SLS), where nylon powder was selectively fused layer-by-layer under laser irradiation. The polydimethylsiloxane (PDMS) prepolymer and curing agent (Sylgard 184, Dow Corning) were mixed at a mass ratio of 10:1, stirred for 10 min, and then degassed under a vacuum for 30 min. The mixture was cast into the SLS-fabricated mold and cured at 80 °C for 2 h. After cooling, the elastomer was peeled off, thereby replicating the microstructured surface. To promote interfacial adhesion, the mPDMS surface was treated with oxygen plasma for 30-60 s to enhance surface wettability and facilitate subsequent deposition.

MWCNTs were dispersed in ethanol at a concentration of 1 mg/mL, followed by ultrasonication to obtain a homogeneous suspension. A conductive MWCNT network was then deposited onto the mPDMS surface via physical vapor deposition (PVD). The deposition was performed in a vacuum chamber with a base pressure of approximately 5 × 10^-3 Pa and a working pressure of approximately 0.5 Pa, while the substrate was maintained at 25 ± 3 °C. The deposition duration was controlled within 5-10 min. Flexible electrodes were fabricated by first depositing copper nanoparticles onto a rigid substrate via drop-casting, after which laser patterning was used to define the electrode geometry. Polyimide (PI) was then cast and thermally cured to form a flexible substrate. After release, the patterned copper layer was embedded onto the PI surface, forming a flexible conductive electrode. Finally, the MWCNT-coated mPDMS layer and the flexible electrode layer were assembled in a face-to-face configuration and encapsulated using a PDMS film, resulting in the fully integrated flexible-pressure-sensor array.

The entire device consisted of a 3 × 3 nine-channel flexible pressure sensor array with a sensing area of approximately 14 × 14 mm. The actual object and size are presented in Supplementary Figures 1 and 2. Notably, the flexible sensor array exhibited an ultralight design, with a mass of 0.15 g and a thickness of 0.43 mm. Supplementary Figure 3A presents scanning electron microscopy (SEM) images of the PDMS surface without SLS treatment at different magnifications. The PDMS surface is shown to be relatively smooth and without a distinct surface structure. Figure 2B shows an SEM image of the mPDMS, wherein the surface observed after SLS treatment exhibited a pronounced microstructured morphology. SEM images of the MWCNT/mPDMS system at different magnifications are presented in Figure 2C and Supplementary Figure 3B. The microstructure remains clearly observable even at a magnification of 10,000×. Figure 2D and Supplementary Figure 4 show three-dimensional topography images of the PDMS and MWCNTs/mPDMS surfaces, and the surface line-scan roughness information is provided in Supplementary Figure 5. The PDMS surface exhibited a small height fluctuation with a roughness of only 0.765 μm, suggesting a relatively smooth surface with minimal topographical variation. By contrast, the MWCNT/mPDMS surface exhibited significant height variation, with a maximum height difference of approximately 65 μm and a surface roughness of 14.32 μm. These characterization results confirm the successful fabrication of a PDMS film with a surface microstructure via SLS. SEM and three-dimensional surface profilometry analyses were employed to elucidate the structural characteristics introduced by SLS. The precise replication of the designed microstructures on the PDMS surface was confirmed, and their effect on the sensing performance was investigated. The mPDMS layer exhibited a significantly higher surface roughness (Ra ≈ 14.32 μm) compared with the pristine PDMS film (Ra ≈ 0.76 μm), indicating an enlarged effective contact area and enhanced interfacial friction at the electrode–polymer interface under compression. This roughened morphology facilitates the formation of multiple tunneling pathways among the embedded MWCNTs, thereby improving the pressure-dependent conductance response. By contrast, the smooth PDMS surface reduced the dynamic modulation of the conductive network, which lowered the piezoresistive sensitivity.

Supplementary Figure 6A and B presents the Raman and Fourier transform infrared (FTIR) spectra of the MWCNT/mPDMS system, respectively, which confirm that MWCNTs were successfully deposited on the surface of mPDMS. The elemental distribution of mPDMS is presented in Supplementary Figure 7A-C, wherein Si is shown to be dominant and C and O are present in relatively low proportions. These results are consistent with those of the energy spectrum characterization [Supplementary Figure 7D]. The elemental distribution of MWCNT/mPDMS is presented in Figure 2E-G, and the corresponding spectral results are shown in Supplementary Figure 8. Here, the C and O contents are significantly greater than those of mPDMS, and the peak intensity of C is slightly increased, indicating the successful deposition of MWCNTs onto the mPDMS surface.

A wireless microcontroller-based acquisition module was developed to enable real-time transmission of the sensing signals [Supplementary Figures 9 and 10]. The performance of the flexible array was first characterized at the unit level. Calibration results from five fabrication batches exhibited exceptional uniformity, yielding a mean sensitivity of 1.364 kPa^-1 and excellent linearity within 0-100 kPa [Figure 3A and B]. This near-linear response is attributed to the progressive contact evolution of the microstructured surface and the continuous modulation of the MWCNT conductive network. The working principle of the sensor is illustrated in Figure 3C, where the surface microstructure enables gradual contact modulation under zero-, low-, and high-pressure conditions, resulting in increased conductive pathways and corresponding electrical signal variations. The relative sensitivity deviation across batches was as low as 2.17%, while the calibrated linearity and repeatability errors remained below 2.6% and 1.3%, respectively [Supplementary Figure 11]. These results confirm the high reproducibility of the preparation process and provide a solid basis for manufacturable large-scale integration. Compared with representative thin-film pressure sensors reported in recent years [Supplementary Table 1], both wide-range detection and high sensitivity were achieved, overcoming the traditional tradeoff between compliance and measurement precision.

The high sensitivity is attributed to the hierarchical microstructured surface of the MWCNT/mPDMS composite. At low pressures, limited contact between the electrode and concave microstructures results in low baseline conduction; as pressure increases, the progressive contact growth reduces the tunneling resistance, resulting in an exponential-to-linear transition. Once the microstructures are fully compressed, further loading leads to a saturated response, forming a quasilinear region for calibration. This mechanism accounts for the excellent sensing precision and stability. The I–V curves display highly linear behavior [Figure 3D] and minimal frequency dependence [Figure 3E], while the cyclic loading tests show reproducible amplitudes with a hysteresis error of only 4.75% F.S. [Figure 3F and G]. The combination of low hysteresis and a high signal-to-noise ratio ensures accurate dynamic transduction for continuous handwriting recognition.

In addition to high sensitivity, the device exhibited a fast dynamic response, environmental resilience, and mechanical durability, all of which are essential yet rarely co-realized in soft-matter sensors. The unit detected pressures as small as 9.74 Pa and exhibited rapid response and recovery times of 12 and 16 ms, respectively [Figure 3H and I]. The sensor response was systematically evaluated under representative temperature conditions (20-45 °C) and bending angles (0°-50°) [Supplementary Figure 12]. The results show minimal signal fluctuation during real-time environmental changes and mechanical deformation, indicating good environmental and structural robustness. After 7 d of water immersion, the signal deviation was maintained below 1.02%, demonstrating the effectiveness of PDMS encapsulation in waterproofing and sweat resistance [Figure 3J and Supplementary Figure 13]. Under 45° cyclic bending at 2 Hz, the response deviation remained within 3.94% [Figure 3K and Supplementary Figure 14], and after applying an overload of 1.3 MPa, the sensitivity was reduced by only 9.84% without significant deterioration in linearity or repeatability [Figure 3L, Supplementary Figures 15 and 16]. Long-term cyclic tests further demonstrated the mechanical endurance of this system, which exhibited a consistent response amplitude and a total deviation of approximately 6.3% after 3,000 loading cycles at 30 kPa [Figure 3M and Supplementary Figure 17]. Importantly, no drift in the baseline or sensitivity was observed, underscoring the excellent fatigue resistance of the MWCNT/mPDMS interface. The plasma-induced surface activation was found to improve the adhesion of the MWCNT network to PDMS, thereby contributing to its stable electrical response under repeated mechanical deformation cycles. Compared with previously reported devices, which often prioritized single metrics (e.g., sensitivity–range tradeoff or short-term response), multifactorial optimization was achieved, including ultrahigh sensitivity, broad detection range, rapid response dynamics, waterproof reliability, and long-term mechanical stability in a unified architecture. Moreover, the coupling between the hierarchical microstructure, conductive percolation, and elastomeric compliance enhanced signal fidelity and ensured stable performance under diverse operating conditions. This synergistic design principle allows for the precise spatiotemporal mapping of subtle muscle deformations, providing a robust foundation for array-level FMG sensing and high-accuracy handwriting decoding.

The sensitivity of the nine channels of the sensor array is presented in Supplementary Figure 18A and B, and an average value of 1.385 kPa^-1 was obtained. The sensitivity of the nine channels relative to the average value [Supplementary Figure 18C] showed a maximum deviation of approximately 5.97%, indicating that the different channels demonstrate high consistency. The response curves of each channel of the flexible sensing array to a single-point external force are presented in Figure 4A. Clear responses to varying applied forces are shown, which enabled the differentiation of the location and magnitude of the applied force [Figure 4B-D]. The crosstalk of the flexible sensing array was calculated [Figure 4E-G], and the maximum crosstalk of the remaining channels was 4.34%, 4.58%, and 2.75% when only channels 1, 5, and 8 were subjected to an external force [Supplementary Note 2]. The crosstalk of each channel remained within a reasonable range under single-point loading, ensuring the measurement accuracy of the flexible sensing array. The flexible sensing array also demonstrated an effective mapping capability for multipoint forces, as illustrated by the response schematics and curves for objects of different shapes [Supplementary Figure 19A-D]. The calculated response heatmaps of the array are shown in Supplementary Figure 19E-G, wherein the flexible sensing array exhibited a reliable mapping performance for multipoint force distributions.

To evaluate the performance of the wearable flexible sensing array, 10 healthy volunteers (five males and five females aged 22 ± 2 years, with no history of neuromuscular disorders) were recruited to participate in gesture- and handwriting-recognition tests [Supplementary Note 3 and Supplementary Figure 20]. The 3 × 3 sensing array was conformally attached to the dorsal forearm using a polyurethane film to prevent displacement [Supplementary Figure 21A]. Each sensing channel was connected in series with a 5 kΩ resistor and sampled at 50 Hz via a wireless microcontroller module [Supplementary Figures 9 and 10]. The deformation of local muscles during finger or wrist motion altered the pressure at the skin–sensor interface, resulting in measurable voltage changes across the array. Furthermore, by combining experimental measurements of the proposed system with reported data from representative wearable interfaces, the system-level performance (including the latency and user calibration time) was evaluated, thereby enabling a comprehensive assessment of real-time applicability [Supplementary Figure 21B and C]. The recorded time-domain signals exhibited periodic fluctuations consistent with the voluntary motion frequency (< 3 Hz), as confirmed by the fast Fourier transform and short-time Fourier transform analyses [Figure 5A and B, Supplementary Figure 22]. After applying low- and high-pass filters at 3 Hz, the low-frequency components reflected voluntary control, while the high-frequency residuals corresponded to intrinsic mechanomyographic signatures. The filtered signals retained distinct amplitude and phase patterns for different gestures [Figure 5A₂ and A₃], demonstrating the ability of the array to resolve muscle deformation dynamics at multiple temporal scales.

Each of the nine channels selectively responds to specific gestures, forming characteristic spatiotemporal response maps [Figure 5C and D]. Although certain channels exhibited partially similar patterns among hand gestures “0”, “1”, and “8”, cross-channel comparison enables unambiguous classification. Repeated measurements showed highly consistent response timing and amplitude over 10 consecutive cycles [Figure 5E], confirming excellent repeatability. The corresponding power spectral density (PSD) maps further distinguish different gestures based on their spectral energy distribution [Figure 5F and G, Supplementary Figure 23]. Dimensionality reduction through principal component analysis and t-distributed stochastic neighbor embedding revealed compact clustering within identical gesture classes and clear separation between distinct gestures [Figure 5H].

Statistical evaluation of the root-mean-square (RMS) values and signal standard deviations (SDs) [Figure 5I and J, Supplementary Figures 24 and 25] showed that the motion states exhibited larger variance than the stationary states, confirming robust channel activation and effective noise suppression. Collectively, these analyses demonstrate that the array captured muscular deformation with a high temporal resolution and specificity, thereby forming a stable biomechanical signature for each gesture.

To quantify recognition performance, a deep neural network (DNN) comprising three hidden layers (50–30–20 neurons) was trained on the acquired dataset, with an 80%/20% training/testing split [Figure 6A and Supplementary Note 4]. The recognition accuracy increased with the number of active channels, reaching 84.3%, 90.4%, 95.2%, 97.4%, and 99.6% for three-, four-, five-, six-, and nine-channel configurations, respectively [Supplementary Figure 26 and Figure 6B]. Even when reduced to a four- or five-channel array, accuracies above 90% and 95% were obtained, respectively, suggesting that high recognition reliability can be achieved with fewer sensing sites. Channel arrangement had a negligible influence on recall and precision [Supplementary Figure 27]. By contrast, signal preprocessing affected classification outcomes, as high-pass-filtered data yielded poor prediction accuracy due to low signal amplitude and environmental noise, whereas low-pass-filtered data achieved comparable or slightly superior accuracy compared with raw data [Figure 6C-F and Supplementary Figure 28]. The effect of dataset size was further investigated by varying the number of motion repetitions. The recognition accuracy significantly improved from one to two repetitions but plateaued thereafter [Figure 6G, Supplementary Figures 29 and 30], indicating that high performance can be maintained with reduced data volume and computational cost.

The gesture recognition robustness was examined across different movement speeds [Figure 6H]. Extremely fast gestures resulted in reduced accuracy, while the inclusion of multispeed data during training restored and further enhanced the overall performance [Figure 6I and J, Supplementary Figure 31]. These results demonstrate the adaptability of the DNN model to intrasubject speed variations and its potential for generalized motion decoding.

Overall, the wearable array exhibited precise, repeatable, and high-fidelity detection of muscular deformation during voluntary motion. By coupling dense spatiotemporal sensing with machine-learning-based decoding, near-perfect gesture recognition (≈ 99.5%) was achieved without recalibration, establishing a scalable strategy for noninvasive human–machine interaction and neuromuscular monitoring.

As significant differences in hand-force generation are encountered when writing different symbols and words, the proposed flexible sensing array may be applied to the recognition and correction of handwriting. Figure 7A shows a schematic of the writing process, while Figure 7B presents the corresponding symbols and their stroke orders, such as counterclockwise and orthogonal circular motions. Letters, numbers, and Chinese characters with different stroke orders and continuous strokes were used to evaluate the performance of the proposed system [Figure 7C-E and Supplementary Figure 32]. Figure 7F shows the response curves of each channel when writing different shapes, demonstrating clear distinctions among shapes. The action time required for writing different shapes is presented in Figure 7G, which indicates that both shape and stroke sequence produced distinguishable temporal features for symbol identification. The RMS results obtained for channel 6 when different writing shapes are shown in Figure 7H, wherein clustered distributions for identical shapes and a clear separation between different shapes are observed. These results indicate that the proposed flexible sensing array effectively distinguishes different writing symbols and different writing sequences of the same symbol. The response curves generated during the writing of different characters are presented in Figures 7I and J, which demonstrate distinct multichannel signal patterns corresponding to different handwriting tasks.

As pen diameter and weight vary in practical applications, the effects of these parameters were investigated considering the writing-recognition accuracy. The flexible sensing array exhibited good response as the mass of the gripped object was varied [Supplementary Figure 33]. Here, a single pen was weighted using a tennis-racket weight to simulate different pen masses. The recognition performance obtained under different pen weights was further evaluated [Supplementary Figure 34 and Figure 7K]. The classification accuracies for the normal pen, slightly heavier pen, and heavier pen were 98.25% ± 1.47%, 97.93% ± 1.76%, and 98.37% ± 1.70%, respectively. Statistical analysis using a paired t-test showed no significant difference between these conditions (P > 0.05), indicating that the proposed FMG-based recognition system was robust toward variations in pen weight. By contrast, the responses of each channel when holding objects with different diameters significantly differed [Supplementary Figures 35 and 36]. The confusion matrix for symbol classification using pens of different thicknesses, achieved by wrapping the pen with tennis grip tape, is shown in Supplementary Figure 37, while the recall and classification accuracy for each pattern are presented in Figure 7L. Although the classification accuracy slightly decreased (from 98.13% to 97.25%) as the pen became thicker, the overall prediction accuracy remained high.

The action times of the written words are presented in Supplementary Figure 38, and the RMS of channel 6 is shown in Supplementary Figure 39. The action times and RMS significantly varied across different words, while variations within identical actions remain small. The classification confusion matrix for writing different words is shown in Supplementary Figure 40, and the recall and accuracy are provided in Supplementary Figure 41, demonstrating a high recognition accuracy of 96.35% ± 0.53%. Effective differentiation was also achieved for different downstroke orders for the same word. Moreover, variances across shapes and stroke sequences enabled the effective discrimination between different writing symbols. The RMS value of the signal detected by channel 6 differed significantly across writing shapes, exhibiting clustered distributions for identical shapes and clear separation between different shapes. These results indicate that the proposed flexible sensing array effectively distinguishes different writing symbols and different writing sequences of the same symbol. To assess scalability, a 26-letter handwriting dataset was collected under natural writing conditions. The model achieved an accuracy of ~98.8% with errors mainly occurring among structurally similar letters, indicating strong discrimination under an expanded vocabulary [Supplementary Figure 42].

We further explored the impact of cursive writing on recognition accuracy. As shown in Supplementary Figures 43 and 44, the overall recognition accuracy reached 97.83% ± 0.22% and was unaffected by cursive writing. The confusion matrix and recall rate of different character classifications for datasets composed of mixed stroke orders and cursive variations of standard writing (for example, B includes B1, B2, and BL) are shown in Figure 7M and Supplementary Figure 45. An accuracy of 99.54% ± 0.16% was obtained when using mixed data for the classification and prediction of handwritten characters, which is significantly higher than that obtained using a single dataset (P < 0.05). Therefore, when constructing a dataset of written characters, the inclusion of different stroke orders and cursive writing methods is necessary to improve prediction accuracy. To evaluate the performance of different machine-learning models in the handwriting recognition task, four models [DNN, gated recurrent unit (GRU), long short-term memory (LSTM), and Transformer] were compared [Supplementary Table 2 and Supplementary Figure 46]. Among them, GRU, LSTM, and Transformer represent temporal modeling approaches capable of capturing time-dependent features in FMG signals, whereas DNN relies on multichannel instantaneous features for classification. All models achieved accuracies of > 99%, with Transformer achieving the highest accuracy and DNN achieving a comparable accuracy. Notably, DNN exhibited the smallest number of parameters, lowest model size, and fastest inference speed. Considering both the recognition performance and computational efficiency, the DNN model was selected for further investigation considering real-time and lightweight wearable applications.

Ablation studies were performed on network depth and regularization to validate the selected DNN model [Supplementary Tables 3 and 4]. Increasing the network depth improved feature representation, and an optimal performance was achieved using a three-layer architecture. Batch normalization and dropout both enhanced stability and generalization, and their removal resulted in performance degradation. To further assess model effectiveness, DNN, one-dimensional convolutional neural network (1D-CNN), LSTM, and an ensemble model were compared under the same protocol [Supplementary Figure 47]. DNN achieved the best overall performance, offering the highest accuracy despite having a slightly higher relative RMS error (RRMSE), and it was therefore selected for further handwriting recognition evaluation. This system also showed strong robustness in cross-subject evaluations. A leave-one-subject-out validation showed that the recognition accuracy remained above 97%. In addition, signal drift induced by repeated reattachment (10 cycles) remained below 1.6%, confirming the reliability of the system for practical wearable applications [Supplementary Note 5 and Supplementary Figure 48]. To evaluate robustness under dynamic motion, tests were conducted using different walking speeds. The model maintained a high accuracy under both slow (98.74% ± 0.37%) and fast (98.63% ± 0.42%) walking, with no significant difference compared with the baseline (99.54% ± 0.16%) [Supplementary Figure 49].

In addition, cross-age validation was performed using data from two participants that were over 40 years old. Without retraining on this group, the model achieved accuracies of 98.63% ± 0.42% and 98.71% ± 0.38%, with no significant difference compared with the baseline (P = 0.067 and 0.074; Supplementary Figure 50), demonstrating a strong generalization capability. To further assess robustness under a cognitive load, a dual-task experiment was conducted in which participants performed continuous subtraction during handwriting. The system maintained a high accuracy of 97.65% ± 0.34% (P = 0.068; Supplementary Figure 51), indicating a stable decoding performance under attention fluctuation. Cross-continuity generalization experiments were conducted to evaluate the robustness of the model under different handwriting styles [Supplementary Figure 52]. When the training and testing styles were inconsistent, the recognition accuracy slightly decreased to 96.5% ± 0.21% (Normal → Legato) and 96.0% ± 0.17% (Legato → Normal). After introducing mixed handwriting samples for training, the accuracy significantly improved to 98.5% ± 0.13% (Mixed → Legato) and 99.09% ± 0.09% (Mixed → Normal), demonstrating enhanced cross-style generalization ability. To evaluate temporal robustness, time-sequential cross-validation was performed by training the model using early writing samples and testing it on later samples within the same session [Supplementary Figure 53]. The recognition accuracy was 99.34% ± 0.15%, with no significant difference compared with the Early → Early condition (99.54% ± 0.16%, P = 0.071). The analyzed system also maintained robust handwriting recognition across different writing modalities and scales. The recognition accuracy remained above 97% even for cursive and high-speed writing, with a < 1.5% deviation from discrete characters. Moreover, it exhibited minimal dependence on handwriting size (5-25 mm), confirming the geometric and temporal adaptability of the proposed FMG-based sensing array [Supplementary Note 6 and Supplementary Figure 54].

The findings of representative studies are listed in Supplementary Tables 5 and 6 to compare the proposed system with recently developed flexible-sensor gesture-recognition systems. The competitive advantages of our system are highlighted in terms of its accuracy and wearable-device integration. The proposed array uniquely integrates both gesture and stroke-order recognition tasks, demonstrating array-level reliability and handwriting capability beyond existing FMG or sEMG systems.