Device fabrication
We first introduced the design concept behind the thermoelectric spacer fabric (TESF). Figure 1 illustrates the production process and device fabrication schematic of the TESF. Low-cost, high-rigidity 36 dtex polyester monofilament fibers were used as spacer yarns, which were combined with softer 135 dtex and 90 dtex polyester monofilaments through 3D knitting technology for continuous mass production on our machine (Fig. 1a). Warp knitting machines operate without generating direct emissions, carbon emissions, pollutants, or wastewater discharges, thus having a minimal impact on the environment. The loop formation process of the specific double-needle bar warp knitting machine is shown in Supplementary Fig. 1, and detailed knitting procedures are supplemented in the “experimental details” section in the Supplementary Information. The digital diagram of the inlay yarn and its movement are shown in Supplementary Fig. 2. Due to its ease of water processing, non-toxicity, and excellent thermoelectric properties, carbon nanotubes (CNTs) were selected as the thermoelectric fillers for the fabric (Fig. 1b). Oleylamine was used as an n-type dopant for CNTs due to its excellent air stability40, while poly(3,4-ethylenedioxythiophene):poly(styrene sulfonate) (PEDOT:PSS) and CNT have been reported to exhibit excellent p-type thermoelectric properties41,42. Thanks to the cuttability and flexibility of TESF, it can be easily shaped into various forms and labels without leaving marks, making it suitable for integration into everyday wearable devices (Supplementary Fig. 3). As shown in Supplementary Fig. 4, photographs of diverse types of fabrics display that both p-type and n-type thermoelectric fabrics can be woven on a large scale. To understand the doping mechanism of polymers on thermoelectric fabric, we examined the microstructure of the fabrics. The scanning electron microscope (SEM) images of the initial spacer fabric reveal the smooth surface of individual fibers. The interlayer structure between the fabrics is clearly defined, and the gaps between the fibers are distinct (Supplementary Fig. 5). After the introduction of PEDOT:PSS, the individual fibers in the p-type TESF become rough (Supplementary Fig. 6). The characteristic composite layered structure of PEDOT:PSS and CNTs can be observed in the magnified SEM images. In contrast, the surface of the n-type TESF remains smoother, but the oleylamine-doped CNTs wrap, increasing the diameter of the CNT edges (Supplementary Fig. 7). The basic morphology of the fabric shows no significant changes. Although the initial spacer fabrics can be mass-produced using industrialized machines, the preparation of TESF remains challenging. In future work, we plan to integrate an impregnation bath directly at the back end of the knitting process, enabling large-scale and continuous preparation of TESF.
Our TESF is compared to recently reported wearable multifunctional thermoelectric devices in terms of breathability, response speed, accuracy, κ, and durability, and it meets several leading metrics in the field (Fig. 1c)43,44,45,46,47. Notably, the permeability of TESF, a crucial factor for wearable devices, reaches 1250 mm s−1. This fully supports heat and moisture exchange during wear, ensuring a comfortable user experience. The ultra-low κ of 0.043 W m−1 K−1 allows TESF to consistently maintain a stable temperature gradient across the fabric, facilitating accurate temperature signal detection. When compared to the latest thermoelectric-based temperature sensors, TESF-based sensors exhibit a rapid temperature response of up to 240 ms and precise temperature difference recognition of 0.02 K, both of which represent the high performance levels in available thermoelectric devices (Fig. 1d, see Supplementary Table 1 for detailed comparisons)21,43,45,47,48,49,50. Additionally, when evaluated against a recently reported 3D thermoelectric compression sensor, our device outperforms in all metrics (Fig. 1e)44,51,52,53,54,55,56. The TESF endures up to 10,000 compression cycles, the maximum reported for thermoelectric devices, and achieves a compression response time of 20 ms, setting a record for response speed in thermoelectric devices (see Supplementary Table 2 for detailed comparisons).
Sensing performance
The 3D structure of the spacer fabric endows the thermoelectric fabric with excellent breathability, which is crucial for heat and moisture exchange during wear (Fig. 2a). The samples were placed on a custom-built electrical testing device to investigate their thermoelectric properties. As shown in Fig. 2b and Supplementary Fig. 8, the original TESF, which contained only CNTs, exhibited a positive Seebeck coefficient (S) of approximately 48.83 μV K−1 due to the presence of oxygen impurities, indicating that the charge carriers were holes. After adding PEDOT:PSS as a p-type dopant, the electrical conductivity (σ) and power factor (S2σ) increased significantly, consistent with previous studies on CNT/PEDOT:PSS composites42,57,58,59. When oleylamine was introduced as an n-type dopant to the CNTs, the S of the fabric switched to a negative value. Additionally, the unique structure of the 3D fabric resulted in a large amount of static air within, allowing TESF to achieve an ultra-low κ superior to other thermoelectric materials with different structural designs. The out-of-plane ΔT makes it easier to maintain a stable output signal when used as a wearable device. The sensitivity of TESF to different temperature differences was first evaluated, and as seen in Supplementary Fig. 9, both p-type and n-type TESF responded sensitively to the respective ΔTs. They were also sensitive to small ΔTs, with the minimum temperature response for p-type/n-type TESF being approximately 0.02 K (Fig. 2c). The response time to temperature changes was just 240 ms (Fig. 2d). Moreover, the output voltages of both n-type and p-type TESF demonstrate remarkable stability under cyclic temperature variations (Supplementary Fig. 10).
a Air permeability of p-type thermoelectric spacer fabric (TESF) and n-type TESF. All samples were measured three times and averaged. b S and thermal conductivity (κ) of carbon nanotubes (CNTs)-based, p-type, and n-type TESF. All samples were measured three times and averaged. c Minimum discernible ΔT for p-type and n-type TESF. d Temperature sensing response time for p-type and n-type TESF. e Resistance response of TESF under different compressive strains, where R is the resistance change value under strain, and R0 is the initial resistance. f Response and recovery speed of TESF under compressive strains. g Stability of resistance changes in TESF after 10,000 compressive strain cycles. h S of p-type and n-type TESF under different compressive strains.
Piezoresistivity is an important parameter for evaluating the practical application potential of wearable electronics. The 3D supporting structure of the spacer fabric also lends potential for its use as a piezoresistive sensor. The strong polyester monofilaments in the fabric bend under external pressure and return to their original state once the force is removed (Supplementary Fig. 11a). Supplementary Fig. 11b shows the compression strain-stress curves for different samples. Compared to the initial spacer fabric, the stress in both p-type and n-type TESF significantly increased, due to the rigid effects of the CNTs in the dopants. We selected p-type TESF as a reference to closely examine its mechanical and electrical changes during compression. As shown in Supplementary Fig. 11c, the fabric exhibits no significant hysteresis or fatigue under 10%−50% compression strain, indicating high resistance and recovery (Supplementary Fig. 12).
During dynamic compression and recovery, the spacer yarns in the fabric make contact and separate from the internal surface of the fabric, resulting in a piezoresistive effect. We constructed a 3D compression model, as shown in Supplementary Fig. 13, to reveal the resistance changes during this dynamic process. During compression, the contact angle between the spacer yarns and the two planes decreases, increasing the contact area and creating more conductive pathways. Based on this concept, we evaluated the resistance of the fabrics under different compression deformations to demonstrate their piezoresistive sensing potential (Fig. 2e). The gauge factor (GF) during compression can be categorized into three regimes: A, B, and C. When the compression deformation exceeds 20%, the gauge factor reaches its maximum, approximately 1.7, which corresponds to our previously described equivalent resistance model. Additionally, the response time and recovery time of the fabrics under large deformations are only 20 ms and 60 ms, respectively, fully meeting the real-time requirements for wearable devices (Fig. 2f). Mechanical stability and durability are critical indicators for wearable devices. As shown in Fig. 2g, the piezoresistive signal of the fabric remains stable after 10,000 cycles under a 40% compression strain. Additionally, the resistance and S values of both n-type and p-type TESFs remain unchanged after 10,000 compressions, demonstrating that the 3D fabric structure ensures reliable and stable support (Supplementary Fig. 14). Moreover, the thermoelectric properties of both n-type TESF and p-type TESF remain very stable after bending cycles and washing (Supplementary Figs. 15, 16). To investigate the coupling relationship between internal resistance and output voltage during dynamic processes, we evaluated the S under different compression strains (Fig. 2h). Remarkably, the S of TESF remained unaffected as strain increased. This prompted further investigation into the changes in thermoelectric voltage and thermal resistance under dynamic conditions.
Decoupling mechanism
To explore the decoupling characteristics of TESF, we designed a series of experiments. Taking p-type TESF as an example, we applied independent and coupled stimuli to the fabric and observed changes in the electrical signals. Initially, we measured the I–V curve of the fabric in its baseline state (Fig. 3a). When pressure was applied to the fabric, the resistance signal significantly decreased due to the piezoresistive effect, resulting in a noticeable shift in the I–V curve (Fig. 3b). Next, while keeping the pressure constant, we applied a ΔT to the fabric, which caused the I–V curve to shift to the right with the same slope (Fig. 3c). According to the typical thermoelectric mechanism, the output thermoelectric voltage is S × ΔT. This result indicates that the internal resistance of the fabric is not affected by the inherent temperature variation during the output thermoelectric voltage. Finally, when the compressive stress was removed, the I–V curve returned to a slope parallel to the initial sample, with only a lateral shift in the thermoelectric voltage due to the ΔT (Fig. 3d).
I–V curves of TESF (a) without external stimuli, (b) under compressive strain alone, (c) under both compressive strain and ΔT, (d) under ΔT only, (e) at different ΔT values (ΔT = 0, 1, and 3 K), (f) under various compressive strains (0%, 30%, 60%), and (g) under different strains (0%, 30%, 60%) with a constant ΔT of 1 K. h Finite element simulation of temperature distribution and output thermal voltage of TESF under compressive strain.
To refine the experimental process, we conducted continuous testing across multiple gradients. As shown in Fig. 3e, the I–V curves shift significantly to the right when the fabric is subjected to different ΔTs, indicating clear thermoelectric properties. The I-V curves of the fabrics under different compressive strains at a constant temperature difference exhibit varying slopes (Fig. 3f). The slope decreases as the pressure increases, indicating a corresponding reduction in resistance. Notably, the voltage remains unaffected under different pressure conditions. Additionally, we performed compression tests on the fabric under a constant ΔT of 1 K (Fig. 3g). Obviously, applied pressure does not affect the thermoelectric voltage of the fabric; instead, its lateral shift remains consistent with the origin. From these observations, we conclude that the piezoresistive and Seebeck effects of the fabric are fully decoupled. We then connected the fabric to a temperature-variable compression testing facility, as illustrated in Supplementary Fig. 17a, to systematically investigate its decoupling properties under various ΔTs. The ΔT was controlled in real time using a pair of Peltier elements. Initially, we evaluated the internal resistance of the fabric under different compressive strains across a ΔT range of 0 to 30 K (Supplementary Fig. 17b). The results show that the temperature has a minimal effect on resistance, irrespective of the applied compressive strain. Similarly, monitoring the output thermoelectric voltage of the fabric under different compressive strains revealed that pressure has a negligible influence on the thermoelectric voltage (Supplementary Fig. 17c). These findings confirm that in this 3D material, piezoresistive signals and thermoelectric voltage signals change independently of each other. Furthermore, we constructed a 3D equivalent model of the fabric and validated our strategy using finite element analysis. As shown in Fig. 3h, a stable ΔT of 10 K was first applied to the fabric. With the application of compressive strain, the temperature distribution within the fabric remains unaffected, as observed from the contour plot, due to the low κ of the fabric. Since the S remains constant at different compressive strains, the thermoelectric voltage output remains stable. This effectively demonstrates the decoupling characteristics of Seebeck and piezoresistive effects in the 3D fabric.
Irregular deformations caused by human movement significantly impact the output power of the device, and no dedicated work has yet focused on decoupling the signals of this device to enhance its stability. Based on the decoupling characteristics of TESF, we investigated the decoupling performance of integrated devices. The experiments were conducted under conditions with and without thermal sources and pressure, simulating various stimuli coupling caused by temperature transfer and motion during human wear. The device was fabricated by seamlessly integrating eight pairs of p-type and n-type TESF into a 3D fabric substrate (Supplementary Fig. 18). The circuit connections of the device are shown in Supplementary Fig. 19. The entire device measures approximately 5.5 × 5.5 cm, allowing it to be flexibly deformed and making it suitable for everyday wear. As illustrated in Supplementary Fig. 20a, we applied three different conditions to the device: plastic board pressure, palm close, and palm press, using a plastic board and a palm without a thermal source. When pressing the device with a plastic board, the device resistance generates a compressive signal, while the voltage signal remains unchanged (Supplementary Fig. 20b and Supplementary Movie 1). When the palm approaches the device, the resistance signal remains constant. Due to the ΔT caused by the human body and the environment, the device produces a noticeable thermoelectric voltage response (Supplementary Fig. 20c and Supplementary Movie 2). As shown in Supplementary Fig. 20d, when repeatedly pressing the device with the palm, both resistance and voltage signals are output in a decoupled manner due to the piezoresistive and Seebeck effects (Supplementary Movie 3). This indicates that not only TESF, but also the flexible devices fabricated with integrated TESF, exhibit decoupling of resistance signals caused by strain and thermoelectric voltage signals induced by ΔTs.
Sensing applications
We first evaluated the output performance of the device. As shown in Supplementary Fig. 21a, the maximum output power of the device is achieved at different ΔTs when the connected load resistance (Rload) is 110 Ω, which matches the internal resistance of the device. This is further validated by the maximum output power at various load resistances, as presented in Supplementary Fig. 21b. For energy harvesting tests, the device was worn on the user’s wrist (Supplementary Fig. 21c). An infrared diagram indicates a ΔT of approximately 8 K between the top and bottom of the device, attributed to the low κ of the 3D fabric, which effectively prevents the transfer of body heat to the top surface. Consequently, the device generates a stable and consistent voltage of 5.44 mV (Supplementary Fig. 21d).
The flexibility and high integration of this device make it easy to incorporate into everyday wearable garments. As shown in Fig. 4a, the device was sewn into an N95 mask, creating a wearable smart mask for real-time breath monitoring. According to infrared images, temperature changes at the nose area during inhalation and exhalation cause variations in the ΔT across the device, enabling real-time breath monitoring. When users wear the smart mask, the real-time output thermal voltage distinguishes between facial actions, including normal breathing, coughing, sighing, and deep breathing (Fig. 4b). Importantly, the thermal voltage signal generated by breathing is unaffected by resistance changes caused by airflow vibrations. The accurate recognition of the device allows differentiation and identification of these signals based on the frequency of inhalation and exhalation, the rate of change in the signal, and the peak values of the thermal voltage. For instance, normal breathing shows a slow frequency and low peak values (Fig. 4c). Coughing produces three distinct thermal voltage peaks with a long duration (Fig. 4d). Compared to other actions, signing results in much higher thermal voltage peaks (Fig. 4e). The difference between normal and deep breathing is reflected in the height and frequency of the peaks (Fig. 4f). In addition to assessing breathing conditions, the smart mask can differentiate facial expressions such as smiling, slight smiling, and laughing (Supplementary Fig. 22).
a Schematic diagram of the smart mask integrated into an N95 mask for monitoring respiratory status, including infrared images showing temperature changes during expiration and inhalation. b Real-time thermal voltage output of the smart mask during various respiratory actions, including expiration and inhalation, and magnified views of (c) thermal voltage signals during breathing, (d) thermal voltage signals during coughing, (e) thermal voltage signals during sighing, and (f) thermal voltage signals during deep breathing.
To explore practical applications of the device, we integrated it with a wireless transmission module to build a wearable wireless signal monitoring system. The working principle of the wearable wireless sensing system is shown in Supplementary Fig. 23. Based on the thermoelectric effect, the device generates real-time voltage signals, which are then captured by an analog-to-digital converter and sent to an operational amplifier. The microcontroller collects and analyzes these signals, extracting the waveforms, frequencies, and peak values of the thermoelectric voltage, and transmits the data in real-time to a mobile terminal via Bluetooth. We installed this system into our mask (Supplementary Fig. 24). As shown in Supplementary Fig. 25 and Supplementary Movie 4, when the user breathes normally, the signal sent to the mobile app fluctuates regularly. However, when the user stops breathing, the signal disappears, indicating that the user is in danger and requires immediate assistance. Similarly, we sewed the fabric onto insulating gloves and integrated it with a wireless transmission module to explore temperature warning applications (Supplementary Figs. 26, 27). When the user wears gloves at varying distances from a heat source, the phone collects corresponding signals (Supplementary Fig. 28). As shown in Supplementary Fig. 29 when the gloves are near a flame, a high-temperature threshold is detected; exceeding this threshold poses a burn risk to the user. Therefore, real-time data observed on the phone can provide pre-warning to avoid burns and scalds.
The resistive sensor signals of wearable devices often suffer from temperature-induced interference due to direct contact with the skin, which can reduce the accuracy and scientific validity of the sensing signals. Our TESF, however, can shield against temperature interference in resistance measurements. Therefore, we designed a multifunctional sensor based on TESF (Supplementary Fig. 30). We installed the sensor on the wrist, elbow, and finger joints, and the sensor accurately differentiates various motion amplitudes through the piezoresistive effect (Fig. 5a–c). To monitor the independence of resistance and thermoelectric voltage signals, we also measured the inherent thermoelectric voltage during the sensing tests and found that the thermoelectric voltage is unaffected by motion, which aligns with our conclusions (Supplementary Fig. 31). People with speech impairments require coordinated finger gestures and high interaction for learning sign language. We designed a smart glove capable of converting finger movements into electrical signals, aimed at enhancing sign language learning for individuals with speech impairments (Fig. 5d). We attached a compressive strain sensor to the joints of the glove to capture the pressure generated during finger movements and transmitted the captured signals to a personal computer for performance analysis. Such a smart glove not only addresses the initial challenges of sign language learning for individuals with speech impairments but also provides a method for remote communication for people with disabilities. The output signals of the glove when performing sign language numbers 1-5 are shown in Supplementary Fig. 32.
The pressure sensor-based TESF is used to signal bending activity in the (a) wrist, (b) elbow, and (c) finger areas. d Schematic of smart glove-assisted sign language training and transformation. e Schematic diagram of the corresponding transformation of sign language and sensing signals, where the test signals are all ΔR/R0. f Confusion matrix of gesture recognition with an accuracy of 100%.
Additionally, Fig. 5e and Supplementary Fig. 33 display complex sign language gestures, including I LOVE YOU, OK and COOL glove signals. By distinguishing the arrangement and combination of finger movements, specific meanings can be assigned. The collected signals of the smart glove can be interpreted to understand sign language by analyzing these finger movement patterns. To achieve intelligent recognition of sign language gestures, we designed and implemented a machine learning model based on SVM for precise classification and recognition of complex device signals and developed a machine learning-assisted sign language training strategy (Supplementary Fig. 34). In the implementation process, we first extracted key attributes of the signals through feature engineering and then used a radial basis function (RBF) kernel to enhance the non-linear classification capabilities of SVM. Additionally, we optimized the hyperparameters of the model using grid search and cross-validation strategies, which significantly improved the classification accuracy. The model can achieve a 100% classification accuracy across multiple independent test datasets (Fig. 5f and Supplementary Fig. 35).