Robust hydrogel sensors for unsupervised learning enabled sign-to-verbal translation

Abstract

Highly stretchable and robust strain sensors are rapidly emerging as promising candidates for a diverse of wearable electronics. The main challenge for the practical application of wearable electronics is the energy consumption and device aging. Energy consumption mainly depends on the conductivity of the sensor, and it is a key factor in determining device aging. Here, we design a liquid metal (LM)-embedded hydrogel as a sensing material to overcome the barrier of energy consumption and device aging of wearable electronics. The sensing material simultaneously exhibits high conductivity (up to 22 S m⁻¹), low elastic modulus (23 kPa), and ultrahigh stretchability (1500%) with excellent robustness (consistent performance against 12 000 mechanical cycling). A motion monitoring system is composed of intrinsically soft LM-embedded hydrogel as sensing material, a microcontroller, signal-processing circuits, Bluetooth transceiver, and self-organizing map developed software for the visualization of multi-dimensional data. This system integrating multiple functions including signal conditioning, processing, and wireless transmission achieves monitor hand gesture as well as sign-to-verbal translation. This approach provides an ideal strategy for deaf-mute communicating with normal people and broadens the application of wearable electronics.

INTRODUCTION

Wearable sensors that transducing a physical stimulus into electrical signals can track human status and surrounding information by translating these physical stimuli into electrical signals,^1–5 which play a vital role in the development of the Internet of Things,^6–8 soft robotics,^9–11 bioelectronics,^12–14 and artificial intelligence.¹⁵ For high detection performance of wearable sensors, many efforts have been devoted to improving the sensitivity by tunneling effect,¹⁶ crack generation mechanism,^17,18 textile technology,^19–21 and so on. Even many researchers work on solving these barriers, the energy consumption and device aging of wearable sensors are still in infancy.

Energy consumption depends on the conductivity of wearable sensors and is also one of the most important factors for device aging. The high resistance of wearable sensors often significantly increases the device temperature, causing them to burn out easily.^22,23 Despite the availability of many high-conductivity inorganic material-based sensors^24–26 which can address this issue, they still face challenges due to their poor stretchability and high Young's modulus. To overcome the poor mechanical properties, various hydrogels with conductive fillers (e.g., carbon nanotube,²⁷ graphene,²⁸ MXene,^29,30 and metallic particles³¹) have been developed. These rigid and conductive fillers slip from the surrounding hydrogel matrix during the stretching process, leading to increased toughness, limited deformation,^32,33 and poor stability.^34–37 The current rigid fillers cannot simultaneously increase the conductivity and robustness of the hydrogels.³⁸ Liquid metal (LM) is the most commonly used conductive filler and is rapidly gaining popularity due to its distinctive attributes such as high conductivity,³⁹ fluidity at room temperature, environmental stability, and biocompatibility.^40–42 Because the hydrogels and LM are compatible and complementary, integrating LM with hydrogels may potentially overcome this limitation.⁴³

Hereby, we report a soft and highly stretchable LM-embedded hydrogel (LM-H) with long-term stability by encapsulating LM particles into polymeric network of hydrogels. The LM-H exhibits low elastic modulus (23 kPa), high stretchability (1500%), and high electrical conductivity (22 S m⁻¹). Benefiting from their excellent electrical and mechanical properties, the LM-H readily serves as a strain sensor (GF = 3.86) to real-time record physiological strains with excellent robustness (consistent performance after 12 000 cycles). To take full advantage of their excellent properties, the LM-H strain sensor can apply to gesture monitoring, electronic screens, and marine animal behavior monitoring. Furthermore, a sign language translation system with self-organizing map (SOM) is fabricated by connecting the LM-H with a wireless circuit, which realizes the recognition of different gesture signals. This system can rapidly identify simple and efficient gesture recognition with a response time of 0.21 s.

RESULTS AND DISCUSSION Design and preparation of the LM-H

The LM-H with significant improvements in conductivity and stretchability compared with existing conductive hydrogels^44,45 is developed by encapsulating LM particles into sodium alginate (SA) and acrylamide (AAm) polymeric network (Figure 1A,B). The LM particles are dispersed in the AAm monomer solution by ultrasonication to form a uniform and dark gray AAm/LM suspension (Figure S1), which attributes to the surface interaction between the NH₂ groups of AAm and the metallic oxide layer of the LM. The morphology of AAm modified EGaIn exhibits spherical nanodroplets of approximately 50 nm diameter by scanning electron microscopy (SEM) (Figure S2) and transmission electron microscopy (TEM) (Figure S3). Figure S4 shows the diameter of 60 nm tested by dynamic light scattering (DLS). When the EGaIn concentration is higher than 1.4 wt.%, the suspension of the EGaIn is unstable (Figure S5) due to the coordination between Ga of excess EGaIn and NH₂ groups of AAm. The corresponding elemental mapping images can identify the distribution of Ga and In elements in the nanodroplets (Figure S6). The hydrogel forms a double network which can improve the electrical property via continuous connected pathways without compromising the excellent stretchability and mechanical robustness (Figure 1C). Initially, the SEM image of the LM-H shows relatively smooth and porous structure with a pore size of 5–10 μm (Figure 1D). Such structures often prevent from crack propagation of the LM-H⁴⁶ and further contribute to improve the robustness of LM-H. When the strain increases, the conductive pathway formed by EGaIn particles gradually disconnect, resulting in an increase in resistance of the LM-H (Figure 1E). The corresponding x-ray photoelectron spectroscopy (XPS) identifies the distribution of Ga, In, Ca, and Na elements in the hybrid hydrogel (Figure S7), which further confirms that EGaIn particles are homogeneously dispersed in the network of the PAAm-SA hydrogel (Figure 1F). Figure S8 shows peaks of C 1s, O 1s, and N 1s, which attribute to various reactants (sodium alginate, acrylamide, and N, N′-methylenebisacrylamide) and the peaks of Na 1s and Ca 2p are attributed to the Na⁺ in sodium alginate and Ca²⁺. The Ga 2p at 1118 eV and the In 3d at 443 eV indicate the presence of EGaIn. The elemental mapping (Ga) of the LM-H at a strain of 0%, 40%, and 100% supports the fracture mechanism. In the original state (Figure 1F), 40% stretching condition (Figure 1G), and 100% stretching condition (Figure 1H), the network of the LM-H gradually become long and small.

FIGURE 1. Design of the LM-H. (A) Schematic diagrams of liquid metal. (B) Structure of PAAm and SA. (C) Polymeric network inside the LM-H. (D) SEM images of the LM-H with double network (Scale bar, 10 μm). (E) Illustration of fracture mechanism. (F–H) Elemental mappings of the LM-H at a strain of 0%, 40%, and 100% (Scale bar, 1 μm). The distance between Ga elements (yellow) gradually increases with the stretch, indicating the broken conductive path of the LM-H.

Mechanical properties of the LM-H

After the cross-linked dual network of the LM-H formed, these LM-H can be stretched over 13 times (Figure 2A), and the stretching process of the LM-H is shown in Figure S9. The mechanical strength of the LM-H significantly increases due to the coordination effect between the LM and PAAm.⁴⁷ However, as a soft filler, the EGaIn cause relatively higher elastic modulus and lower stretchability. The elastic modulus of the LM-H increases from 16 to 23 kPa (Figure 2B) and the stretchability decreases from 1500% to 1350% (Figure 2C) as the EGaIn content increases from 0 to 1.4 wt.%. This phenomenon is widely encountered in the field of nano-reinforced soft materials⁴⁸ and is typically caused by mechanical stress build-up and the associated devastating energy release. The water content of the LM-H with different EGaIn concentration is measured by lyophilization and the results are over 80% (Figure S10), which results in the softness of the LM-H. The swelling ratio (W_s) reveals the structural properties and cross-linking density of the network and is calculated by Equation (1).^49,50 [Image Omitted. See PDF]where W_t and W₀ are the weights of the swollen hydrogel at a specific time and the weights of dry hydrogel, respectively. The W_s displays a maximum number at 400 min, followed by leveling off (Figure S11). Furthermore, when the EGaln content increases from 0.2 to 1.4 wt.%, the W_s decreases from 11 to 8. The different W_s numbers for the LM-H are caused by different EGaIn concentrations. Consequently, it induces a change in the hydrogel from a glassy state to a rubbery state as fluid water penetrates the hybrid hydrogel network. Such state transition is also proved by the rheological characteristics of the LM-H which are determined by frequency sweeps at constant stress by comparing the magnitude of loss modulus (G″), storage modulus (G′), and the loss factor (tan θ = G′/G″). As shown in Figure 2D, the angular frequency-dependent progress for G′ and G″ displays a rapid increase up to a maximum (0.1–40 rad s^–1) followed by leveling off (40–100 rad s^–1). G″ is relatively higher than G′ in the whole frequency range, therefore, tan θ ˂1, indicating that the LM-H can keep elasticity (the typical gel phase character) in all range of shear rate from 0.1 to 100 rad s^–1 both with and without LM. The value of tan θ for the LM-H is apparently lower than that of the PAAm-SA hydrogel (Figure 2E), demonstrating that the LM-H behaves as a more viscoelastic solid. Hence, it can be concluded that the LM-H maintains its elasticity under high shear rates, which is a testament to its mechanical robustness.⁵¹

FIGURE 2. Mechanical properties of the LM-H. (A) Typical stress–strain curves of the LM-H with different EGaIn contents. (B) Elastic modulus of the LM-H with different EGaIn contents. (C) The maximum stretchability of different the LM-H. (D) The frequency dependence of modulus measurement of the LM-H embedded with 0 wt.% and 1.4 wt.% LM, and (E) the corresponding tan θ. (F) Loading and unloading resistance responses of the LM-H strain sensor with a strain of 400%.

In addition to robustness, the excellent anti-puncture property of the LM-H (Figure S12) further ensures its stability in practical applications. The SEM images of the LM-H at a strain of 0%, 40%, and 100% support fracture mechanism. As shown in Figure S13, we perform SEM images to analyze the antifracture nature of the LM-H. The polymeric network exhibits a favose structure in the original state of LM-H. As the LM-H stretches from 0% to 100%, the favose structure of the LM-H gradually transforms into cracks. This phenomenon suggests that the polymeric network of LM-H can change from a chaos state into a linear state. Within a certain stretching range, the crosslinked network not only remains unbroken but also keeps a reversible stretching property. These properties give the mechanical robustness to the LM-H.⁵²

The hysteresis under deformations is an important consideration for wearable electronics as their application scenarios inevitably encounter various mechanical deformations which may cause device failure.⁵³ Figure 2F shows the loading-unloading curves of LM-H with 1.4 wt.% EGaIn content at room temperature at a tensile strain of 400% in which LM-H strain sensors exhibit low hysteresis of 7.5%. The reason is that the broken noncovalent interactions can recover automatically due to the dissociation and association of the synergetic dynamic cross-linking of the fluid EGaIn particles, electrostatic interactions, and hydrogen bonding.

Electrical properties of the LM-H

Apart from the brilliant mechanical properties, the introduction of EGaIn also endows the LM-H with excellent electrical and strain-sensitive conductivity and long-term durability. As shown in Figure 3A, the conductivity of LM-Hs with consistent strip shape (25 mm in length, 5 mm in width, and 1 mm in thickness) increases up to 22.4 S m⁻¹ and the sheet resistance of the LM-H decreases to 2.5 Ω with the higher EGaIn content. Benefiting from the enhancement of EGaIn to the electrical conductivity, the conductivity of LM-H with 1.4 wt.% EGaln shows a significant change compared to bare PAAm-SA hydrogel (Figure 3B and Table S1) and the gauge factor (GF) of the LM-H sensor reaches 3.86 with a linear value of 0.997 under large strain (Figure 3C). This phenomenon verifies the fracture mechanism (Figure 1D) that aggregated EGaIn nanodroplets are gradually separated and the resistance of the LM-H increase with the LM-H gradual stretching.

FIGURE 3. Electrical Properties of the LM-H. (A) The resistance and conductivity of the LM-H with different EGaIn contents. (B) The relationship between conductivity and tensile strain of the LM-H with 0 wt.% and 1.4 wt.% EGaIn. (C) The strain sensing property of the LM-H with 1.4 wt.% EGaIn content from 0% to 400%. (D) The relative resistance changes of the strain sensor with 12 000 load/unload cycles at a fixed strain of 400%. (E) The relative resistance changes of the hydrogel sensor response to different bending angles of the finger. (F) Comparison of conductivity, maximum stretchability, and maximum number of cycles between the LM-H and various hydrogels.

The excellent electrical conductivity of the LM-H enables its application as conductors of circuits containing a 3 V power supplier (Figure S14 and Video S1), while its strain-sensitivity can also be clearly demonstrated through hydrogel stretching induced darkening of LED in the circuit. Furthermore, the resistance change (decrease within 6%) of the LM-H exhibits excellent stability and repeatability during continuous stretching for 12 000 times at a strain of 400% (Figure 3D and Figure S15). It indicates that the LM-H possesses excellent reproducibility and durability of upon cycled stretching and releasing processes. Moreover, As shown in Figure S16, LM-H shows decent adhesive capacity to both engineering materials and tissues (e.g., PET, PI, and porcine skin). Such mechanical and electrical properties of the LM-H broaden their application range including physiological signal detection, dynamic environment induction, and so on. To demonstrate the strain sensing ability, we use the LM-H constructed wearable sensors to record a variety of motion signals including finger bending (Figure 3E), wrist bending (Figure S17), swallowing, and coughing (Figure S18). When interfacing with a volunteer's throat, the device can capture subtle vibrations of the skin as signatures of a wide range of physiological processes. In addition to human signals, we have also expanded the application of LM-H in different marine organisms such as shrimp (Figure S19) and crabs (Figure S20) by recording their motor behavior. The ability to record the signal in small deformation has also been proved by writing letters and numbers on the LM-H (Figure S21). By optimizing the composition and gel conditions, the designed LM-H can achieve high conductivity, stretchability, and durability which are superior to the most existing hydrogel strain sensors (Figure 3F and Table S2).

Health monitoring system for sign language translation

The excellent electrical and mechanical characteristics of the LM-H make them serve as good candidates for wearable electronics such as epidermal sensors.⁵⁴ To construct a wearable strain sensor, we fabricate an unsupervised health monitoring system (Figure S22) by connecting the LM-H to a wireless circuit for sign language translation. Figure 4A illustrates the remote health monitoring system by wireless communication with a smart phone which consists of five LM-H-based sensors and a wireless circuit board. In this system, all gestures are converted into electrical signals by the LM-H-based sensors. The sign language translation system can be further improved by simultaneous using multiple gel devices as a sensing array. These combined sensors can perfectly attach with the skin as they exhibit lower Youngs' modulus than tissues. Combined with the skin adhesion of the LM-H, the electrode can withstand finger and joint movements. The wireless monitoring system can wear on the wrist with multiple functions, such as signal conditioning, processing, wireless transmission, and data visualization. This wireless circuit system is small, which facilitates its convenience in the application process (Figure 4B). The assembly process of the health monitoring system consists of four components (Figure 4C): (1) ESP32-Circut layer-PI-Electronic as the important sensing element, (2) a Bluetooth low energy system-on-a-chip (SoC) for data collection, system control, and wireless connectivity, and (3) a small volume flexible printed circuit (PFC) for integrated system functions. As shown in Figure 4D, the process flow begins with analog signal acquisition, then signal conditioning and processing, and the finally signal wirelessly transmit to a customized mobile phone. We utilize a methodical and automatic self-organizing map (SOM) that does not require a set of prelabeled patterns for robust translation of sign language hand gestures to verbal. As shown in Figure 4E, the SOM involves a neuron with the index (i, j) in the 2d array, labeled input vector $X=\left\{{\mathrm{x}}_i:\mathrm{i}=1,\dots ,5\right\}$, weighted vector $W=\left\{{\mathrm{w}}_{\mathit{ji}}:j=1,\dots ,169;i=1,..,5\right\}W=\left\{{\mathrm{w}}_{\mathit{ji}}:j=1,\dots ,169;i=1,..,5\right\}$ where 169 is the total number of neurons, and the discriminant function between the input vector X and the weighted vector ѡ_j of each neuron j is established as the Euclidean distance Dj(x). When interfacing with the volunteer's fingers, the device captures the bending of fingers as signatures of a wide range of physiological processes with a response time of 0.21 s (Figure 4F). To demonstrate real-time sign-to-verbal translation, 10 sign language hand gestures including phrases, alphabets, and numbers (e.g., Thank you, Hello, Help, B, F, Z, 6, 7, 8, and 9) are selected from American Sign Language to represent the fundamental elements of communication (Video S2). We verify the confusion matrix of predicted and true labels in the testing set. Figure 4G demonstrates that the algorithm can distinguish different categories of gestures significantly, realizing simple and efficient gesture recognition function. Thus, the health monitoring system exhibits great potential as the next generation sensor for dynamic interfacing applications.⁵⁵ The capabilities of the health monitoring system based on the LM-H provide a solid platform for future humanoid robots,⁵⁶ wearable devices,^57,58 and so on.

FIGURE 4. Demonstration of the sign-to-verbal translation. (A) Photograph showing the wearable sign-to-verbal translation system translating the sign for “Help” to speech and real-time display on a commercial mobile phone application interface. (B) Photograph of the wireless circuit system. (C) System-level block diagram of the wearable sign-to-verbal translation system, showing analog signal acquisition, processing/wireless transmission, and machine-learning recognition paths from the sensor array to the custom-developed mobile application. (D) Flow chart of the health monitoring system, showing an overview of the process flow of analog signal acquisition, conditioning, processing, and transmission. (E) Schematic diagram of SOM architecture. (F) Photographs of the sign language hand gestures according to American Sign Language and the corresponding generated ΔR/R0 profiles as recognition patterns, which can express short phrases, letters, and numbers. (G) Confusion map for the 10 types of respiration pattern recognition. Target class refers to the collected 10 respiratory signal types and output class refers to the recognized results with the assistance of deep learning.

CONCLUSIONS

In summary, a LM embedded hydrogel is developed to overcome the trade-off between the mechanical and electrical properties in existing conductive hydrogels. Furthermore, we solve challenges of the energy consumption and robustness for the practical application of wearable electronics. The wireless monitoring system is capable of accurate recording body motion signals. We further construct a sensor network with unsupervised learning for motion monitoring to detect different gestures and simultaneously transmit them into language with a response time of 0.21 s. We envision that this hydrogel and its derived sensor network will become a versatile and user-friendly platform to fabricate wearable electronics and create an effective way for human-being communication.

EXPERIMENTAL SECTION Materials

Eutectic Gallium-Indium (EGaln, 75.5 wt.% Ga/24.5 wt.% In. >99.99%), sodium alginate (SA, 99%), acrylamide (AM ≥99%), N,N-methylenebisacrylamide (MBAA, 99%), ammonium persulfate (APS, ≥98%), N,N,N′,N′-tetramethyl ethylenediamine (TEMED, 99%), calcium sulphate dihydrate (CaSO₄‧H₂O, ≥99%) were purchased from Sigma Aldrich (Shanghai, China). A dielectric polyethylene layer (VHB 4905) was obtained from Minnesota Mining and Manufacturing Company (Minnesota, USA). All the chemicals were used without further purification. The deionized water (18 MΩ cm) was used in the experiment.

Preparation of the LM-H

SA aqueous solution (4.8 wt.%) was prepared by dissolving 4.8 g SA particles into 95.2 g deionized water (DI). EGaIn suspensions (0, 0.4, 0.6, 0.8, 1.0, 1.2, and 1.4 wt.%) were prepared by suspending EGaIn, 0.2 M APS, and 2 wt.‰ MBAA solution into 19 wt.% AM aqueous solution, followed by being mixed with SA aqueous solution and de-bubbling for 50 min in a vacuum box. CaSO₄ solution (1 mol L⁻¹, 300 μL) with 10 μL TEMED was injected into above blending, followed by being cured for 1 h in an ultraviolet light chamber (365 nm). MBAA, TEMED, and APS served as the crosslinker, accelerator, and photo-initiator, respectively.

Characterization

The scanning electron microscope (SEM) and elemental mapping images of the hydrogels were taken by a Sigma HD filed scanning electron microscope (ZEISS, Germany). Before measurements, EGaIn nanodroplets and LM-H with 0%, 40%, and 80% deformation were freeze-dried by X0-18S vacuum (ATPIO, China). To enhance image contrast, we dispersed the freeze-dried hydrogel on a flat silicon wafer with a layer of gold sputtering. An XFlash 6130 energy-dispersive spectroscopy (Bruker, USA) and a JEM-2100F high-resolution transmission electron microscopy (JEOL, Japan), and x-ray photoelectron spectra (Thermo Fisher, USA) were used to perform the EDS, TEM, and XPS, respectively. A nanoparticle size analyzer (Nanolink SZ900, China) was used to analyze the particle sizes of EGaIn nanodroplets. The linear variation in the shear rate was measured from 0.1 to 100 Hz at 25°C in rotational mode.

Mechanical test

The mechanical and adhesive properties of the LM-H were evaluated using a ZQ-990LB general mechanical testing machine (Chitake, China). The tensile properties of the LM-H with dumbbell shape were measured using pure-shear tensile tests (testing position of 36 mm in length, 18 mm in width, and 2.5 mm in thickness). All tests were conducted with a constant tensile speed of 100 mm min⁻¹ at room temperature. Young's modulus was calculated according to the slope of the stress–strain curves (within 0%–5% of strain values).

Swelling ratio test

The swelling degrees of hydrogels were evaluated via the changes of weight (W_s). The hydrogel was weighed and then immersed into phosphate buffered solution (PBS) at ambient temperature. At certain intervals, the sample was taken out from PBS and weighed after removing excess solution on it. This process was repeated until the weight of hydrogel reached an equilibrium state.

Electrical conductivity test

The electrical conductivity was measured by the standard four-probe method. The LM-H (25 mm in length, 5 mm in width, and 1 mm in thickness) surface were attached by two copper wire electrodes (diameter, 0.5 mm) via silver paste, followed by being coated with dimethicone oil to prevent water evaporation. The sheet resistance of LM-H (5 mm in length, 1 mm in width, and 1 mm in thickness) was measured by keeping the distance between the two electrodes at 1 mm. The conductivity of the hydrogel (σ) was obtained by E4980AL LCR digital bridge tester (Keysight Technologies, USA) and given by Equation (2).⁵⁹ [Image Omitted. See PDF]where L, R, and S represent the distance between test electrodes, the resistance of hydrogels, and the cross-sectional area of hydrogels, respectively.

Fabrication of the LM-H strain sensor

A LM-H strain sensor was fabricated using LM-H rectangular shape (5 mm in wide, 1 mm in thickness, and 3–6 cm in length) by connecting conductive tapes on both ends. All sensors were encapsulated in a 3 M VHB tape to prevent water evaporation. A LCR was used to monitor the resistance changes in response to body movements. The gauge factor (GF) of the LM-H hydrogel was defined as Equation (3).⁵⁰ [Image Omitted. See PDF]where ΔR, R₀, and ε represent the resistance change of the hydrogel, the initial resistance, and the tensile strain of the hydrogel, respectively.

Fabrication of FPC signal acquisition system based on ESP32

Polyimide (PI) was used as the main substrate material, and a layer of copper was deposited on the whole surface of Flexible Printed Circuit (FPC). The photosensitive material was attached to the FPC to complete the pattern transfer, followed by chemical etching to prepare the circuit layer. The FPC was formed by high-temperature, high-pressure bonding, and surface technology.

Mechanical connection characteristics of the signal acquisition system

Conductive copper foil was placed on our fingers. Because this hydrogel was sticky, it was easy to attach to the fingers, and the electrical signal was transmitted through the copper foil.

Electrical output characteristics

The LM-H was easy to attach to the finger because its good adhesion and the electrical signal was transmitted through the copper foil. The resistance changes of the LM-H caused by gesture switching was real-time captured by the signal acquisition system. The system sent gesture signals through wireless transmission and gestures were recognized by what we developed gesture recognition software.

Implementation of gesture recognition software

The human-machine interface was written by Android integrated development environments to realize the real-time reception, algorithm processing, and prediction of gesture signal data. The on-line speech synthesis technology (http://www.peiyinwa.com/) was used to form the spoken file, which can provide users with a voice reminder function that recognizes gesture results in actual use.

AUTHOR CONTRIBUTIONS

Baoyang Lu, Jun Chen, and Lidong Wu. conceived the idea. Hude Ma, Haiyang Qin, and Na Liu carried out the experiments, analyzed the data, performed demonstration, and wrote the main manuscript text. Xiao Xiao and Xiaofang Pan provided advice on developing the health monitoring systems. Junye Li, Shuqi Dai, and Peiyi Li developed the health monitoring system. Lidong Wu and Jun Chen supervised this study and provided intellectual and technical guidance. Xiao Xiao, Shaolei Wang, Sophia Shen, and Mingjun Huang reviewed and commented on the manuscript. All authors discussed the results and commented on the manuscript.

ACKNOWLEDGMENTS

This work was supported by National Natural Science Foundation of China (22176221, 51763010, 51963011) and the Central Public-interest Scientific Institution Basal Research Fund (CAFS) (2020TD75). Jun Chen acknowledges the Henry Samueli School of Engineering & Applied Science and the Department of Bioengineering at the University of California, Los Angeles for the startup support. Jun Chen also acknowledges the Hellman Fellows Research Grant, the UCLA Pandemic Resources Program Research Award, and the Research Recovery Grant by the UCLA Academic Senate. Hude Ma and Baoyang Lu acknowledge the National Natural Science Foundation of China (51763010, 51963011), Jiangxi Provincial Double Thousand Talents Plan-Youth Program (JXSQ2019201108), Jiangxi Key Laboratory of Flexible Electronics (20212BCD42004), and the National College Student Research Training Program (202211318003). The authors would like to thank Dr. Shaoting Lin from MIT for insightful discussions and Mr. Zhilin Zhang from Jiangxi Science and Technology Normal University for the mechanical stability test.

CONFLICT OF INTEREST STATEMENT

The authors declare no conflict of interest.