Ion Migration‐Modulated Flexible MXene Synapse for Biomimetic Multimode Afferent Nervous System: Material and Motion Cognition

The biomimetic afferent nervous system (ANS) is significant for transporting external stimuli into intelligent robots; however, it is still far from bionics due to traditional multisensor and single‐cognition strategies. Herein, a flexible biomimetic ANS with multimode fuzzy perception and brain‐like cognition is developed, by fusing the ion migration‐modulated MXene synapse and machine learning (ML). First of all, the elementary perceptual ability to mimic biological neuroreceptors is demonstrated. Motion artifact in the ionic conductive elastomer (ICE) current is eliminated. Furthermore, the multimode perception and brain‐like cognition are accomplished by synaptic currents (SCs) and nine ML algorithms. Finally, the cognitions of materials, gestures, and motions are conducted by distributing the ANS devices across body joints, and with the optimal ML method, the accuracies can reach 80%, 100%, and 90%, respectively. This synapse‐based ANS may provide a new idea for developing next‐generation neuromorphic intelligent robots.


Introduction
[3] The synapse is a hub point in nerve fibers and has a significant role in ANS by relaying the neural signals from the receptor to the central nerve. [4][19][20][21] Furthermore, with the aid of the wearable synapses, the human gestures [22] and limb motions [23] were monitored using the synapse-based ANS; even the hardness of materials was also felt. [24][27][28][29][30] However, the normal strategies to mimic the biological receptors are mainly focused on building multitype sensor arrays, [31,32] the geometric size is still the obstacle in the development of emerging intelligent technologies, like electronic skin, intelligent prostheses, and neural repair, and it is necessary to pursue a new solution.
In the meantime, to endow the ANS with advanced cognition functions for analyzing and processing perceived external information, [13,33] extensive and in-depth studies of intelligent algorithms are carried out. [12]For example, the artificial neural network (ANN) was exploited to build an optoelectronic spiking afferent nerve, which was capable of not only distinguishing pressure inputs, but also recognizing Morse code, braille characters, and object movement; [34] the deep learning technique was introduced into an artificial tactile skin device to classify the surface textures of 20 fabrics; [35] and the employment of ANN in building the gesture recognition system also demonstrated the great prospection of the biomimetic ANS. [22]However, the single cognition model used in the current biomimetic ANS is still difficult to simulate the generation of optimal options in the brain during the parallel processing of multiple events which has been discovered by modern neuroscience. [36]Since there are various consciousnesses (like gestures, hardness, and coldor-hot) when forming the ultimate cognition, it is necessary to develop a strategy based on the optimal option for dealing with the signals perceived from multiple channels in ANS.If this idea can be accomplished and combined with the synapse-based ANS device, it may promote the evolution of biomimetic ANS from primitive perception to intelligent cognition.
Herein, a flexible ion migration-modulated MXene synapsebased biomimetic multimode ANS is proposed to realize the cognition for tactile, gestures, and motions, which are implemented and demonstrated at the level of hardware device and algorithm, as depicted in Figure 1b.Our previously developed MXene and ionic conductive elastomer (ICE)-based ANS devices [26] are used to perceive the exerted stimuli and directly convert them to the neural electrical signals.Then they are subsequently treated by the optimal cognition-based machine learning (ML) approaches, like AdaBoost, Bagging, Voting, etc. First, the sensitivity and reliability of the synaptic current (SC) of the wearable ANS are testified by measuring their SC variations in response to the applied physical stimuli, and the noise inhabitation is found in SC signals by comparing them with the currents through ICE.Moreover, the deduction for the underlying physical mechanism in this ANS is presented according to the conductive filamentbased synapse working principle [37] and the ion migration in the ICE layer like in the tactile organ of the organism. [38]hen, the ANS device worn on the finger is used to sense different materials, the measured SCs are found to be altered in accompany with the contacted material types, and it is deduced to be a comprehensively multimode sensibility to the different hardness, roughness, or the thermal conductivity at the interface with ICE, which is analog to the tactile organ on the biological receptor.The collected SCs are recorded in a database and processed by a variety of ML classifiers to train and test the material cognition model, which is similar to the cognition process in the central nerve.The assessments of nine ML methods in the view of the train accuracy, the test accuracy, and running time indicate the AdaBoost is an optimal algorithm for material cognition.In addition, using the trained models, material cognition is examined and presented in the form of a confusion matrix.Third, gesture cognition is examined by wearing the glove equipped with the proposed ANS devices.The responding intensities of SCs are found to be in positive correlation with the bending degrees of the joints; we deduce the devices can be deformed in accordance with the extension and flexion of phalangeal joints, then ions in the ICE layer can be redistributed; consequently, the conductive filament can be modified with a result of SC changing.Using similar data processing methods mentioned above, the gesture cognition models are trained, and it is found that the accuracy can reach 100% using the naive Bayesian classifier (NBC); meanwhile, the shortest training time can also be obtained, which is beneficial to reducing the systematic cost.Last but not least, motion cognition is tested by distributing the devices on the joints of limbs and body, so that the collected SC responses are used to identify the joint actions.The degrees of different joint movements can be distinguished with accuracies higher than 86%, and identification of the motion position can also be achieved with an accuracy of 90.21%.Conclusively, the proposed synapsed-based ANS is demonstrated to be capable of simulating the integral processes in the ANS of organisms including information perception and brain-like cognition.We believe this work is helpful for constructing the biomimetic neurological system by merging synapse-based multimode bionic receptors with artificial intelligence algorithms, which may pave the way for the coming age of intelligent robots.

Synapse-Based ANS Device
The synapse-based ANS is based on our recently developed synaptic sensors. [26]In this work to probe the methods for weaving them as an ANS neural mesh, more investigations about elementary characterizations of the as-prepared devices should be conducted.The architecture and intrinsic principle for perceiving the external stimuli are outlined.On the flexible polyethylene terephthalate (PET) substrate, the sandwiched MXene layer between the bottom (carbon past) and the top synaptic electrode (Ag) (the ANS device structure and fabrication processes are depicted in Figure 2a and S1, Supporting Information) is the functional layer of the device.Most areas of its top side are coated with ICE layer, which has the uniform surface (Figure 2b) and is microporous, as shown by uniform holes, with diameters of about 10 μm (Figure 2c), which has been acknowledged to play an active role in constructing tactile receptors. [37,38][41][42] It is generally accepted, in the nanostructure of the MXene layer, that the conductive filament can be formed by driving Ag atoms from the top synaptic electrode under an appropriate voltage; then the electron current avenue from top to bottom (or vice versa) can be set or reset with a result of the plastic synaptic current, namely SC. [43] Herein, an ion migration-modulated conductive filament model is proposed, as diagramed in Figure 2g.The electrostatically driven ionic transportations of EMIM þ and BF 4 À in the ICE layer are similar to the ion migration in tactile organs, [44] the distributions of negative and positive ions in ICE can be altered by the external stimuli (pressure or heat), and then the charged vacancies in its underneath MXene layer may be attracted or expelled, so that the formation of the conductive filament can be adjusted by the outside stimuli.The typical resistive switching behavior of the as-prepared device is testified by the shifted SC with the swept voltages applied on the synaptic electrode, as shown in Figure 2h.The long-term potentiation/depression characteristics of the ANS device are also displayed in Figure S2, Supporting Information.These experiment results are in accordance with the published literature [45][46][47] and suggest that the aggregation and dissipation of Ag in the conductive filament region are brought about in the ANS device.According to the set current and voltage in Figure 2h, the set power consumption is calculated as 21 μW and is lower than other related works (Table S1, Supporting Information), which displays that the ANS device has the advantage of low power consumption.The modulation of ICE on SC and stability of currents under different voltages applied on the ICE electrode can be found in Figure S3, Supporting Information.Furthermore, the responses to the external stimuli, which are the weights put on the ICE layer (Figure 2i) and the bending degrees of the devices (Figure 2j), are examined.Additionally, small deformations generated by the weight of 1 g and the pulse beating can also be detected, as shown in Figure S4 and S5, Supporting Information.It is demonstrated that the plasticity of SC can be made by applied stimuli as our theoretical anticipation, and its repeatability is proven (Figure S6 and S7, Supporting Information), so it should be a good candidate as an elementary component in ANS.However, in our previous work, the currents through the other electrode (i.e., ICE electrode in this work, named ICE current here) also exhibited sensitivity to environmental variations. [26]xtra comparisons between SC and ICE currents are performed by wearing the device on the fingertip and rhythmically clicking desktop; the recorded currents in Figure 2k-n clearly denote there are fewer noises in the SC signals (orange curves) than ICE currents (blue curves), and the signal-to-noise ratio (SNR) of SC is calculated as 34 dB, which is higher than that of ICE current (28 dB).This result can be understood by the ICEmodulated conductive filament from two points.1) The ionic migration is triggered by clicking, because the bias voltage remains constant during the test; the current changes can only be evoked by the disturbed ion distribution in the ICE layer.
2) The inhibited noises in SCs are inherent from its conductive filament-based current forming mechanism, because under the controlled working conditions, SC can only be changed by the perturbed vacancies; though there may be some randomly weak vibrations as measured by ICE currents, they just flash across faintly and cannot produce disturbances to SC.That is to say, the motion artifact [48] in the ICE current, which is caused by the over-sensitivity of ion migration to external disturbances, [49][50][51] can be eliminated by coupling the ion migration into the conductive filament of MXene synapse with the aid of trapassisted tunneling effect.By the way, the lower values of SC (10-12 mA) than ICE currents (25-50 mA) suggest that using SC as the "fake tactile" signal is conducive to reducing power costs.

Material Cognition
As is well known, humans can intuitively judge the types of touched materials, like metal, plastic, cloth, and so on, according to comprehensively analyzing the basic feelings of (but not limited to) hardness, roughness, or thermal conductivity.Normal strategies to fulfill the aim of ANS depend on the integration of multisensors and the algorithm to imitate human identification processes; the geometric size of the sensor group and the cost are the main obstacles for implanting the massive artificial ANS into the finite skin area of the robots or human beings.The interesting multimode sensitive synapse for temperature, pressure, and deformation developed in our previous work [26] provides a new chance to be free from the multi-sensor-based receptor; in this section the experiments and calculations for distinguishing nine kinds of materials are presented using the above-tested synapse-based ANS devices to evaluate the possibility of the proposed ANS, as well as optimize the algorithm.The scheme of the work in this part is illustrated in Figure 3a; the MXene and ICE-based ANS is mounted on the glove, and the finger is contacted with nine different materials, which are PET, glass, metal, cotton, wood, skin, polyvinyl chloride (PVC), plastic, and cloth (composed of cotton and nylon), as listed in the first line in Figure 3b.The signals of SC are collected during the contact actions; then the database of the acquired data is set up and used to train the models for material cognition by applying different ML methods including decision tree (DT), support vector machine (SVM), NBC, logistic regression classifier (LRC), random forest (RF), voting classifier, bagging classifier, AdaBoost classifier, and Multilayer Perceptron (MLP).The classification performance of these methods is determined by their hyperparameters (the detailed meanings of these hyperparameters are listed in Table S2, Supporting Information), for example, the hyperparameter C in SVM is determined by the balance between model accuracy and complexity, because its increase can improve the accuracy, but cause overfitting.To further optimize these models, Grid Search (GS) method is introduced to search the optimal hyperparameters of different models.Then these algorithms are evaluated by comparing their train accuracy, test accuracy, and running time.Finally, the results for identified material types are output.
The experimental results in Figure 3b demonstrate there are clear deviations among the SC signals when the finger contacts nine kinds of materials in the same manner.We think these diversities of SCs are the synthetical responses of the device to the differences in objects like hardness, surface smoothness or elasticity, as well as the thermal conductivity.To confirm this idea, the data of material characteristics (roughness, hardness, conductivity, elastic modulus, and thermal conductivity) are looked up and listed in Table S3, Supporting Information, and the main features (amplitude variation, response time, and recovery time) in the SC response signals when touching different materials are extracted and displayed in Table S4, Supporting Information.The Pearson correlation coefficients (PCCs) between material characteristics and SC features are calculated and plotted in Figure S8, Supporting Information.According to the classification of correlation degree listed in Table S5, Supporting Information, it can be found there are correlations between material features and SC signals.It is demonstrated that the material features including hardness, roughness, elasticity, conductivity, and thermal conductivity may all cause the ion migration in ICE layer and then affect the conductive filament as illustrated in Section 2.1.These measured results suggest the proposed synapse-based ANS device is very similar to the receptor in the biological ANS (Figure 1a).It can perceive external information, though there is no conventionally assessed selectivity in this device.Fortunately, the perceived information can be treated by artificial intelligence algorithms, just like the central nerve in the biological ANS.The similar contact tests with different materials are repeated, the acquired database is composed of 947 datasets with 600 data points in each of them, and then randomly divided into two parts (the train set and test set, with a ratio of 8:2) for model training.The nine ML models are optimized by applying the GS method to search the optimal hyperparameters, and the searching ranges and the values of searched hyperparameters are provided in Table S6, Supporting Information.After that, the assessments of the nine optimal ML models are conducted and presented in Figure 3c.The best test accuracy of 80% belongs to the AdaBoost algorithm; it indicates the AdaBoost-based cognition method is an optimal one for material recognition.This experimental result is in conformity with the theoretical principles and application conditions of AdaBoost provided in Table S7, Supporting Information.At last, a new dataset is imported into the trained AdaBoost model; the material cognition results are shown by the confusion matrix in Figure 3d.The labels on the horizontal ordinate are the true types of objects, while the ones on the vertical ordinate are the predicted types using AdaBoost.The highest accuracy of 100% means all the SC signals for contacting wood and plastic are successfully identified as wood and plastic, respectively.On the contrary, 60% of the perceived data for samples of glass can be successfully distinguished, while 25% and 15% of them are wrongly grouped as cloth and metal, respectively.The average accuracy of our limited test is 80%; we think it can be dramatically increased with the expansion of data volume.Besides, the confusion matrixes of the other algorithms are shown in Figure S9, Supporting Information, which reveal that the AdaBoost method can accurately classify the largest number of material types.This synapse-based ANS strategy is an artificial duplicate of biological ANS and may be a good candidate to realize tactile cognition.

Gesture Cognition
Benefitting from the studies on material cognition, the gesture cognition based on the proposed ANS device is executed using a similar technical route as given in Figure 4a.Extension and flexion of finger joints are monitored by wearing the glove mounted with the prepared devices, as shown in the inserted photo of Figure 4a.Their SC signals in the actions of nine international common gestures (named A-I) are collected and used to build the database for training and testing the models of gesture cognition which are produced using nine ML methods given in Figure 3a 2j, that is to say, the intensities of SC signals from the fingers may be shaped by the gestures.Subsequently, the gesture dataset is assembled by integrating SC signals from five fingers in the order of thumb, index, middle, ring, and little.There are 115 tests for each of the nine gestures, so the database is composed of 1035 gesture datasets, which are randomly divided into the train set and test set with the ratio of 8:2.The number of the train and test set samples is presented in the right of Figure 4a.Furthermore, the models are built using different ML methods, GS method is also introduced to search the optimal hyperparameters of each model, and the searching ranges and the values of searched hyperparameters using the GS method are listed in Table S8, Supporting Information.Then they are assessed by comparing their train accuracy, test accuracy, and running time as shown in Figure 4c.It can be seen that the accuracies of seven algorithms in nine are up to 100%; considering efficiency and cost, the NBC algorithm is selected as an example for the following tests because of its shorter running time.This experimental result is in conformity with the theoretical principles and application conditions of NBC provided in Table S7, Supporting Information.By feeding new datasets in the database, the trained NBC model is used for gesture cognition, the output results are presented in the confusion matrix as shown in Figure 4d, which demonstrate that all gestures can be distinguished accurately, and more confusion matrices of the other algorithms are provided in Figure S10, Supporting Information.It can be found that seven of nine methods can accurately discriminate the gestures, and NBC is preferred due to its shorter running time, which is beneficial to reducing the systematic cost.The proposed ANS device has advantages compared to other flexible sensors (Table S9, Supporting Information) because it has a more biologically similar device structure, sensing mechanism and cognition strategy, as well as higher classification accuracy (100%) and shorter response time (11.85 ms, as shown in Figure S11, Supporting Information).

Motion Cognition
According to the above-mentioned works, the integration of the multimode sensibility and intelligent algorithms of the proposed biomimetic ANS device is expanded to the limb joints; thereby, more tests for motion cognition are conducted.Joint extension and flexion are fundamental to motions; thereby, the prepared devices are distributed on the joints marked by (b) to (i) in Figure 5a; then the triggered SC signals are recorded, which clearly demonstrate that joint states (exemplified as the bending degrees of 15°and 30°) on limbs (like finger (Figure 5b), wrist (Figure 5c), elbow (Figure 5d), ankle (Figure 5e), knee (Figure 5f ), hip (Figure 5g)) and body (like neck (Figure 5h) and shoulder (Figure 5i)) can all be monitored.These data indicate the ANS devices can catch not only the obvious motions happening on limb joints, but also the relative slight changes of motion degrees at the same joint.Meanwhile, these experiments also show the uniformity between the amplitudes of SCs and the deformation degrees of the articulus, which has been found in gesture recognition.Thereby, using the same data processing flow (Figure 4a) and the method for comprising the gesture dataset, the posture dataset is formed by the collected SCs data from distributed positions (Figure 5a); then the database is composed of repetitively acquired SC data (115 times).Furthermore, using the same nine ML approaches and assessing methods given in gesture and material recognitions, the studies of motion identifications are executed.The accuracies of nine ML approaches for recognizing the motion degrees at each of the eight joints are compared and presented in Figure 5j, which are up to 86%, even 100% at the positions of neck, finger, elbow, and knee.Meanwhile, the optimization of the hyperparameters in different ML methods is also conducted as mentioned in material and gesture cognition, and the searching ranges and the values of the searched hyperparameters values using the GS method are displayed in Table S10, Supporting Information.Based on the results presented in Figure S12, Supporting Information, AdaBoost is demonstrated as the optimal ML algorithm for recognizing the position of the moving joint.The confusion matrix of AdaBoost for motion position recognition is presented in Figure 5k, and the average accuracy can reach 90.21%.
In addition, the comparisons of the classification accuracies by respectively using the ICE currents and SCs for material, gesture, and motion cognitions are provided in Table S11, Supporting Information.It can be found that the classification accuracies using SCs are higher than ICE currents, which demonstrate that the SCs have advantages in cognition due to their stability mentioned in Section 2.1.The horizontal comparisons with other literature about neuromorphic devices are given in Table S12, Supporting Information.Compared with the other literature, which mainly focused on biomimetic reproduction of hand tactile, the application scope of the proposed ANS device in this work is expanded to almost all the body joints, including not only the joints of finger, but also the knee, neck, elbow, wrist, ankle, hip, and shoulder.It is believed that the proposed ANS device may be a good candidate to substitute the conventional ANS strategy and pave the way for the coming age of intelligent robots.

Conclusion
In summary, a biomimetic multimode ANS by integrating the biological receptor, neural transmission, and cognition is designed and demonstrated step by step.First, the multimode plasticity of SC is evidenced to be owned by the proposed ANS.The motion artifact in the ICE current can be eliminated by the ion migration-modulated MXene synapse.Second, the multimode perception ability of the ANS can be demonstrated using material cognition as a proof of concept.The variation of SC is found to be related to the material type and is the comprehensive response to the deviations of materials in hardness, roughness, thermal conductivity, etc.Then, the brain-like cognition is carried out by evaluating nine different ML methods; it can be found that accuracy for material recognition can reach 80%.Furthermore, the proposed ANS is utilized for the identification of gestures and motions; the average accuracies are higher than 90%.In a word, this synapse-based ANS can biomimetically implement the multimode sensation and thinking process, just like the biological ANS in organisms.Together with its merits of low power consumption and noise inhibition, the proposed ANS may be a good candidate to substitute the conventional ANS strategy.

Experimental Section
Synthesis of the MXene Nanosheets: First, 0.333 g HF, 2.5 mL 12 M hydrochloric acid, and 0.5 g Ti 3 AlC 2 powder were mixed; then 2.5 mL deionized water was added and purged with N 2 for 20 min.After stirring in silicon oil bath (40 °C) for 24 h, the resulting solution was centrifuged (4000 rpm min À1 ) in 40 mL deionized water for 5 min.After centrifugation for five times, the obtained material was dried in the constant temperature drying oven at 80 °C for 24 h.Then the multilayer Ti 3 C 2 T x nanosheets were obtained.Ultrasonic dissociation for 10 h was conducted after adding 80 mL deionized water to 100 mg multilayer Ti 3 C 2 T x nanosheets.Subsequently, the mixture was centrifuged at 4000 rpm for 10 min.After centrifugation, the supernatant was collected to obtain uniform few-layer Ti 3 C 2 T x nanosheets solution.
Preparation of the ICE Layer: First, 1 g thermoplastic polyurethane elastomer (TPU) and 2 g 1-Ethyl-3-methylimidazolium Tetrafluoroborate (EMIMBF 4 ) were dissolved in 30 mL N-Dime-thylformamide (DMF) at 80 °C overnight to obtain a homogeneous solution.Subsequently, 2 mL of the EMIMBF 4 /TPU mixed solution was dropped on a 1.5 cm Â 1.5 cm glass substrate and dried at 80 °C for 3 days to form ICE layer.Finally, the conductive carbon paste was uniformly placed on the surface of the ICE layer and dried at 80 °C for 2 h to form the electrode layer.
Fabrication of the ANS Device: In the manufacturing process, the ANS device was placed on a 1 cm Â 1 cm flexible PET substrate.The conductive carbon paste was coated on the PET substrate by the microelectronic printer (Shanghai Mifang Electronic Technology Co., Ltd., China) and used as bottom electrode.Subsequently, the electrode layer was spin coated three times and annealed at 80 ºC to form the dense MXene-Ti 3 C 2 T x film.The silver electrode of the artificial synapse was prepared by the microelectronic printer on part of the MXene film and used as synaptic electrode of the ANS device.Then the prepared ICE layer was cut to the size of the PET substrate and coated on the surface MXene layer.Finally, the as-prepared device was insulated with insulating ink and dried at 80 ºC for 3 h.
Test Conditions: The tests mentioned in this work were all conducted with the ambient temperature remaining around 25 °C.For the purpose of imitating human real actions to feel material types, the pressure of the finger in material recognition was maintained at a relatively stable level using the similar touching action, instead of some specific instrument.
Machine Learning Models: The cognition strategy used in this work was similar to the optimal option generation in the brain during the parallel processing of multiple events.Nine ML algorithms (including DT, SVM, NBC, LRC, RF, Voting, Bagging, AdaBoost, and MLP) were introduced to simulate the different thinking ways in the brain.These ML algorithms models were implemented on Jupyter Notebook (Cité des Sciences, Paris) platform using Python (Python Software Foundation, Netherlands) programming language based on the scikit-learn tool kit.GS method was used to search the optimal hyperparameters for the nine algorithms; the detailed meanings of hyperparameters are listed in Table S2, Supporting Information.The searching ranges and the searched values of hyperparameters using the GS method are listed in Table S6, S8, and S10, Supporting Information.The response SCs were directly read from Agilent semiconductor parameter analyzer and stored as a dataset.947 datasets were used in material cognition, 1035 datasets were used in gesture cognition, and 1680 datasets were used in body motion cognition.All these datasets were split into the train-test set with a splitting ratio of 80-20%.These ML methods were evaluated according to their test accuracy first; for the same accuracy the running time was considered.
Apparatus: The transmission electron microscope (TEM) images were carried out using Tecnai G2 F20 (FEI, USA).Scanning electron microscope (SEM) spectrum was obtained using Quanta 200 (FEI, USA).All electrical measurements are performed using the Agilent semiconductor parameter analyzer.
The Statement for Human Subjects: A disclaimer of signed informed consent from the person who participated in the experiments with human subject was obtained.

Figure 1 .
Figure 1.Comparison diagram of biological and biomimetic ANS.a) Biological ANS is composed of receptors, nerve fibers, and the central nerve, the synapse is a hub point in nerve fiber.b) Biomimetic ANS is composed of the MXene and ICE-based synapse to mimic multimode sensitive tactile receptors in biological ANS, and the ML approach-based calculations to imitate cognition processes in the biological central nerve.

Figure 2 .
Figure 2. a) The schematic diagram of the synapse-based ANS device, which is composed of an insulation layer, an ICE electrode, an ICE layer, MXene, a synaptic electrode, and a bottom electrode on the flexible substrate.b) The SEM image of the surface topography of the ICE layer.c) The SEM image of the cross-section of the ICE layer.d) The TEM surface morphology of the MXene layer.e) The enlarged area of the white dashed box in (d).f ) The SEM photo of the MXene layer.g) Illustration of the physical mechanism of the proposed device.h) Cyclic voltammetric curves of SC, the swept voltages are exerted on the synaptic electrode, ICE electrode is unbiased.i) SC responses when different weights (10, 20, and 50 g) are put on the ICE layer.Voltages applied to the synaptic electrode and ICE electrode are 0.5 and 4 V, respectively.j) SC responses when the bending angles are 30°, 45°, 60°, and 90°.k-n) The comparisons of ICE currents and SCs by wearing the device on the fingertip and rhythmically clicking the desktop under the same bias conditions given earlier.

Figure 3 .
Figure 3. Conceptual demonstration for material cognition.a) Scheme for material cognition including data acquisition of SCs, the training, testing, and evaluation of models.The prepared ANS device is mounted on the glove, and contacted nine kinds of materials with a same manner are shown in the first line in (b); the acquired data are incorporated into a database, which is randomly divided into two parts, the train set and the test set, with the ratio of 8:2.Nine ML methods are tested, which are DT, SVM, NBC, LRC, RF, Voting, Bagging, AdaBoost, and MLP.b) The recorded signals of SC during the actions of contacting nine materials, PET, PVC, glass, metal, cotton, wood, skin, plastic, and cloth.c) The evaluations of different ML methods for material cognition, by comparing train accuracy, test accuracy, and running time.d) The confusion matrix for the accuracy of material cognition classification using the AdaBoost algorithm.
, individually.All nine gestures are repeatedly tested 115 times, the signals given in Figure 4b are instances of each gesture.It could be found that the SCs detected from different fingers are changed with the bending of different fingers.These extension and flexion of joints related SC signals are in agreement with the basic feature of the device provided in

Figure 4 .
Figure 4. Gesture cognition based on the synapse-based ANS device.a) Flow chart of data process for gesture cognition system, in which nine gestures (nominated as A-I) are tested.b) SC signals of different gestures, while the voltages applied on the synaptic and ICE electrodes are controlled at 0.5 and 4 V, respectively.c) The classification accuracies and the running time of nine ML algorithms applied to gesture cognition.d) The confusion matrix for the accuracy of gesture cognition classification when using the NBC algorithm.

Figure
Figure2j, that is to say, the intensities of SC signals from the fingers may be shaped by the gestures.Subsequently, the gesture dataset is assembled by integrating SC signals from five fingers in the order of thumb, index, middle, ring, and little.There are 115 tests for each of the nine gestures, so the database is composed of 1035 gesture datasets, which are randomly divided into the train set and test set with the ratio of 8:2.The number of the train and test set samples is presented in the right of Figure4a.Furthermore, the models are built using different ML methods, GS method is also introduced to search the optimal hyperparameters of each model, and the searching ranges and the values of searched hyperparameters using the GS method are listed in TableS8, Supporting Information.Then they are assessed by comparing their train accuracy, test accuracy, and running time as shown in Figure4c.It can be seen that the accuracies of seven algorithms in nine are up to 100%; considering efficiency and cost, the NBC algorithm is selected as an example for the

Figure 5 .
Figure 5. Motion cognition experimental results.a) The distribution of joints equipped with ANS devices.The motions on b) finger, c) wrist, d) elbow, e) ankle, f ) knee, g) hip, h) neck, and i) shoulder are monitored by the synapse-based ANS device, and the triggered SC signals are recorded in the last line of (b-i).j) The accuracies of nine ML approaches for recognizing the motion degrees at eight joints, and the highest ones are marked.k) The confusion matrix for classifying the position of the moving joint when using the AdaBoost algorithm.