Integration of Flexible Touch Panels and Machine Learning: Applications and Techniques

The rapid advancement of mobile devices and human–machine interaction technologies has ushered in a new era for flexible touch panels as a novel input interface. This article reviews the historical evolution and technical progress of flexible touch panel technologies, from rudimentary single‐point touch to sophisticated grid‐free touch systems. Additionally, the working principles and mechanisms that underpin these advanced systems, including capacitive, resistive, piezoelectric, and triboelectric nanogenerator technologies, are explored. Following this, the integration of machine learning methods into these panels is discussed, offering new avenues for enhancing user experience and expanding functional capabilities. Various machine learning algorithms such as support vector machines, artificial neural networks, convolutional neural networks, and k‐nearest neighbors are examined for their potential applications in touch panel technologies. Finally, the challenges and prospects for the application of flexible touch panels fused with machine learning are discussed.


Introduction
With the continuous advancement of flexible electronic technology, [1] tactile sensors [2] are seeing increasingly broad applications in various devices.These sensors are capable of not only detecting user touch inputs [3] but also perceiving the shape, [4] texture, [5] and pressure of objects, [6] offering users a more authentic and multidimensional interactive experience. [7]n recent years, as mobile devices, [8] health monitoring, [9] and human-machine interaction [10] have rapidly developed, flexible touch panels have become a focal point.These panels aim to detect and interpret user touch inputs and offer not only the functionalities of traditional touch screens [11] but also remarkable flexibility [12] and adaptability for use in curved and complex-shaped devices.Depending on the tactile mechanisms and technical types of flexible touch panels, a wide range of application scenarios and fields can be accommodated.
Flexible touch panels, [13] as an innovative input interface, are gradually gaining attention.Concurrently, the introduction of machine learning (ML) [2,14] into the domain of touch panels opens up new possibilities for enhancing interactive experiences and functionalities. [15]As a potent tool and methodology, ML has demonstrated tremendous potential across various sectors. [16]In the realm of flexible touch panels, ML can optimize touch recognition algorithms [17] and implement higher level functions like multiuser authentication, [18] touch pattern recognition, [19] and soft biometric recognition. [20]By learning and extracting patterns from extensive touch data, ML enables flexible touch panels to better adapt to user needs and provide personalized interactive [21] experiences.Integrating flexible touch panels with ML not only boosts data processing [22] and decision-making efficiency [23] but also explores new application scenarios. [24]For example, in health monitoring, [25] ML algorithms allow flexible touch panels to more accurately identify [26] and interpret [27] biosignals; [28] in smart homes, [29] learning user behavior [30] and preferences [31] enables more personalized services. [32]herefore, this article aims to explore the close relationship between flexible touch panels and ML, delving into their applications in scientific research.Section 2 reviews the historical evolution and technical advancements of flexible touch panels, from single-point to array-type to grid-free developments.Section 3 introduces the working principles and applications of flexible touch panels, including capacitive, resistive, piezoelectric, and triboelectric mechanisms.Section 4 presents the main types of ML models (support vector machines [SVM], k-nearest neighbors [k-NN], artificial neural networks [ANN], and convolutional neural networks [CNN]) and their applications in flexible touch panels.Section 5 outlines the future challenges and prospects of ML-based flexible touch panels.Through in-depth research, we can anticipate that more innovations and applications will emerge in this interdisciplinary field Figure 1.

Evolution and Technical Advancements in Flexible Touch Panels
Since its inception, flexible touch panel technology has undergone a remarkable historical evolution.Chronologically, there are three main types of touch technology: single-point touch, array touch, and flexible grid-free touch.Single-point touch is suitable for basic applications due to its simplicity and low cost.Array touch has become the mainstream smart device of choice due to its high degree of interactivity and accuracy.Grid-free flexible touch panels, on the other hand, open up new directions of innovation, showing their unique value, especially in wearable technology and new display technologies.
Initially, single-point touch technology [33] was in its earliest form.Flexible single-point touch panels were developed for interactive devices, [34] allowing users to input information, [35] control applications, [36] and execute various operations by touching the screen surface. [37]Unlike traditional rigid touch panels, these flexible single-point versions are manufactured using flexible materials, offering greater adaptability and portability. [38]As illustrated in Figure 2a, Zamanpour et al. [39] added BaTiO3 [40] to organosilicon polymers as a charge generation layer and deposited it onto a conductive layer based on silver nanowires to form a composite layer, creating a transparent, flexible touch panel based on triboelectric nanogenerators [41] (TENGs).It was employed as a transparent, elastic, flexible, and ultralight touch on/off switch, and used for activating/deactivating LEDs, thermometers, and oximeters.However, the limitations of this single-input modality spurred researchers to explore additional interactive approaches.
Subsequently, array-type touch technology [42] emerged, introducing multitouch capabilities that allow users to perform more complex gestures using multiple fingers, such as controlling the movement of game characters or displaying text. [43]This advancement profoundly altered human-machine interaction methods. [44]Li and co-workers [45] developed a flexible array-type touch panel based on TENGs using ink writing technology [46] through an all-printing method, encapsulating a silicone layer with silver electrodes (Figure 2b).When a plastic board in the shape of "SZU" was placed on the touch panel, the sensor array sensed touch contours based on voltage modulation.The tactile sensing information capability of this touch panel showed application potential in the field of electronic skin [47] for soft robots [48] to perceive external environments.However, as technology advances, there is an increasing demand for more natural and accurate touch experiences, necessitating the development of new touch panel technologies.
Grid-free touch technology [49] subsequently emerged, liberating itself from the fixed sensor grids [50] on traditional touch panels.This approach employs a more flexible sensor layout, enabling the entire touch surface to sense inputs, thus eliminating dead zones and blind spots while greatly improving accuracy and sensitivity.This technological advancement allows users to interact more freely with fingers, palms, or even stylus tips, paving the way for more creative and personalized methods of interaction.As shown in Figure 2c, Yu et al. [51] utilized silver nanowires [52] (AgNWs) as metal conductive electrodes and thiol-modified nanofibrillated cellulose [53] nanopaper as substrates to fabricate a piezoresistive-based high-precision flexible transparent touchscreen panel.This touch panel allows for handwriting with a finger both in curved and flat states.
In summary, the historical evolution of touch technology, from single-point touch to array-based and then to grid-free systems, not only showcases rapid advancements in the field of human-machine interaction but also offers researchers an unending stream of opportunities for study and innovation.This ongoing progress is driving the wide-ranging application and development of touch technologies across various sectors.

Principles and Applications of Flexible Touch Panels
Flexible touch panel technologies can be categorized by principle into capacitive, resistive, piezoelectric, and TENG technologies.Capacitive touchscreens dominate smartphones and tablets due to their high sensitivity and multitouch capabilities.Resistive touchscreens are still popular in some industrial and commercial applications due to their cost-effectiveness.Piezoelectric touchscreens are mainly used in special applications where high precision and durability are required.Copyright 2022, Elsevier.b) Array-type touch panel used for displaying touch contours.Reproduced with permission. [45]Copyright 2020, Elsevier.c) Grid-free touch panel employed for font writing.Reproduced with permission. [51]Copyright 2021, Elsevier.

Capacitive Touch Panels
Capacitive flexible touch panels represent a type of touch sensor that incorporates capacitive touch technology [54] into flexible substrates.This innovation combines the merits of capacitive sensing and flexible materials, offering adaptability to curved surfaces and varying device shapes, while also featuring high sensitivity and multitouch capabilities.
The working principle and basic structure parallel those of conventional capacitive touch sensors, relying on changes in capacitance to detect user interactions.Typically, these panels consist of a flexible, transparent conductive layer for touch input and an underlying layer of sensing electrodes.The flexibility of the base material allows the panel to bend and conform to diverse surfaces, making it suitable for curved devices, wearables, [55] and other nontraditional screen shapes.The sensing electrode layer comprises multiple capacitive sensors to monitor capacitance changes on the touch surface.When a user's finger or another conductive object comes into contact with the touch surface, there is a change in capacitance at the point of contact due to the electrical charge carried by the human body; this change is then measured by the sensing electrodes.Utilizing multiple capacitive sensors in the sensing layer enables the detection of multiple touch points, thereby supporting multifinger gestures. [56]s illustrated in Figure 3a, significant work in this area includes research by Hong et al., [57] successfully fabricated a grid-free, single-layer capacitive flexible touch panel with multilayer electrodes deposited on a flexible substrate, capable of multitouch [58] on various surfaces.Noteworthy contributions also come from Wu and co-workers, [59] who used an amorphousindium tin oxide/silver/crystalline-indium tin oxide triple-layer structure as electrodes and deposited them on a colorless polyimide [60] substrate to create flexible and foldable touch panels for multitouch applications, as shown in Figure 3b.Furthermore, a compact array-type touch panel produced by Zhang et al. [61] exhibited high precision and sensitivity, providing accurate pressure mapping results with negligible crosstalk between adjacent sensing units (Figure 3c).
In summary, capacitive flexible touch panel technology not only holds immense potential in the realm of human-machine interaction but also shows broad application prospects in sensing and other fields.

Resistive Touch Panels
Resistive flexible touch panels [62] are sensors that rely on the piezoresistive effect of materials to detect touch.When an external object applies pressure to the surface of a resistive touch panel, the flexible material of the panel deforms.This deformation leads to changes in the distance between electrodes or alterations in resistive values, thereby modulating the panel's electrical characteristics.Signal-reading circuits measure these electrical variations and translate them into interpretable tactile information.Compared to capacitive touch panels, a significant downside of resistive panels is their lower sensitivity. [63]s illustrated in Figure 4a, Lee and co-workers [64] have developed a transparent, flexible friction-resistive touch panel with grid-free touch point recognition based on a single-layer TENG.This innovative panel combines the features of both TENGs and piezoresistive effects. [65]It uses piezoresistive mapping for pressure detection, [66] while also leveraging the high precision of TENGs to create a nine-grid touch panel for robotic control, offering a novel solution to the sensitivity issue typically associated with resistive touch panels.
One of the advantages of resistive flexible touch panels is their ease of fabrication.Kim [67] reported the creation of large-area AgNW films coated using a continuous roll-to-roll slot die coater, aiming for cost-effective, large-area flexible touch screen panels.These panels have been successfully implemented for writing functionalities, as shown in Figure 4b.
Furthermore, resistive touch panels can also excel in intricate pattern drawing with effective electrode processing techniques.Cho and co-workers [68] used highly aligned AgNW network electrodes to fabricate transparent flexible resistive touch panels.These panels demonstrated exceptional performance in both text writing and intricate pattern drawing (Figure 4c).
Therefore, resistive flexible touch panels not only have utility in simple touch-based interactions but also show promise in more complex applications.

Piezoelectric Touch Panels
Piezoelectric touch panels [69] serve as versatile sensors capable of detecting and quantifying the force or pressure exerted upon their surfaces.Utilizing the piezoelectric effect, [70] these panels convert the mechanical force applied to their surfaces into electrical signals, thereby capturing and transmitting tactile information.The piezoelectric effect is a physical phenomenon wherein  [64] Copyright 2022, Wiley-VCH.b) Schematic diagram of a high-precision flexible transparent touch panel and photograph of the writing operation.Reproduced with permission. [67]Copyright 2016, Springer Nature.c) Structural diagram of a high-precision flexible transparent touch panel, photographs of handwriting, and drawing imitation.Reproduced with permission. [68]Copyright 2017, American Chemical Society.
certain materials produce charge separation or polarization when subjected to mechanical stress or pressure.
Generally constructed from piezoelectric materials, electrodes, and support structures, the architecture of piezoelectric touch sensors can vary.However, they primarily include a sensing surface coated with piezoelectric materials and electrodes.Piezoelectric materials may consist of ceramic compounds such as lead zirconate titanate, [71] or polymer materials like polyvinylidene fluoride (PVDF). [72]Electrodes, usually placed on either side of the piezoelectric material, are employed to measure variations in electrical signals.When an external object applies force to the sensor's surface, the piezoelectric material undergoes minor deformations, causing charge separation.This separated charge creates a voltage difference, or electrical signal, between the electrodes.By measuring these signals, one can ascertain the magnitude, location, and distribution of the applied pressure.
Piezoelectric touch panels find extensive applications in various domains, including human-machine interaction.As shown in Figure 5a, Nguyen et al. [73] fabricated a flexible, multilayered piezoelectric coating for tactile feedback in automotive touchscreen interfaces via 3D printing.Their tactile demonstrator consists of two touch-panel buttons for toggling the indicator lights on and off while providing vibrational feedback to the user.
Piezoelectric touch panels offer a novel approach to humanmachine interaction, with rapid response time as one of their key advantages.Bi et al. [74] developed an ultrafast responsive  [73] Copyright 2023, Multidisciplinary Digital Publishing Institute.b) Array-type piezoelectric touch panel with color-changing capabilities under different voltage levels.Reproduced with permission. [74]Copyright 2022, Elsevier.c) Schematic of data acquisition in a high-precision piezoelectric touch panel.Reproduced with permission. [75]Copyright 2020, Elsevier.
piezoelectric-induced color-changing transparent touch panel display.With a transition and recovery time of just 1.4 s from transparent to deep blue and a color contrast of 62.2%, the touch panel can reliably display real-time motion trails of fingers or stylus tips, as shown in Figure 5b.This innovative display method also offers new avenues for the application of piezoelectric touch panels.
However, one of the challenges impeding their further development is the low voltage output.Wang et al. [75] engineered a high-performance, energy-efficient, fully flexible piezoelectric touch screen.They achieved this by integrating piezoelectric materials with the mechanical-to-electrical conversion function of β-PVDF nanorod arrays and the signal amplification capabilities of organic field-effect transistor devices, as depicted in Figure 5c.
In summary, piezoelectric touch panels show great promise in terms of utility and functionality.Especially noteworthy is their superior performance in response time and haptic feedback.

Triboelectric Touch Panels
Triboelectric touch panels [76] operate as sensors that leverage the triboelectric effect [77] to detect and measure touch, friction, and other forces.The working principle of these panels is grounded in the concepts of the triboelectric effect and nanogenerators.When two different materials come into frictional or compressive contact, the distribution of electric charge between them changes, resulting in charge separation.This separated charge can generate a voltage difference, or electrical signal, between electrodes. [78]y engineering an array of nanostructures, triboelectric touch panels induce the triboelectric effect upon touch or force application, thereby converting tactile information into electrical signals.The design of triboelectric touch panels typically comprises two main components: the contact layer and the electrode layer.The contact layer is the surface interfacing with external objects and usually consists of materials with differing electron affinities.The electrode layer collects charges generated by the triboelectric effect.Motion and friction between these layers lead to charge separation, creating a voltage difference that can be measured to determine the intensity, location, and changes in touch or friction. [79]riboelectric touch panels exhibit high sensitivity and good stability.Li et al. [80] introduced a transparent, highly sensitive TENG as electronic skin, which detects pressure through four sensor units and controls game character movement, as shown in Figure 6a.
Ease of fabrication and low cost are also advantages of TENGbased touch panels.As demonstrated in Figure 6b, Wang et al. [81] present a facile, low-cost method for fabricating a metalelectrode-free, fully integrated, soft triboelectric touchscreen array.This touchscreen performs excellently in electrically mapping external pressures exerted during text writing.
In summary, different types of flexible touch panel technologies have their own merits and applicability in human-machine interaction.Capacitive panels are suited for consumer electronics, while resistive types are applicable in high-pressure sensing scenarios.Triboelectric and piezoelectric panels represent emerging technologies with self-powered characteristics, promising unique roles in virtual reality, wearable devices, and other domains.As technology continues to advance and innovate, the application scope of these flexible touch panel technologies is expected to further expand.

Applications of ML in Flexible Touch Panels
ML plays a pivotal role in enhancing the capabilities of flexible touch panels.By employing various algorithms, ML facilitates advanced functionalities such as high-precision touch recognition, gesture identification, and secure multiuser authentication.The choice of specific ML algorithms is guided by their unique strengths and compatibility with the requirements of touch panel technology.

SVM
SVM are ML algorithms commonly employed for classification and regression tasks. [82]They hold significant applications in fields such as touch panel calibration and gesture recognition. [83]VM serves as a generalized linear binary classifier, [84] where its decision boundary is the maximal margin hyperplane that separates the input samples. [85]ne of the advantages of SVM is its capability to handle nonlinear relationships, allowing it to achieve excellent performance in complex input data scenarios. [86]During the validation phase, the SVM model determines which vector inputs correspond to which partitioned hyperplane. [87]As a supervised learning algorithm, SVM is effective in handling both classification and regression tasks. [88]Its superior performance in pattern recognition [89] and classification problems [90] makes it a powerful tool for achieving high-precision touch recognition in flexible touch panels.SVM is a popular classification method, in part because it exhibits relatively high performance even with small training datasets. [91]n a groundbreaking study by Cui and co-workers, [92] a new method for secure multiuser authentication using piezoelectric touch panels was developed.This advanced technology captures unique keystroke dynamics by leveraging the duration and force of user touches.Using SVM for data processing and analysis, a classification accuracy of 97% was achieved, thereby confirming the method's effectiveness for secure multiuser authentication.Researchers used piezoelectric touch panels to collect user touch behavior metrics such as force and duration, serving as input features for SVM.This led to a high-precision model for identifying different users based on their unique touch characteristics.Therefore, touch panels capable of providing more accurate time and force information are expected to yield higher authentication accuracy.SVM is not only applicable to the current design of touch panels but can also be applied to any force-input-based devices or systems, such as smartphones and tablets, to offer secure access and protect users' private information.
Seol and co-workers [93] have provided a precise method to differentiate various touch inputs on touch screens, such as no-touch, finger touch, and stylus touch, by using a SVM. Figure 7a displays the "One-versus-the-rest SVM classification scheme and evaluation setup for the proposed touch system."SVM is employed as a classifier to determine the type of touch on the touch screen.Given that the pulse frequencies produced by the stylus and fingers on the touch screen differ, SVM can accurately differentiate between the two based on these frequency variations.
In an article by Bae and co-workers, [94] a flexible large-area tactile sensor array is introduced, effectively eliminating electrical crosstalk between sensor units.The sensor employs a mesh structure and a composite material of carbon nanotubes and polydimethylsiloxane, achieving high sensitivity and durability.It is successfully applied to Braille readers through ML-based information processing.Braille, consisting of six raised dots arranged in 3 rows by 2 columns, is a tactile writing system for the visually impaired.The proposed Braille recognition system is designed to identify Braille alphabets read by the grid-based sensor array through SVM, with corresponding letters being visualized in real-time.SVM serves as the computational backbone for effectively translating tactile data into recognizable Braille characters.The high-dimensional data classification capability of SVM allows the system to accurately discern complex tactile patterns.As shown in Figure 7b, the SVM process is a critical step that bridges the gap between the original sensor readings and the real-time visualization of actionable Braille characters on the display, rendering the system not only powerful but highly reliable.The SVM process helps translate tactile information into corresponding Braille alphabets, which are then displayed in real time on the screen as depicted in Figure 7d.The system first undergoes initial data collection and signal processing.Artificial Braille is loaded onto the sensor, and the resulting initial signals are converted into vector inputs for SVM training.The array layout and visualization settings are shown in Figure 7c.When pressure is applied to the individual units of the sensor array, electrical  [80] Copyright 2023, Elsevier.b) Demonstration of pattern recognition and corresponding electrical response of the array-type touch panel.Reproduced with permission. [81]Copyright 2020, Springer Nature.
fluctuations in all units are captured.The SVM model analyzes these sensor responses to determine the corresponding hyperplane in the feature space, ultimately achieving real-time character recognition.
The advantages of SVM in handling high-dimensional data [95] and complex classification problems [96] are evident, whether applied to aiding the visually impaired in Braille recognition [97] or to highly secure identity authentication. [98]SVM not only enhances the functionality and reliability of the respective systems but also improves their universality and scalability.The efficacy of SVM is particularly manifested in its adaptability and extensibility.As more data are collected by the system, the SVM model can be fine-tuned to achieve higher accuracy, making the technology highly adaptable to different types of Braille inputs and potentially extendable to other tactile-based languages or codes. [99]This scalability paves the way for broader applications in assistive technologies, setting a high standard in the field.The inclusion of SVM significantly propels the advancement of intelligent systems across multiple domains, not limited to visual or tactile assistive devices and secure authentication.Its wide-ranging applications serve as an exemplary integration of ML and sensor technologies, with profound social and economic impacts.

ANN
ANN are computational models inspired by the structure of natural neural systems. [100]They are designed for pattern Figure 7. Application of SVM models in flexible touch panels.a) One-versus-the-rest SVM classification scheme and evaluation setup for the proposed touch system.Reproduced with permission. [93]Copyright 2019, The Japan Society of Applied Physics.b) The process of Braille recognition through SVM.c) The process of Braille recognition using a grid-based sensor array and visualization settings.d) Sensor responses from the tactile sensor array and corresponding outputs displayed on the monitor.Reproduced with permission. [94]Copyright 2021, American Chemical Society.
recognition or solving optimization problems. [101]Composed of multiple layers, including an input layer, one or more hidden layers, and an output layer, each layer contains one or more "neurons" interconnected through weights.These weights are optimized through iterative training and the backpropagation algorithm, allowing for precise classification or regression of input data.Compared to SVM, ANN is generally more flexible and can adapt to more complex data structures and nonlinear relationships. [102]However, this flexibility comes with drawbacks, including the potential for overfitting and higher computational resource requirements.This is why SVM may be a more efficient choice in certain scenarios, particularly when the dataset is small or has high dimensionality but low complexity.While SVM excels in classification tasks, they may be computationally intense; ANN, on the other hand, offers a way to handle more complex and diverse data, especially when intricate interactions among multiple variables need to be captured.
In a study by Gao et al., [103] they introduced a ML-based technique that trains ANN using capacitance patterns induced by the user's finger touches.Through this technique, they achieved high touch angle classification accuracy (95.7%) and high average force detection accuracy (90%).The fully connected ANN comprises layers embedded with multiple nodes, including an input layer, an output layer, and multiple hidden layers (Figure 8a).ANN plays a crucial role in differentiating various objects during the grasping process in soft robotic hands, especially at different stages of preprocessing and dataset preparation.Through a series of meticulous preprocessing steps, electrical signals collected from multiple channels are transformed into data usable for classifying touch location and force levels.The fully connected ANN used for these two classifications demonstrated extreme flexibility and accuracy.
In another study by Gao et al., [104] they presented a ML technique targeted at piezoelectric force touch panels, utilizing ANN to address the instability in force-voltage responses due to varying touch directions.The research emphasized the high detection accuracy of force sensing based on piezoelectricity in interactive displays.The article highlighted that ANN, as a successful and widely used ML algorithm, shows potential as a powerful tool in providing acceptable position detection accuracy (95.2%) and force level identification (93.3%).The study also optimized ANN specifically for piezoelectric-based touch panels, experimentally enabling conventional smart devices to achieve higher than 97% accuracy in position and force level detection at low computational costs, thereby enhancing user experience.This demonstrates that ANN not only possesses high levels of flexibility and adaptability but also holds immense application potential in the fields of human-machine interaction and touch sensing.Overall, these studies collectively underscore the significant value of ML, particularly ANN, in modern interactive displays and touch panel technologies.

CNN
CNN and traditional ANN are both variants of artificial neural networks, but CNN is specifically designed to handle data with grid structures, such as images. [105]Unlike ANN, CNN employs convolutional layers to automatically and adaptively learn local features, rather than relying on fully connected layers. [106]his ability for local feature extraction allows CNN to have lower computational complexity and higher recognition accuracy when dealing with high-dimensional data such as images, videos, and time series. [107]Therefore, CNN generally outperforms in tasks like image recognition, [108] natural language processing, [109] and other feature extraction tasks [110] from complex data.CNN is particularly well-suited for pattern recognition, image recognition, object detection, and natural language processing tasks.Compared to traditional fully connected networks, CNN can more effectively handle a large volume of input data, thereby reducing computational complexity and improving model performance. [111]n a study by Chang et al., [112] an ultrasonic touch sensing system was proposed that utilizes Lamb waves and CNN for tactile position sensing.The system employs a thin steel plate equipped with multiple piezoelectric sensors to generate and detect acoustic waves.Upon touch, the system's waveforms are disturbed due to absorption and diffraction.These changes are analyzed by a CNN to predict the touch location with an accuracy exceeding 95%, a spatial resolution of 1 Â 1 cm 2 , and a response time of 60 milliseconds.CNN plays a critical role in this touch-sensing system, analyzing waveform data collected by multiple piezoelectric transducers.When the steel plate is touched, waveform signals are disturbed due to the effects of wave absorption and diffraction.These disturbances cause corresponding changes in the output signal waveforms.Figure 8c-e further elaborates on the construction and working principle of this system, including the thin plate structure of the touch panel, the piezoelectric transducers used for emitting and receiving Lamb waves, and the CNN model architecture used for position sensing.CNN's importance in handling complex sensor data, especially in high-precision and quickresponse touch sensing applications, is highlighted.
In their research, Song et al. [113] also employed CNN as a key data analysis tool.In a study by Guarino et al., [114] a transfer learning strategy was employed to create image-based representations of touch gestures performed on mobile devices, which were then used to train a CNN model.They found that this approach could achieve an accuracy rate of up to 94% in gender recognition and even 99% in age group identification.Specifically, they emphasized the effectiveness of the "Scroll" gesture in identifying gender and age groups, with accuracy rates reaching 81% and 96%, respectively.
In multiple application scenarios involving touch sensing and user identification, CNN has proven to be a highly effective and flexible tool.From high-precision identification of different touch patterns like stroking and patting to touch position sensing, and to identification of soft biometric traits like gender and age groups through transfer learning strategies and CNN, these studies collectively showcase the immense potential of CNN in handling complex touch sensing data.

k-NN
While traditional ML models like SVM, ANN, and CNN have excelled in applications involving flexible touch panels, other algorithms like k-NN also hold potential value. [115]The k-NN algorithm is a simple and intuitive ML technique that classifies an object based on the majority class of its "k"' closest neighbors in the feature space. [116]It is highly versatile, being useful for both classification and regression problems, and does not make any underlying assumptions about the distribution of the data.However, the algorithm can be computationally expensive and slow, especially when dealing with large datasets, as it requires storing all training samples and calculating distances between data points during the prediction phase. [117]gure 8. Application of multiple ML models in flexible touch panels.a) Conceptual depiction of multifunctional stacking.b) Schematic of a fully connected ANN.Reproduced with permission. [103]Copyright 2019, Multidisciplinary Digital Publishing Institute.c) Schematic of a Lamb wave ultrasonic touch screen based on a thin plate as the touch panel and disk-shaped piezoelectric transducers as wave emitters and receivers.d) Finger touch at position coordinates (3,5) on the steel plate.e) CNN model architecture for Lamb wave touch screen position sensing.Reproduced with permission. [112]opyright 2020, Multidisciplinary Digital Publishing Institute.f ) Schematic and optical image of an adjustable RFTA.g) Comparison of accuracy for object grasping recognition among various ML methods.h) Schematic of the Sing-ANN classifier and its confusion matrix; training and testing sets consist of 175 independent data groups, with an accuracy of 92.6%.i) Schematic of the Pre-KNN classifier and its confusion matrix; training and testing sets consist of 175 independent data groups, with an accuracy of 93.1%.Reproduced with permission. [118]Copyright 2023, Wiley-VCH.
In a report by Xu et al., [118] a flexible RF-resonator-based tactile array (RFTA), as shown in Figure 8f, was introduced.This array can accurately identify tactile positions through ML and differentiate between various objects during the grasping process by soft robotic hands.The device has shown immense application potential in flexible mobile scenarios and highly integrated tactile sensing systems for soft robots.For example, Figure 8g shows a comparison of the accuracy of various ML methods in objectgrasping identification tasks, which includes the k-NN algorithm as well.Additionally, the study also conducted a comparison of the accuracy of ML methods in object-grasping identification tasks.The research showed that using a single-ANN (Sing-ANN) classifier achieved an accuracy rate of 92.6% on independent sets of 175 training and testing data (as shown in Figure 8h).Meanwhile, utilizing a k-NN with preprocessing (Pre-KNN) classifier achieved an accuracy rate of 93.1% on the same independent sets of 175 training and testing data (as shown in Figure 8i).
In a study by Sagbaş et al., [119] the authors developed a real-time pressure detection system based on smartphone soft keyboard input behavior.Utilizing various sensors on the smartphone, including linear acceleration, gravity, gyroscope, and touch panel, they extracted 172 attributes from the raw sensor data.However, high-dimensional data can negatively impact the performance of ML algorithms.To address this issue, they employed feature subset optimization based on genetic algorithms and ultimately achieved an 89.61% classification accuracy using the k-NN algorithm.
These studies underscore the efficacy and flexibility of the k-NN algorithm in handling touch-sensing data, whether in real-time pressure detection [120] or tactile position identification.Concurrently, these studies also highlight the pivotal role of flexible touch panels and other advanced sensors in these applications.Particularly, in scenarios requiring high accuracy and rapid response, such as medical systems [121] and soft robots, the combination of these technologies offers significant advantages.Not only do these studies further propel the applications of tactile sensing and ML in soft robots and wearable electronic devices, but they also provide valuable directions and data foundations for subsequent research.

Challenges and Prospects
In this literature review, we have offered a holistic examination of the evolution and current advancements in flexible touch panel technologies, tracing their journey from rudimentary singlepoint systems to today's highly sophisticated grid-free touch interfaces.We have also delved into the working principles and mechanisms, like capacitive, resistive, piezoelectric, and triboelectric technologies, that underpin these advanced touch panel systems.Importantly, we have detailed how ML algorithms, such as SVM, ANN, CNN, and k-NN, have been successfully integrated into these touch panels to elevate user experience and broaden their scope of applications.The selection of these ML algorithms is based on their specific strengths and suitability for different aspects of flexible touch panel functionality.ML, as a whole, serves to enhance the sensitivity, accuracy, and versatility of touch panels, making them more intuitive and responsive to user interactions.This integration of ML algorithms with touch panel technology leads to more advanced, reliable, and userfriendly interfaces.
Despite the remarkable progress, there are significant challenges to be addressed.The emergence of multidimensional data and high-velocity data streams [122] poses complexity issues that demand more from ML algorithms, especially in real-time applications.This complexity further accentuates the need for system robustness and reliability, particularly in medical and securitysensitive settings.Concurrently, as more personal information is fed into ML models, maintaining user privacy and data security becomes an increasingly prominent concern.Popular algorithms like SVM, ANN, CNN, and k-NN also come with their own set of limitations, such as high computational complexity and sensitivity to data quality, necessitating further algorithmic refinement and optimization.
As shown in Figure 9, in the era of rapid technological advancements, the fusion of flexible touch panels and ML promises to revolutionize multiple sectors.Flexible touch panels, with their ability to offer more human-centric, customizable user interfaces, could make daily displays much more intuitive and engaging.When applied to computer devices, these panels not only enhance portability but also add a layer of flexibility that traditional rigid screens cannot offer.In the medical field, the incorporation of flexible touch panels can lead to more intuitive and highly sensitive operational interfaces, thereby enabling healthcare professionals to interact with devices more efficiently.Similarly, the automotive industry stands to gain immensely, as these panels could serve as the new standard for in-car infotainment systems and control panels in new energy vehicles.On the other hand, ML algorithms can tailor content recommendations on daily displays, boost performance and security in computer devices, and offer more accurate diagnoses and treatment suggestions in medical devices.These algorithms can also optimize energy efficiency and driving experiences in new energy vehicles and offer health monitoring and personalized settings in wearable devices.Combining flexible touch panels with ML can lead to major improvements in customization, efficiency, and user engagement.This integration promises to transform our interaction with digital devices and open the door to new, innovative applications.As sensor technologies and ML algorithms advance, we can expect even more exciting developments in this area.Whether in human-machine interaction displays or applications related to soft robotics and mobility, each research endeavor demonstrates high levels of accuracy and reliability, further propelling the development of tactile sensing technologies.

Figure 1 .
Figure 1.Overview of flexible touch panels.

Figure 2 .
Figure 2. Evolutionary types of flexible touch panel technologies.a) Switch-type touch panel controlling the on/off state of a thermometer.Reproduced with permission.[39]Copyright 2022, Elsevier.b) Array-type touch panel used for displaying touch contours.Reproduced with permission.[45]Copyright 2020, Elsevier.c) Grid-free touch panel employed for font writing.Reproduced with permission.[51]Copyright 2021, Elsevier.

Figure 4 .
Figure 4. Resistive touch panels.a) Equivalent circuit of a high-precision ultrathin transparent flexible touch panel, a photograph showing touch position detection, and demonstration of various robot control actions.Reproduced with permission.[64]Copyright 2022, Wiley-VCH.b) Schematic diagram of a high-precision flexible transparent touch panel and photograph of the writing operation.Reproduced with permission.[67]Copyright 2016, Springer Nature.c) Structural diagram of a high-precision flexible transparent touch panel, photographs of handwriting, and drawing imitation.Reproduced with permission.[68]Copyright 2017, American Chemical Society.

Figure 5 .
Figure 5. Piezoelectric touch panels.a) Structure of a switch-type piezoelectric touch panel and demonstration of a vibration haptic actuator.Reproduced with permission.[73]Copyright 2023, Multidisciplinary Digital Publishing Institute.b) Array-type piezoelectric touch panel with color-changing capabilities under different voltage levels.Reproduced with permission.[74]Copyright 2022, Elsevier.c) Schematic of data acquisition in a high-precision piezoelectric touch panel.Reproduced with permission.[75]Copyright 2020, Elsevier.

Figure 6 .
Figure 6.Triboelectric touch panels.a) Array-type transparent touch panel used for game character control.Reproduced with permission.[80]Copyright 2023, Elsevier.b) Demonstration of pattern recognition and corresponding electrical response of the array-type touch panel.Reproduced with permission.[81]Copyright 2020, Springer Nature.
Unlike Chang et al., who used CNN and Lamb waves for touch position sensing, Song et al. took it a step further by constructing a hybrid model based on CNN and long shortterm memory networks to identify and classify different contact patterns on the touch sensor, such as stroking, patting, kneading, and scratching.This hybrid model demonstrated extremely high accuracy in experiments, with the lowest reaching 99.07%.This result once again confirms the powerful capabilities of CNN in handling tactile sensing data.

Figure 9 .
Figure 9. Prospects of flexible touch panels and ML in various fields.