Traditional device electronics, based on downscaling digital devices for reliable logic-gate operations, have driven advancements in the Information Age. However, recent research has shifted towards utilizing the analog characteristics of devices, previously considered a drawback. Neuromorphic systems, with their innovative computing architecture designed to overcome the energy inefficiency of the von Neumann architecture, leverage these analog behaviors. Analog devices, with their numerous internal states adjustable with minimal energy consumption and long-term stability (nonvolatility), offer superior data storage and energy efficiency compared to digital devices. Advances in analog devices allow emulation of biological synapses and neurons, with experimental demonstrations of their ability to solve cognitive tasks related to learning and recognition. A key aspect of neuromorphic systems is the volatile behavior exhibited by some analog devices, where a temporal state enhancement decays to the initial state, mimicking aspects of short-term plasticity in biological synapses, such as paired-pulse facilitation. Furthermore, these volatile devices can emulate biological neurons, integrating temporal input stimuli based on the leaky integrate-and-fire (I&F) neuron model. Integrating sensor devices with volatile analog devices as sensory neurons has been proposed to mimic human sensory perception. A novel device characteristic, termed "semivolatile," allows for switchable volatile/nonvolatile behaviors, offering the potential to simulate both synaptic and neuronal functions simultaneously with a single device type. This paper introduces a tactile sensor system where both sensory neurons and the perceptual synaptic network are implemented using a semivolatile carbon nanotube (CNT) transistor. This single device type emulates the sensing receptor, action potential, sensory neuron, and synaptic network through a combination of a tactile sensor, a voltage-controlled oscillator (VCO), a neuronal CNT transistor, and an array of synaptic CNT transistors, respectively. The system converts pressure stimuli into resistance changes, then into frequency variations using the VCO. The neuronal CNT transistor processes this frequency information in volatile mode, generating a leaky-integrating output. Finally, a network of synaptic CNT transistors in nonvolatile mode performs supervised learning for pattern recognition.

The paper reviews existing research on neuromorphic systems and their use of analog devices to mimic biological neural networks. It highlights the advantages of using analog devices for energy efficiency and the ability to emulate synaptic and neuronal functions. The authors discuss previous work on volatile and nonvolatile devices and introduce the concept of semivolatile devices, which can switch between volatile and nonvolatile modes. The literature review emphasizes the lack of studies using a single device type to implement both neuronal and synaptic functions in a neuromorphic system and positions their work as a significant step toward simpler and more scalable neuromorphic computing.

The methodology section details the fabrication processes for both the tactile sensor and the semivolatile CNT transistor array. The tactile sensor is fabricated using a flexible and transparent polydimethylsiloxane (PDMS) substrate, functionalized with poly-L-lysine to enhance CNT adhesion. A metallic CNT network film is created via spray coating, and two copper electrodes are added. A top PDMS layer completes the sandwich-like structure. The semivolatile CNT transistor array is fabricated on p-doped silicon substrates with a thermally grown SiO2 layer. A local back-gate structure is formed using titanium and aluminum oxide layers as gate insulators. Poly-L-lysine is used to improve the adhesion of the semiconducting CNT network channel. Source/drain electrodes (titanium and palladium) are deposited, and photolithography and plasma etching steps isolate the individual transistors. For the crossbar array, copper and SiO2 are used for metal lines and interlayer dielectric layers, respectively. The operational principle of the semivolatile CNT transistor relies on different hole-movement mechanisms in interface and surface traps. The hysteresis in drain current (ID) with respect to gate voltage (VG) is attributed to hole trapping/detrapping. The fast tunneling of holes at the interface trap and the slow diffusion of holes at the surface traps create the semivolatile behavior. The voltage level (Vhigh) applied during a VG pulse determines the operational mode (volatile or nonvolatile). The methodology also describes the experimental setup for testing the tactile sensor system, including the custom-made software used to control signal flow. The tactile sensor converts pressure stimuli to resistance changes, which the VCO converts into frequency changes. These frequency signals are fed to the neuronal CNT transistor (operating in volatile mode) for leaky integration. The resulting output is then sampled and used for pattern recognition by the synaptic CNT transistor array (operating in nonvolatile mode). A circuit model is developed and used for quantitative analysis of the system components and to validate experimental observations.

The key findings demonstrate the successful implementation of a biorealistic tactile sensor system using a semivolatile CNT transistor. The transistor exhibits switchable volatile/nonvolatile behavior depending on the gate voltage, enabling it to function as both a neuron and a synapse. The volatile mode is demonstrated by the temporal ID enhancement and decay following a single VG pulse. The level of Vhigh determines the mode of operation, with higher Vhigh values leading to nonvolatile behavior. The volatile mode emulates neuronal functions, with the short-term ID change (ΔGST) depending on the frequency of presynaptic spikes. Higher frequencies lead to larger cumulative ID enhancement. The circuit model accurately captures the transient ID behavior. The tactile sensor effectively converts pressure stimuli into resistance changes, and the VCO converts these into frequency-modulated signals. The neuronal device successfully integrates temporally correlated tactile stimuli, with distinguishable responses to different tactile patterns. The system's ability to differentiate between patterns based on timing information is highlighted. Finally, the biorealistic perceptual learning and recognition processes demonstrate the system's capacity to classify input patterns based on the sampled ID responses.

The findings directly address the research question by demonstrating the successful creation of a bio-inspired tactile sensor using a single device type (semivolatile CNT transistor). This simplifies fabrication and paves the way for high-density integration of neuromorphic systems. The emulation of both neuronal and synaptic functions within a single device is a significant advancement in neuromorphic computing. The ability to differentiate temporal patterns based on timing information further enhances the system's functionality. The high accuracy of the circuit model provides a strong foundation for further development and optimization. The results demonstrate the potential of this technology for applications in robotics and prosthetics, where energy-efficient and adaptable sensory systems are crucial.

This research successfully demonstrated a biorealistic tactile sensor system using a semivolatile CNT transistor, effectively emulating both neuronal and synaptic functions with a single device type. This simplifies fabrication and enhances scalability for neuromorphic systems. The system's ability to learn and recognize tactile patterns based on temporal information opens doors for advanced applications in robotics and prosthetics. Future research could explore the integration of this technology with other sensory modalities and the development of more complex neuromorphic architectures.

The current study focuses on a simplified tactile pattern recognition task. Further research is needed to evaluate the system's performance with more complex and varied tactile stimuli. The size and complexity of the synaptic network might need optimization for real-world applications requiring higher resolution. The long-term stability of the semivolatile CNT transistors over extended periods requires further investigation.