Humans and other biological organisms use tactile feedback to interact with the environment (Dahiya et al., 2010). Inspired by nature, numerous research groups are harnessing the technological advances to develop artificial e-skin with features mimicking human skin (Boland, 2010; Tee et al., 2012; Bauer, 2013; Hammock et al., 2013; Wang et al., 2013; Yogeswaran et al., 2015; Núñez et al., 2017). These works find application in prosthetics, potentially to bestow lost sensory feelings to amputees (Raspopovic et al., 2014) and robotics to provide the touch sensory capability allowing them to interact physically and safely with real-world objects (Dahiya et al., 2013b). Thus far, the major focus of e-skin research has been on the development of various types of sensors (e.g., contact pressure, temperature, humidity, etc.) and their integration on large-area and flexible/conformable substrates (Dahiya and Valle, 2013; Dahiya et al., 2013a, 2015; Hammock et al., 2013; Lee et al., 2015; Yogeswaran et al., 2015; Polishchuk et al., 2016; Núñez et al., 2017). However, processing of a large amount of data from e-skin has remained a challenge, especially in the case of large area skin where number of touch sensors increase rapidly. As an example, to develop human inspired e-skin for robotic and prosthetic limbs, an estimated 45 K mechanoreceptors (MRs) will be needed in about 1.5 m² area, as shown in Figure 1 (details in Supplementary Section 1; Johansson and Vallbo, 1979; Boniol et al., 2008; Mancini et al., 2014; Goldstein and Brockmole, 2016). This number of sensors will be much higher if we consider e-skin to have equivalents of thermo-receptors and nociceptors (Goldstein and Brockmole, 2016). The large number of receptors in the skin indicates that the tactile data will multiply rapidly, and therefore one can understand the challenge associated with its compiling and processing. With the recent shift in the focus of tactile skin research in robotics from hands to whole-body tactile feedback, a need has arisen for new techniques to manage the tactile data. Currently, limited solutions are available to deal with large data generated in tactile skin, let alone for the resulting touch based perception, which is another dimension of tactile data handling. For example, in the case of prosthesis, it is important not only to collect the tactile data for critical feedback, but also to decode the user's intentions in real time (Raspopovic et al., 2014). Perhaps a neuron-like inference to handle the tactile data early on could help as indicated by a significant downstream reduction in the numbers of neurons transmitting stimuli in the early sensory pathways in humans (Barlow, 1981; Buck, 1996; Barranca et al., 2014). Research suggests that distributed computing takes place in the biological tactile sensory system (Barlow, 1981; Dahiya et al., 2010, 2015). For example, the ensemble of tactile data from peripheral neurons is considered to indicate both the contact force and its direction (Johansson and Birznieks, 2004; Johansson and Flanagan, 2009). Such distributed local processing of tactile data is advantageous in practical terms as sending reduced data to higher-perceptual level releases some pressure in terms of complex and bulky sensory hardware. Thus, the hardware implemented neuromorphic tactile data processing along with neural networks like algorithms will be helpful. Currently, the neuromorphic hardware is primarily targeted for vision and hearing related applications. Since, vision and hearing are not as distributed as tactile sensing, the neuromorphic hardware developed for them is not optimal for tactile sensing and dedicated solutions are needed. Few works on tactile sensing have used software based neural networks approaches for tasks such as object recognition via texture or materials (Decherchi et al., 2011; Kaboli et al., 2015). However, due to the lack of large-scale parallel processing, the software-programmed neural networks are slower and less energy-efficient (Ananthanarayanan et al., 2009; Misra and Saha, 2010) and hence the HNN implementations will be interesting.

The hardware neuromorphic architecture implementations reported in literature thus far are based on devices such as memistor (Widrow and Hoff, 1960), spin-logic (Sengupta et al., 2015; Grollier et al., 2016), memristor (Jo et al., 2010), neuron MOSFET (Ishii et al., 1992; Kotani et al., 1998), analog circuit based neurons (Mead and Ismail, 2012), field programmable gate array (FPGA) (Misra and Saha, 2010) and software-programmed neural networks (Cotton and Wilamowski, 2010). So far, these technologies have not been used with tactile skin. But, they could offer alternative to the υ-NWFET approach presented here—even if υ-NWFET has many inherent advantages such as possibility of printing devices on large area as discussed in the next section. The above alternative technologies have their own advantages and challenges in terms of complexity, scalability, speed, reliability, repeatability, cost, non-bendability, power consumption etc., which limit their use in the emulation of biological systems. For example, the memistor, a 3-terminal electrochemical cell element achieved limited success because of scalability issues (Widrow and Hoff, 1960). Likewise, the spintronic neurons are energy efficient (Grollier et al., 2016) but it is challenging to realize practical large-scale neuromorphic architectures and read-out. Recently, two-terminal memristive devices have gained significant attention as the state of their internal resistance could indicate the history of the voltage across and/or current through the device (Yang et al., 2013). The memristive approach is promising in terms of low-energy, high-density memories and neuromorphic computing (Courtland, 2016), but as memristors are two terminal devices it may not be possible to simultaneously execute the signal transmission (computation or reading phase) and learning functions (writing phase). Metal NWs finds application mainly as interconnects and junction elements in crossbar memristors. Use of inorganic semiconductor NWs for HNN is an interesting direction.

Addressing the need for tactile HNN in e-skin, this paper presents a novel Neural Nanowire Field Effect Transistor (υ-NWFET) structure as the functional building block. The focus of the paper is on modeling, simulation and first validation with fabricated υ-NWFET structure prior to practical realization of a large area e-skin. This paper is organized as follows: The υ-NWFET device structure, working principle related to a biological neuron, and specific advantages of using NWs for HNN are presented in the Section υ-NWFET Based Neuro-Mimicking e-skin. Various modeling and simulation tools, and device fabrication methodology are presented in Section Methods. The results of modeling, simulation and fabrication are presented in Section Results along with experiment of e-skin integrated on a 3D printed robotic/prosthetic hand. Section Discussion discusses overall implementation and study of impact of fabrication related variability over HNN performance. The results are summarized in the concluding Section Conclusions.

A simplified representation of biological and artificial neurons are shown in Figures 2A,B, respectively (McCulloch and Pitts, 1943; Goldstein and Brockmole, 2016). The υ-NWFET devices (symbol in Figure 2C and structure in Figure 5A) imitate the working of biological neuron in a simplified manner. The structure of υ-NWFET (Figure 5A) is a variant of a neuron MOSFET with NWs providing the functional channel region (Ishii et al., 1992; Shibata and Ohmi, 1992; Taube et al., 2016). The main floating gate, modulating the channel current is capacitively coupled to several gates. The overlap width of the individual gates over the floating gate determines the initial synaptic weight of the neural input on which further schemes of plasticity operates. This imitates the synaptic summation of weighted inputs in the cell body (soma) of the biological neuron or the artificial neuron. The activation function is performed at circuit level as discussed later in Section Circuit Modeling. It may be noted that the biological neural systems also have the plasticity or synaptic modulation, which reflects their ability to strengthen or weaken the synaptic weights over time. This modification of weights results in various forms of memory [namely, Sensory Memory, Short Term Memory (STM) and Long Term Memory (LTM), (Atkinson and Shiffrin, 1968)] at different hierarchical levels of the neural network. The sensory memory lasts for fraction of a second and is associated with local distributed computation involved in the tactile, smell or vision sensory system. Related to tactile perception, sensory memory plays a role in the local distributed computation such as force direction estimation, local curvature estimation, downstream reductions. Sensory memory on further hierarchical levels leads to STM which typically lasts for few seconds to a maximum of 30 s. These STMs gradually get transformed to LTMs at higher perceptual levels of neural network which can last up to lifetime. The υ-NWFETs based circuits presented here could exhibit similar behavior as discussed later in Section Circuit Modeling.

The working of υ-NWFET can be explained with modulation of the source-drain output curent by the voltage mode weighted summation of all input voltages to individual gates (V_Gwi). Voltage-mode summation through an insulating dielectric has significant advantage over current-mode summation as the standby power dissipation can ideally be avoided. The voltage in the floating gate (V_FG) is given by:

Where C_wi corresponds to the capacitance between the individual gate and the floating gate and determines the weighing factor w_i. V_Gwi corresponds to the voltage at each gate, V_Offset arises from any non-ideal charges such as fixed-oxide or interface trap charges (Shashank et al., 2011; Robinson et al., 2013).

The schematic illustration of proposed biomimetic tactile sensory neuro-system is given in Figure 2D. It consists of a prosthetic hand covered with a tactile e-skin; the simulated/measured tactile signal are read out by the receptive sensors to the input sensory layer of the υ-NWFET based neural network for sparse coding of tactile information. In bio-mimicking hardware, the sparse-coded output should comprise of encoded information such as pressure/force and temperature spatial and time distribution, force direction, local curvature, vibration, slip, humidity, comfort feeling, proprioception, reflex signals, pain, gestures. As a demonstration, in this paper, sparse coding of gesture direction has been presented based on υ-NWFET array architecture. Modeling was carried out to understand whether NW could be effective as a channel material for a neuron MOSFET as compared to the implementation with bulk material (Shibata and Ohmi, 1992). A combination of p-and n-channel υ-NWFET cascaded with a CMOS inverter has been demonstrated to represent a neuronal element. With multiple synaptic inputs and an output, the proposed structure will contribute toward building a computational architecture inspired by biological systems. An integrated hardware-realized neuromorphic tactile system could mimic or simulate biological activity and lead to unidirectional or bidirectional bio-electronic interfaces.

The viability of Si-NWs as an active material for HNN has been investigated through modeling and simulation, followed by the fabrication of first υ-NWFET device and tactile e-skin interface. The work flow of methodology used in this paper is summarized in Figure 3 and described below in detail.

The υ-NWFETs based approach for realizing HNN has several advantages for tactile data processing in electronic skin (e-skin). For example, it allows implementation of neural circuits in a compact array architecture (DeHon, 2003). With good subthreshold control of a tri-gated or gate all around NWFETs (Kuhn, 2012), it would also be possible to develop highly power efficient devices or circuits. Further, the possibility of printing NWs (Shakthivel et al., 2015; Yogeswaran et al., 2015; Navaraj et al., 2017) means with υ-NWFETs it will be possible to develop bendable or conformable systems, which is much needed for better integration of e-skin on curved surfaces such as the body of a robot or prosthetic hand (Dahiya et al., 2015). Such e-skin could have printed υ-NWFETs in the backplane (Yogeswaran et al., 2015; Shakthivel et al., 2016) to communicate with higher perceptual levels. It is possible to have 3D integration or stacking of NWs based circuit (Javey et al., 2007) and if such work is extended for e-skin then we may see more advantages, particularly, in terms of mimicking biological tissues and brain. The integration of υ-NWFET based neural processing circuits with NW-based neural recording/mapping and stimulation circuits is another direction that could significantly advance the research in neuro-prosthetics, bio-neuro interfaces and electroceuticals (Patolsky et al., 2006; Qing et al., 2010; Robinson et al., 2013; Thomson et al., 2017).

While the direction is interesting and promising, there are significant challenges associated before realizing a fully biomimicking artificial tactile skin. The tactile data processing in biological e-skin is much complicated and has complex pathways. The notion of neurons being represented as entities performing weighted summation followed by actuation itself is a significant approximation far from a real neuron, and a slightly more closer approximation uses time-domain differential equations to explain biological neuron's membrane dynamics and interaction (Marder and Taylor, 2011).

Biological neurons are highly energy efficient compared to most artificial implementation of neurons (Boahen, 2017). To achieve energy efficiency and better performance, a υ-NWFET should have lower leakage current, higher drive current and higher on-to-off ratio. High-K dielectric used as a gate dielectric to avoid gate leakage while still having lower EOT ensures better coupling and control of gates over the channel. Further, the gates formed in a trigate configuration around the NW offers better subthreshold performance. On-to-off ratio of up to 2.6 × 10⁴ was achieved for the simulated device structure with all gates ON vs. all gates OFF. For packing more neurons per unit area and improving performance further, the υ-NWFET must be scaled in all dimensions such as the width, length of the NW, the gate span of each gates.

When such intense scaling is carried out, the process variations such as variations in doping concentration, NW dimension, gate width, line edge roughness may influence the synaptic weights and performance of the neural network. However, neural networks are known to be inherently fault tolerant and robust. In the presented work, comparison of designed capacitances with experimental capacitance shows a maximum fabrication induced weight variation ~0.1%. Allowing up to 1% change in the weights of the system model lead to misclassification of the data set by only ~0.06% which shows the robustness of NN for such applications.

The various circuit approaches presented could be used as sub-components for neuro-mimicking tactile e-skin and based on system requirement, choice can be made between hardwired-neuron with no plasticity (as in Figure 7), neuron with plasticity having sensory memory or STM (as in Figure 9) or LTM (as in Figure 8) to be used at different hierarchical levels of the tactile sensing NN. Synaptic plasticity finds application both in data storage/memory as well as neural computation. In the first approach, i.e., hardwired-neurons, learning and circuit/layout synthesis will be through software tools which will be followed by practical fabrication forming a hardwired neural network. By introducing additional plastic synapse schemes (as in Figures 8, 9), the weight could be modulated over the initial value set by the capacitances. In this case the synaptic weight initially set by the capacitances could be considered as a phyletic memory (Fuster, 1997) because it is hardwired over which further schemes of plasticity operates. Such an approach could be considered as a semi-plastic neural network. For such a network, the quantization arising from layout synthesis (For example rounding-off the weight equivalent capacitance dimensions to 1 μm) followed by lithography process for fabricating the capacitors will lead to k-levels of possible discrete synaptic weights (Obradovic and Parberry, 1992; Obradović and Parberry, 1994). The approach proposed in Figure 9 results in a programmable synapse, which could be used to implement hardware-in-the-loop learning. It is to be noted that in the initial stage of tactile sensing the sensory data need to be stored only for a short time and hence a neural circuit with sensory memory is sufficient for earlier tactile data handling as in Figure 9. This circuit could be further modified for higher hierarchical levels to have STM and LTM associated plasticity. A transitionary circuit from sensory memory to STM can be achieved either by replacing the C_SM with an element of higher value (as in Figure 9C) or by replacing Q_forward with a nanoionic-like transistor (Pillai and De Souza, 2017) for use in higher hierarchical level of neural network beyond tactile skin. Beyond that, increasing C_SM may not be a practical option as it becomes bulkier in the process of realizing longer times. Possible strategies for transition from STM to LTM at different stages of the network include replacing Q_forward by a nanoionic-like transistor (Pillai and De Souza, 2017), or with a NW-based programmable floating gate transistor (Yan et al., 2011), or replacing R_forward with a memristive device (Yang et al., 2013).

The system model interfaced with a flexible and transparent touch sensitive e-skin with 6 × 6 tactile elements and tested in real-time demonstrates the working of the proposed approach. However, since the SimBrain model involves software NN programmed in java to mimic the HNN, a delay of ~1.12–1.54 s was observed per cycle for the implementation with Intel® Core™ i7-4500U CPU @ 2.4 GHz with 8GB RAM and a delay per cycle of ~0.394–0.470 s was observed for implementation in Intel® Core™ i7-4790K CPU @ 4 GHz with 32GB RAM. While this could be improved by approaches such as use of GPU, dedicated HNN such as the proposed approach will be optimal for real-time tactile data processing. The delay with software NN will be much substantial if the number of tactile elements are increased for example to a human palm ~18,675 MRs as shown in Figure 1. Further advancements in the system model is required toward advanced tactile perception tasks such as schematically shown in Figure 2D.

One of the potential application of this technology could be in an industrial task such as fruit or object sorting, where the bio-mimicking neural networks in the skin of robotic hands could classify and held objects based on the physical parameters such as pressure, temperature etc. as well as optical parameters from special optical sensors from tactile e-skin (Dahiya and Valle, 2013; Dahiya et al., 2015).

A novel υ-NWFET based approach for realizing hardware neural networks has been presented and validated through device, circuit and system-level modeling and simulation. Two different approaches, STM and LTM, have been simulated to implement the memory or neuroplasticity. Fabrication of a υ-NWFET has been carried out with a Si-NW as the channel material. The I-V characteristics of the υ-NWFET demonstrates the neuronal function of the device with synaptic weights modulating the output current. For example, for a given drain (V_DS = 4 V) and gate voltage (V_GS = 6 V), the drain current at output was reduced by ~54 times with a gate weight of 5/14 as compared to 2/14. The proposed structure is a step toward realizing flexible power-efficient bio-inspired neural sensing and circuit architectures as a backplane for tactile e-skin in robotics or prosthetics. To this end, the system model interfaced with a flexible and transparent touch sensitive e-skin (having 6 × 6 tactile elements) and tested in real-time demonstrates the working of the proposed approach. Up to 1% change in the weights of the system model lead to misclassification of the data set by <0.06% which shows the robustness of NN for tactile sensing application. In principle, the approach could be adapted for spiking neural networks and further exploration in that direction should be extremely interesting. Multilayer or deep learning hardware neural networks could be used for further additional sparse coding to enable advanced tactile perception tasks such as schematically shown in Figure 2D. Future work will include large area fabrication of the proposed e-skin system in a flexible form factor and its subsequent testing.

WT and RD conceptualized the idea. WT, FL, and VV contributed to the simulation. WT, CG, and DS contributed to the fabrication and analysis of device. All the authors were involved in the manuscript preparation. RD provided overall supervision of this research.