The human brain possesses unparalleled cognitive capabilities, occupies a small footprint (roughly the size of a football), and yet consumes very little power (approximately 20 W). Neuromorphic computing seeks to emulate the structure and functions of the human brain using electronic counterparts, thus replicating its area- and energy-efficiency1-5. The human brain is a biological neural network (BioNN) composed of neurons, dendrites, and synapses, which has inspired the development of artificial neural networks (ANNs) that have had transformative impacts on computer vision, speech recognition6, and bioinformatics7. However, ANNs currently run mostly on digital hardware, which suffers from the slowdown of Moore's law and the von Neumann bottleneck. Emerging non-volatile (or long-term memory, LTM) memories and their crossbar integration have shown advantages as hardware synapses8-15. Nevertheless, the performance of ANNs still lags behind that of the brain, because the brain's computing power largely relies on the complex chemical cascades that occur within neuron cells16. This necessitates a more faithful emulation of the structure and dynamic behaviors of neurons to unleash the throughput and efficiency of neuromorphic computing17,18.
To address this grand challenge, there is a tremendous effort underway to devise new building blocks for neuromorphic systems. For instance, basic neural functions, such as integration-and-fire, have been implemented on short-term memory (STM) devices19-23. In BioNN, a neuron interacts with over 1000 adjacent neurons through its dendrites, processing massive spatiotemporal signals through its dendrite24,25. This dendritic computing paradigm endows the neuron with complex behaviors and computing power26,27. For example, there is evidence that dendritic computing is responsible for the sound localization capabilities of the brain28. Therefore, to emulate neural computing with new fidelity, there has been a recent surge of interest in devices that have both STM and multi-dendritic structures9,29.
Neurotransistors are among the devices that have demonstrated important computational functions of neurons19,20,22,23. For example, a neurotransistor derived from a proton-conducting graphene oxide electric-double-layer transistor (EDLT), with a footprint of 80 µm × 240 µm (channel length × channel width) and a read power of tens of nW, mimicked the dendrite integration, orientation tuning, and gain control of neurons19. Moreover, a multi-terminal neurotransistor that can emulate the dendritic discrimination of neurons for different spatiotemporal signals was also developed from EDLT, which employs coplanar source-drain electrodes and gate and has a closest gate-to-channel distance of 565 µm20. The energy consumption of this neurotransistor in response to a single gate voltage pulse is approximately 1 nJ. However, these three-terminal or multi-terminal devices, which emulate how neurons compute, suffer from a large footprint, high energy consumption, and poor scalability, which significantly undermine their advantages in area and energy efficiency.
In this study, we employed material and structure engineering to develop a vertical EGT (V-EGT) based neurotransistor that simultaneously exhibits STM, a compact footprint (with a channel length of 30 nm), low energy consumption (read power ~3.16 fW, read energy ~30 fJ), and dual gates, making it an ideal choice for hardware implementation of dendritic computing, such as dendrite integration, digital and analog coincidence detection. Moreover, by integrating neurotransistors into a prototypical hardware BioNN, we demonstrated the ability to emulate the sound localization function of the brain, including sound azimuth and distance recognition. Our small size, low power neurotransistor, and the proof-of-concept neural network for sound localization highlight their potential in developing neuromorphic systems that can achieve the energy-efficiency of the brain.
Figure 1a depicts the schematic of the fabrication process flow of V-EGT, which comprises three main steps. The first step (i) involves electrode/spacer/electrode/spacer stack deposition, the second step (ii) is one-step etching to the substrate, and the third step (iii) is channel/electrolyte/gate stack deposition. Thanks to the exposed vertical sidewall, constructing multi-gate V-EGTs is easily achievable by depositing multiple channel/electrolyte/gate stacks side-by-side along the vertical sidewall (the right half of Fig. 1a iii), which is a challenging task for planar EGTs. Furthermore, this V-EGT fabrication process enables an ultra-short-channel without significantly increasing the fabrication complexity compared to planar EGT (Supplementary Fig. 1). Our V-EGTs are designed to be compatible with CMOS processes, making them particularly suitable for large-scale, low-cost fabrication (Supplementary Fig. 2). This is due to the absence of organic/2D channels and liquid/gel-like electrolytes.
The power consumption bottleneck in current neurotransistors is caused by the relatively high conductivity of semiconductor channel materials (such as p-Si30, ITO31, IZO32, IGZO33, ZnO34, and In2O335) or lithium-ion battery electrode materials (such as LiCoO236,37 and LixTiO238). In our previous work, we demonstrated that the simple binary metal oxide Nb2O539,40 is a promising solution to meet the requirement of high channel resistivity. Additionally, the stable thermodynamic properties of Nb2O5 also contribute to stable and reliable electrical operations41. It should be noted that although VO242, SrCoOx43, and SmNiO344 meet the high channel resistivity requirement, they are less favorable than Nb2O5 because VO2 is toxic, and SrCoOx and SmNiO3 are complex in their compositional elements. To achieve all-solid-state neurotransistors, LixSiO2 was utilized as the solid electrolyte, which can also be conveniently deposited by physical vapor deposition (PVD). Although there exists an incompatibility issue between lithium electrolyte materials and CMOS technology, the industry is actively exploring solutions to this challenge45.
The device structure was examined using cross-sectional transmission electron microscopy (TEM), confirming the vertical structure of our device (Fig. 1b). In contrast to the single-layer vertical device depicted in Fig. 1a, a total of four layers of electrode/spacer bilayer were deposited in this work, essentially creating a bilayer V-EGT (Supplementary Fig. 3). The spacer and electrode have thicknesses of 30 and 20 nm, respectively, resulting in a minimum channel length of 30 nm. This is a key feature that highlights the main difference between V-EGTs and planar EGTs. Specifically, V-EGTs address the issue of device scaling that has hindered planar EGTs by ingeniously defining the channel length of EGTs with the thickness of thin films. Furthermore, the vertical stackability of V-EGTs, made possible by the easy cycling deposition of electrode and spacer, endows them with the merit of variable channel length, which will be discussed in detail later. The multi-gate nature of V-EGTs also provides an opportunity to implement various functionalities that exist in the brain, further enhancing the superiority of V-EGTs over planar EGTs.
Elemental mapping (Fig. 1c) corresponding to Fig. 1b confirms the presence and spatial distribution of the Nb2O5 channel, LixSiO2 electrolyte, TiN source-drain, Ti/Au gate, and SiO2 spacer, revealing the sharp boundaries between them (Fig. 1c and Supplementary Fig. 4). The zoomed-in high-resolution cross-sectional TEM image illustrates the vertical structure of a single V-EGT (Fig. 1d). Unlike planar EGTs with exposed source/drain electrodes, the 3D stacked TiN electrodes of different layers in the TEM are interfaced with an array of non-overlapping probing pads for electrical testing. In addition to minimizing the EGT channel length by adopting a vertical structure, the channel width of the EGT was also scaled down as much as possible, resulting in the smallest EGT (channel length × channel width = 30 nm × 2 µm = 0.06 µm2). An elemental line scan along the V-EGT channel/electrolyte/gate stack direction (dashed line in Fig. 1d) indicates that our V-EGT has a channel thickness of less than 10 nm and an electrolyte thickness of less than 20 nm (Fig. 1f). Therefore, our V-EGT not only features a significantly reduced horizontal dimension (down to 30 nm) but also a vertical dimension (down to 20 nm).
To highlight the advantages of our V-EGT, we have compared it with recently reported V-EGTs (Supplementary Table 1)46-50. Our V-EGTs demonstrate three key advances. Firstly, we have optimized the etching process of the vertical sidewalls, resulting in steeper and smoother sidewalls (Fig. 1d). Secondly, we have demonstrated dual-gate V-EGTs. Lastly and most importantly, our V-EGTs employ different types of electrolytes, enabling STM electrical characteristics. The compact dual-gate V-EGT, with STM and ultra-low power, is an ideal building block for high area- and energy-efficient neuromorphic systems.
Like planar EGTs, V-EGTs are also operated by biasing both the gate and drain while grounding the source (Fig. 2a,b). Furthermore, both V-EGTs and planar EGTs rely on the migration of electrolyte ions, specifically the hybrid electric double layer and ion intercalation/deintercalation mechanism. This unique feature of EGT (Supplementary Fig. 5)51 is also the origin of the observed coexistence of STM and LTM in our case, with the former dominating (Fig. 2c and Supplementary Fig. 6). Regarding the obvious STM of our V-EGT, Li's analysis on the correlation between the retention and device size of Li-ion-based EGTs (Li-EGTs) can be applied. According to Li's analysis, the retention (STM) of Li-EGTs decreases (increases) with decreasing channel area52. Specifically, the retention of EGT depends on the discharge speed of EGT in the gate-source circuit after experiencing a gate pulse (τRC = R × C) from a circuit perspective. Here, R comprises the EGT electrolyte resistance and the external resistance in series with EGT, and C contains the gate and channel capacitance of EGT (Fig. 2d and Supplementary Fig. 7)53. For Li-EGT that is in series with an electronic switch (1S1E), the R in the discharge circuit is mainly determined by the OFF state resistance of the electronic switch, which does not change with the size of Li-EGT. However, the C of the discharge circuit decreases with the reduction of Li-EGT size. Therefore, the retention of 1S1E decreases with decreasing channel area (Fig. 2e). Although this situation applies to the retention of 1S1E or Li-EGT in open circuits, it remains true for our case or Li-EGT in short circuits. This is because the short-circuit condition has a smaller discharge load resistance than the open-circuit condition (which only contains the electrolyte resistance of Li-EGT and no external resistance), resulting in a faster discharge speed and more obvious STM.
Benefiting from the stackability of our V-EGTs, the channel length can be flexibly tuned. For instance, EGTs with channel lengths of 30 nm (using the first and second layers of TiN electrodes), 80 nm (using the first and third layers of TiN electrodes), and 130 nm (using the first and fourth layers of TiN electrodes) were demonstrated by selecting different source-drain electrodes. All of them exhibited typical electrolyte-gated behaviors (Supplementary Fig. 8). We also examined the symmetry of the source-drain electrodes and observed that the transfer characteristics of different source-drain electrode combinations were consistent, without any impact from vertical sidewall tilt (Supplementary Fig. 9).
Figures 2f and 2g display the current-time response of V-EGT under pulses of varying amplitudes and widths. Regardless of the amplitude or width of the gate pulse, the channel current of V-EGT decays over time after the pulse excitation, indicating the presence of STM. Furthermore, as the pulse amplitude or width increases, the device exhibits a gradually increasing residual channel current. We quantified the change in conductance as ((G1-G0)/ G0) × 100%, where G0 represents the channel conductance before the pulse application and G1 is either the channel conductance right after the pulse application or after 30 seconds, corresponding to the STM and LTM, respectively. Figures 2h and 2i show the conductance change ratios of V-EGT as a function of the write voltage pulse amplitude (Vw) and width (tw), respectively. When Vw ≤ 3 V (tw = 1 ms) or tw < 1 ms (Vw = 3.5 V), the LTM conductance change ratio is 0. As the amplitude and width of the write voltage pulse increase, the LTM becomes increasingly obvious. To enable the V-EGT to work as an artificial neuron, we engineered the amplitude and width of the gate voltage pulses to produce nearly complete STM.
The energy consumption of EGTs during write and read operations is a crucial factor to consider. Since the gate-source path of EGTs is capacitive in nature, the read operation dominates the energy consumption54. While the read operation of EGTs typically involves applying a DC voltage to the drain, resulting in a finite read energy consumption being unavailable, the read power (PR=VR × IR) is a more meaningful metric to evaluate EGT performance.
Planar EGTs have a larger channel length (Supplementary Fig. 10), leading to a higher read voltage and, consequently, higher read power consumption. Conversely, V-EGTs can significantly reduce their read voltage and read power consumption because the channel length is no longer limited by lithography technology but instead by material thickness (Fig. 3a). With the channel thickness, width, and current remaining constant, VD is calculated to decrease linearly with the decrease in channel length (Fig. 3b). Therefore, V-EGTs can achieve a substantial reduction in their read voltage, and subsequently, their read power consumption. In this study, we demonstrated an ultra-low read voltage of 0.1 mV for a 30 nm channel length V-EGT (Fig. 3c). Moreover, the channel current was found to be significantly larger than the gate leakage at VD = 0.1 mV (~150 pA Vs. ~50 pA), thereby validating the effectiveness of the 0.1 mV read voltage (Fig. 3d). We also compared the ultra-low read voltage of our V-EGT with other EGTs (Supplementary Fig. 11). While some EGTs have achieved read voltages of 0.1 or even 0.01 mV, they consist of organic channel and liquid electrolyte, posing difficulties in large-scale manufacturing. In contrast, our V-EGTs not only offer easy and cost-effective fabrication at a large-scale but also possess the lowest read voltage (0.1 mV) among all inorganic all-solid-state EGTs54-63. Furthermore, reducing the gate width of the EGT can further decrease the read current of the device, thereby reducing the read power consumption (Fig. 3e and Supplementary Fig. 12).
To calculate the read power of V-EGTs, we characterized the current-time response of our devices upon voltage pulses at different read voltages. The curve corresponding to the 1 mV read voltage is clean and free of fluctuation (Fig. 3f). Furthermore, we evaluated the read power of the device (Fig. 3g). Thanks to the small read voltage and the resulting small read current, we achieved a minimum measured read power of 110 fW (Vw (4 V, 10 ms), Vr (0.5 mV)). Moreover, the use of narrower write voltage pulses and smaller read voltages leads to a read power of the V-EGT in the fW-level, specifically ~ 3.16 fW (Supplementary Fig. 13). To the best of our knowledge, such a read power is the lowest among various EGTs (Fig. 3h). In addition to analyzing the read power, we have also computed and compared the read energy consumption of our devices with that of others. Initially, we formulated an equation to calculate the energy of both the STM and LTM EGTs. For STM EGTs, we estimated the time for calculating the read energy consumption by measuring the time required for the decay of STM EGTs from the maximum channel current to the initial current after undergoing a pulse. To ensure a fair comparison with STM EGTs, we used the write-read delay of LTM EGTs as the time for calculating read energy consumption (refer to Supplementary Note 1 for further details). We then computed and compared the read energy consumption of various EGTs (Supplementary Table 1 and Fig. 3i). Despite the relatively long decay time of our V-EGT (~10 s), we still achieved a read energy consumption of approximately 30 fJ due to the ultra-low read power, which is comparable to the energy consumption of biology (1-10 fJ). By expediting the decay process of V-EGT (refer to Supplementary Note 2 for further details), it is possible to decrease the decay time and hence the read energy consumption of the V-EGT. With a projected decay time of 10 ms, the V-EGT could potentially achieve a read energy of approximately 30 aJ (Fig. 3i).
In addition to low energy consumption, the vertical structure of V-EGTs also offers inherent advantages in device density. Compared with planar EGTs, V-EGTs boast a 2.5-fold increase in device area (4F2 vs. >10F2, where F refers to the feature size)49. Furthermore, when the vertical stackability of V-EGTs is exploited in the future, their dimensions could be further reduced by a factor of N, where N represents the number of V-EGT layers64. It is important to note that the significantly decreased channel length of V-EGTs will not only result in a substantial reduction in the read voltage, but also facilitate the potential enhancement of the write voltage and speed of the EGT. Through an analysis of the current composition of EGT and the interdependence between them65, it was discovered that the electrolyte thickness has the ability to synergistically scale with the channel length of EGT (Supplementary Fig. 14a-c). Thanks to the significantly reduced channel length of V-EGT, the electrolyte thickness of V-EGT is thinner than that of planar EGTs (~20 nm vs. ~50 nm for planar EGTs), which consequently improves the write voltage and speed (~5-fold increase compared to planar EGTs) (Supplementary Fig. 14d,e). Ultimately, the overall scaling down of V-EGT (including the channel length and width, as well as the electrolyte thickness) results in concurrent low write and read voltages (Supplementary Fig. 15).
The compact size and multi-gate structure, along with the features of STM and low energy consumption, make our V-EGTs highly suitable for use as high-density and low-power artificial neurons. In this regard, we first demonstrate that V-EGTs and biological neurons exhibit similar relationships that govern the input-output signals by using a single dual-gate V-EGT (Fig. 4a). In this setup, the gate (drain) of a dual-gate V-EGT is considered as the dendrite (axon) of a neuron, and our device can mimic signal propagation pathways in the brain. Furthermore, the short-term response of the channel current to gate voltage pulses and the integration function of the drain current to multiple gate voltage pulses reflect the short-term information processing capability of the neuron and the dendrite's spatial integration of multiple input signals. Figure 4b illustrates the electrical testing setup for the dual-gate V-EGT, which is used to characterize the artificial neuron properties of our device. Firstly, we examined the dendrite integration of our dual-gate V-EGT neurons (Fig. 4c). Since the channel current controlled by each gate is independent, the resulting dendrite integration is linear, i.e., Ids = Ids1 + Ids2. It should be noted that although the dendrite integration of our dual-gate V-EGT artificial neurons lacks the nonlinearity observed in BioNNs, this linear integration mechanism is well-suited for scenarios requiring precise computations, such as logical operations. In biology, different dendrites of a neuron form distinct synaptic weights with neighboring neurons. However, for simplicity, we assume here that each dendrite of a neuron shares the same signal-receiving capacity. Correspondingly, each gate of the dual-gate V-EGT should also have equal ability to regulate the channel current. To verify this, we investigated the consistency between different gates. Owing to the highly uniform etching and thin-film deposition processes, the electrical properties of different gates in the dual-gate V-EGT exhibited good consistency when each gate was controlled individually (Supplementary Fig. 16).
In addition to realizing the dendrite integration function of neurons, our dual-gate V-EGT can also perform AND logic by utilizing the difference between the drain current when the two gates act simultaneously and separately. This functionality enables the realization of the coincidence detection function of neurons. To achieve this, a current threshold Ith is set, and the drain current is only larger than Ith when both gates act simultaneously. Vice versa, the current is smaller when the gates act separately, which implements the AND logic, as illustrated in Fig. 4e. Interestingly, the peak channel current varies with the time interval between voltage pulses applied to gate 1 and 2, gradually decreasing with the increase of the time interval. Here, we use high1, high2, and high3 to represent the peak of channel currents corresponding to t = 0, 1, and 7 s, respectively, where high1>high2>high3. This characteristic can be utilized for coincidence detection by biological neurons in response to two events. For biological neurons, receiving two input signals simultaneously results in the largest output signal, while receiving them separately results in a weaker output signal. In order to improve the coincidence detection function of our dual-gate V-EGT artificial neurons and apply it to sound distance detection (which will be discussed later), we have expanded the definition of neuronal coincidence detection by adopting a new physical quantity to benchmark the degree of coincidence of two events (Fig. 4f). Specifically, we define the interval between two events as Δt, where Δt can be any real number. With the dual-gate paired-pulse facilitation (PPF) characteristics of our device, coincidence detection can be achieved in other situations (e.g. Δt = 0.1, 0.5, and 2 s). To differentiate it from the PPF of a single-gate EGT, we refer to the PPF of dual-gate V-EGT as dual-gate PPF, and an example waveform of dual-gate PPF is shown in the inset. The intensity of such coincidence detection can change continuously (non-linear decay) (Fig. 4f), which is crucial for sound distance detection.
To better demonstrate the coincidence detection capabilities of our device for sound distance detection, we present the A2/A1 values obtained from a series of input pulses with different time intervals (Δt = 0.1, 0.5, 2, and 7 s) in Fig. 4g. By defining a coincidence threshold of Δt = 0.1, 0.5, and 2 s respectively, and a non-coincidence threshold of Δt = 7 s, we were able to observe coincidence detection similar to that observed in biological neurons, as indicated by the vertical trend in Fig. 4g. Specifically, we observed a significantly larger coincidence degree at Δt = 0.1, 0.5, or 2 s compared to that at Δt = 7 s, which is indicative of digital coincidence detection. In addition, by varying the criteria of coincidence detection from Δt = 0.1 s to Δt = 0.5 s and Δt = 2 s along the horizontal direction of Fig. 4g, we observed a gradual decrease in the A2/A1 coincidence degree for the corresponding coincidence detection, which is a manifestation of analog computing-based coincidence detection.
The neuronal properties of our device facilitate the reemergence of complex brain functions. Sound localization serves as an example. By defining the front of the human eye as the reference direction, the angle between the propagation direction of sound and the normal is referred to as the azimuth of sound. When sound is emitted, the presence of the azimuth generates a sound path difference, causing the left and right ears to perceive the sound at different times (interaural time difference, ITD) (Fig. 5a). As ITD and sound path difference have a one-to-one correspondence, the brain can determine the sound azimuth through ITD (Fig. 5b).
To replicate the sound localization abilities of the brain accurately, it is essential to predict the distance of the sound source in addition to recognizing the sound azimuth. For instance, consider the human brain's ability to estimate the distance of a truck driving on the road ahead. As the intensity of sound decreases proportionally with the square of the distance, the brain can determine the distance of the truck based on the intensity of sound it receives (Fig. 5c). When the truck is to the listener's left and moves from left to right, the listener perceives a gradual increase in the volume of the truck's honk. As the truck moves towards the right and away from the listener, the sound of the honk gradually diminishes (Fig. 5d).
To distinguish the ITD of sound and identify its azimuth and distance, we constructed a neural network (Fig. 5e). The blue spheres in the figure represent neurons that sense sound in the left and right ears, respectively. These neurons transmit sound information to post-sound information processing neurons (green or yellow spheres) through axons (blue lines), synapses (blue nodes), and dendrites (green or yellow lines). The sound signal is assumed to propagate through the network in the form of spikes, where the amplitude of spikes is modulated by synaptic weight. By configuring the synaptic weight matrix between the pre-neurons and two green post-neurons into a diagonal matrix, the outputs of the two post-neurons can be different. The difference between the outputs of the two neurons varies with the ITD and is symmetric with respect to the ITD (specifically, the output of neuron 1 (3) at ITD = t is the same as the output of neuron 3 (1) at ITD = -t). The network can recognize sound azimuth accordingly (see Supplementary Note 4 for a more detailed working principle analysis). The post-neuron situated in the center of the three post-neurons exhibits equivalent connection strength to the two neurons that sense sound in the left and right ears, as shown in Fig. 4 while introducing the coincidence detection function of neurons. As the distance between the truck and the listener increases, the intensity of the honking sound heard by the listener diminishes. This precisely corresponds to the fact that a larger ITD results in a smaller coincidence degree A2/A1. Therefore, this neuron functions as a detector of sound distance. Notably, the ITD and the travel time of the truck's honk differ, and A2/A1 is not inversely proportional to the square of ITD. Nonetheless, a clear mathematical relationship between them exists, as depicted in Supplementary Fig. 17. Hence, it is possible to determine the distance to the sound source based on our neurons' coincidence detection with an analog coincidence degree A2/A1 (refer to Supplementary Note 5 for further details).
In accordance with Fig. 5e, a hardware neural network was constructed (Fig. 5f). To emphasize the sound localization capability of our network, we neglected the section that emulates human ears to convert sound signals into electrical signals. Instead, we stimulated the output signals of sound sensing neurons with electrical signals that have different time intervals. Additionally, we substituted the modulation of synaptic weights on the spike amplitude of the sound signal with the modulation of voltage pulse amplitude using voltage divider circuits while retaining the diagonal synaptic weight matrix. The feasibility and efficacy of this approach, the modulation of synaptic weight on pulse spike amplitude, is depicted in Supplementary Fig. 18. Three dual-gate V-EGTs serve as the three post-neurons, with two green post-neurons working together on sound azimuth recognition and the yellow post-neuron functioning in sound distance recognition. The gates of all post-neurons function as dendrites.
Based on the aforementioned hardware neural network, whose equivalent circuit diagram is presented in Fig. 5g, we replicated the brain's recognition of sound azimuth (Fig. 5h-j) and distance (Fig. 5k), respectively. As an illustration of sound azimuth localization according to ITD, we displayed the intensity and sequence of sound signal spikes received by each dendrite of post-neurons 1 and 3, as well as the output signal when ITD = 2 s (Fig. 5h,i). The Ipost1/Ipost3 value corresponds to a unique ITD and can be utilized to determine the sound azimuth, thus simulating the brain's sound azimuth recognition function. By altering ITD, we further demonstrated the identification of all potential azimuths (-90° - 90°) (Fig. 5j).
Aside from recognizing sound azimuth, we also demonstrated the capability of our hardware neural network to recognize sound source distance, based on the response of the yellow post-neuron to the left and right ears. By creating a symmetric dual-gate PPF of our dual-gate V-EGTs around ITD and integrating the inherent analog characteristic and nonlinear decay, our hardware neural network replicated the sound distance recognition function of the brain (Fig. 5k). Therefore, our constructed hardware neural network has the ability to simultaneously recognize both sound azimuth and distance.
The biological ITD typically falls within the range of 1 ms. However, due to the relatively long current decay time (~10 s) of our current V-EGT, our ITD for sound localization ranges from 0.1 to 10 s, which is significantly larger than its biological counterpart. Nonetheless, given that our current V-EGTs are capable of recognizing ITD within the range of 0.1 - 10 s, a future V-EGT with a decay time of 10 ms (refer to Supplementary Note 2 for further details) will be able to accurately identify ITD with smaller time differences (e.g., identify ITD with a 0.1 ms difference).