Ayoub H.
Jaafar
ab,
Li
Shao
c,
Peng
Dai
a,
Tongjun
Zhang
a,
Yisong
Han
d,
Richard
Beanland
d,
Neil T.
Kemp
b,
Philip N.
Bartlett
c,
Andrew L.
Hector
*c and
Ruomeng
Huang
*a
aSchool of Electronics and Computer Science, University of Southampton, Southampton, SO17 1BJ, UK. E-mail: r.huang@soton.ac.uk
bSchool of Physics and Astronomy, University of Nottingham, Nottingham, NG7 2RD, UK
cSchool of Chemistry, University of Southampton, Southampton, SO17 1BJ, UK. E-mail: a.l.hector@soton.ac.uk
dDepartment of Physics, University of Warwick, Coventry, CV4 7AL, UK
First published on 10th November 2022
Memristors are emerging as promising candidates for practical application in reservoir computing systems that are capable of temporal information processing. Here, we experimentally implement a physical reservoir computing system using resistive memristors based on three-dimensional (3D)-structured mesoporous silica (mSiO2) thin films fabricated by a low cost, fast and vacuum-free sol–gel technique. The in situ learning capability and a classification accuracy of 100% on a standard machine learning dataset are experimentally demonstrated. The volatile (temporal) resistive switching in diffusive memristors arises from the formation and subsequent spontaneous rupture of conductive filaments via diffusion of Ag species within the 3D-structured nanopores of the mSiO2 thin film. Besides volatile switching, the devices also exhibit a bipolar non-volatile resistive switching behavior when the devices are operated at a higher compliance current level. The implementation of mSiO2 thin films opens the route to fabricate a simple and low cost dynamic memristor with a temporal information process functionality, which is essential for neuromorphic computing applications.
A reservoir computing (RC) system is a framework derived from RNNs and is a suitable candidate for temporal and sequential data processing.11,14 It normally consists of massive nonlinear elements that transform the input pulses’ spatiotemporal patterns in a high-dimensional space within the reservoir, which are then analyzed by readout function using a simple learning algorithm.15 In the reservoir network, only the node states connecting the reservoir layer and the output layer are required to be read out through linear weighted sums for recognition. As a result, the reservoir itself does not need to be trained but only the output connections with linear weights do, which makes the training process simple, fast, and energy-efficient.11,16–18 The implementation of physical reservoir computing requires a physical system to act as a reservoir to produce a nonlinear and high-dimensional response to the external stimuli. To qualify as a suitable physical reservoir, the response of the system must have many degrees of freedom, and ideally the system should also allow some control parameters that may be tuned to change the response of the reservoir.12
To process temporal information, the reservoir states need to be determined by both present and past inputs within a certain period of time. Physical memristors featuring the properties of STP and nonlinear dynamics are therefore desirable for the implementation of RC systems.11 Compared with the conventional memory devices (e.g. DRAM), a filamentary-based memristor offers much higher complexity as its temporal dynamics originated from the random formation and spontaneous dissolution of the conductive filament(s) in response to external stimuli. In a resistive memristor, the STP behavior normally relies on the spontaneous dissolution of the conductive filament(s) after the withdrawal of stimuli.5 A faster dissolution can therefore be extremely advantageous to enable a faster processing time and lower power consumption.19 However, this significantly depends on the ion transfer dynamics in the dielectric layer. Different dielectric materials compositions, qualities and morphologies can all contribute to the different ionic dynamics, resulting in diverse synaptic properties.6 Realizing STP behavior normally requires careful tuning of the external programmed voltage pulse. Several memristor devices have since been applied in implementing physical RC systems using dielectric materials such as WOx,13,17 Ta2O5,20 TiOx,16 HfO2,21 and lead-free Cs3Sb2Br9 perovskite19 as the active electrolyte switching materials to enable STP behavior.
Mesoporous nanocomposite materials feature strongly in the field of energy storage due to their high surface to volume ratio as well as fast electron transport and reduced resistance of charge transfer.22 Such advantages could also be capitalized in neuromorphic switching to achieve the desired synaptic function. Through the introduction of porous structures into the electrolyte layer of a memristor, the ionic dynamics and the growth of conductive filaments can be regulated through control over the pore geometries.23,24 Several mesoporous film-based memristors have been reported to demonstrate resistive and synaptic switching behaviors.25–29 Choi et al. reported the analog and multilevel conductance with a low variation using a TiO2-based memristor with pre-defined one-dimensional (1D) pore structures.30 Li et al. reported the mediation of the STP behavior from a memristor based on oriented 1D mesoporous SiO2 (mSiO2) thin films with a thermally evaporated Ag2S layer.31 They discovered that the existence of mesopores facilitates the growth and dissolution of Ag filaments due to reduced atom strain energy, leading to a shorter relaxation time compared with amorphous silica electrolytes. Nevertheless, further reduction of relaxation is limited by the grain boundaries in the 1D pore structure which serve as the energy barrier for diffusion.32 Mesoporous structures with higher dimensions, therefore, possess great potential to further improve ion diffusion through multiple drift and diffusion pathways, enabling nonlinear and high-dimensional responses to external stimuli requested for a physical reservoir. However, a physical RC system based on memristors with three-dimensional (3D) mesoporous structures, to the best of our knowledge, has never been reported.
In this work, we demonstrate the implementation of a physical RC system using a 3D-structured, highly ordered mSiO2 based memristor that is fabricated via a vacuum-free, low cost and fast sol–gel solution processing technique. The memristor exhibits both volatile and non-volatile resistive switching behaviors that are tunable by controlling the applied compliance current (CC). The fast mobility and diffusivity of Ag+ ion species migrating within the 3D porous structures enable the demonstration of the STP behavior and temporal dynamics. A RC system based on this mSiO2 diffusive memristor has been experimentally demonstrated to perform a pattern recognition task with the 3D-structured mSiO2 layer serving as a reservoir matrix for the system. This work opens the route for simple and low cost memristor devices for building neuromorphic and physical RC systems.
The same mesoporous silica was deposited onto a Si/TiN substrate for memristor fabrication. The schematic of the fabricated mSiO2 memristor is shown in Fig. 1d (details of the fabrication process are in the Experimental section). The mSiO2 layer is sandwiched between an inert TiN bottom electrode and an active Ag top electrode. Fig. 1e presents the cross-sectional TEM image of a pristine mSiO2 memristor. The highly-ordered 3D-structured mesopores are clearly visible from the higher magnification TEM image in Fig. 1f. Energy-dispersive X-ray (EDX) mapping (shown in Fig. 2d–g) over the cross-sectional area also confirms the composition of each layer.
Modulating the CC level is a commonly used approach to control the resistive switching properties of memristor devices between volatile and non-volatile behaviors.35–37 Besides the volatile (temporal) memory behavior, our mSiO2 memristor also demonstrates non-volatile (non-temporal) memory behavior when the devices are operated at higher CC (1 mA and above). Fig. 2e shows the successive I–V characteristics of a DC swept between ±2 V with a current compliance (CC) of 5 mA. The graph clearly shows the typical clockwise bipolar resistive switching behavior. The direction of the current sweep is indicated by the arrows. After switching the memristor to the LRS under a CC of 5 mA, the LRS was maintained after the bias is removed. To switch the memristor back to the HRS, a RESET process was required. The DC endurance of the memristor is plotted in Fig. 2f with over 180 cycles. The resistance ratio between the HRS and LRS is about 1 order of magnitude at 0.5 V, which is sufficient for memory applications. Both the HRS and LRS showed excellent uniformity without significant degradation. To elucidate further the switching stability of the cell, cumulative probability distributions for the HRS and LRS are plotted in Fig. 2g. The figure clearly shows that the HRS and LRS are uniformly distributed with the coefficients of variation (σ/μ, σ is the standard deviation and μ is the mean value) calculated to be 15.95% and 8.57% for the HRS and LRS, respectively. Furthermore, high uniformities were also observed for both VSET and VRESET as shown in Fig. 2h, where the coefficients of variation were 4.31% and 2.95%, respectively. The high uniformities of both programming voltages are essential for large-scale ReRAM applications. Similar bipolar switching characteristics were obtained from multiple devices (Fig. S4†), showing good reproducibility of our mSiO2 based-memristor. Space-charge-limited-current (SCLC) was found to be the dominating conduction mechanism for the memristor at the HRS while ohmic conduction was observed in the LRS (Fig. S5†).
To elucidate the switching mechanism of our mSiO2 memristor, TEM and EDX characterisation studies (shown in Fig. 3a–e) were conducted on a device that has been switched several times before it was left at the HRS. In comparison with the as-fabricated device (shown in Fig. 1e), the switched device contained Ag conical dendrites that had grown from the Ag top electrode towards the TiN bottom electrode (Fig. 3a and b, more TEM images in ESI Fig. S6†). This suggests that the switching behavior of mSiO2 can be largely ascribed to the formation and rupture of Ag conductive filament(s) across the device. Similar mechanisms have also been reported in other works.38,39 It is very interesting to notice that the Ag dendrites are found within the highly-ordered mesopore structure, which acts as a pathway for the drift and diffusion of Ag ions. Based on the experimental results, the switching mechanism at different CCs is proposed and schematically presented in Fig. 3f. Upon the electroforming process, which is induced by the application of positive voltage on the active Ag top/TiN bottom electrode, the active Ag atoms are oxidized to Ag+ ionic species and migrate through the 3D-structured mSiO2 thin film following the electric field direction. After capturing electrons, these Ag+ ions are reduced and precipitate into nanoclusters at or near the TiN inert electrode (Fig. 3f-ii). These nanoclusters grow as more Ag atoms accumulate on top of them and form conductive filaments within the mSiO2 layer. When the grown conductive filament(s) connects the top and bottom electrodes, the device switches to the LRS (Fig. 3f-iii). In the case of a small CC (100 μA), the grown conductive filament is weak and the Ag atoms can spontaneously diffuse through the 3D mesopores, causing an automatic transition from the LRS to the HRS without the RESET process (Fig. 3f-iv). It has been suggested that the internal atomic rearrangement of filaments is strongly related to the interface between the filament and dielectric film during the electrochemical metallization process.40 Formation of voids in the SiO2 layer has been proved to facilitate the movement of Ag species in the dielectric layer due to reduced energy.41 The spontaneous movement of Ag atoms within our mSiO2 is therefore not surprising, considering the tremendous number of voids generated by 3D mesoporous structures. These voids contribute to the lower interfacial energy and atom strain energy,31 leading to low switching voltages (ca. 0.5 V for volatile switching and ca. ±1.5 V for non-volatile switching). On the other hand, such 3D mesopore structures also induce spontaneous dissolution of unstable and weak Ag filament(s) after the withdrawal of bias, resulting in the volatile switching behavior. Similar instability of nanowires in mesoporous silica films has been reported before.42 In the case of a large CC (5 mA), the grown conductive filament(s) across the mSiO2 thin film is large and stable (Fig. 3f-v), resulting in non-volatile switching behavior (the device retains its LRS). In this case, a sufficiently large voltage of opposite polarity is required to rupture the conductive filament(s) and therefore to switch the device to its initial HRS (Fig. 3f-vi). Given the high mobility of Ag in the mSiO2 film, it is likely for the Ag ions at the bottom mSiO2/TiN interface to drift swiftly back to the top electrode under the negative bias, resulting in residual Ag dendrites at the Ag/mSiO2 interface as suggested by the TEM image in Fig. 3a.
A synapse can be considered as a two-terminal device whose synaptic weights can be dynamically modified and stored using voltage spikes as illustrated in Fig. 5a. Upon the arrival of an action voltage pulse at the pre-synaptic terminal, the weights in biological synapses are changed by the release of Ca2+ ions from the pre-neuron. This leads to dynamic modulation of the synaptic connection intensity between the pre- and post-synaptic neurons which forms the biological basis for brain information processing.6 After stimulation, the Ca2+ concentration restores back to its original level via extrusion through the plasma membrane and back to the synaptic cleft. This biological behavior characterizes a typical STP where the intensity of synaptic connection will return to its initial value in hundreds to thousands of milliseconds after the pre-synaptic activity ceases. STP is thought to be critical in performing computational functions relevant to spatiotemporal information processing in biological neural systems, and is critical for the implementation of hardware-based reservoir computing.11,13 This essential biological synaptic plasticity can be emulated by using the two-terminal mSiO2 based memristor proposed in this work. The top and bottom electrodes are regarded as the pre- and postsynaptic neurons while the drift and diffusion of the Ag+ ions mimic the release and restoration of Ca2+ ions. The memristor conductance can be regarded as the weight of the synapse while an increase or decrease in conductance corresponds to the enhancement or suppression of the synaptic connection. To demonstrate similar temporal dynamics of the mSiO2 memristor, electrical pulses were applied to the device and the postsynaptic current (PSC) was recorded using a read voltage of 0.1 V. As illustrated in Fig. 5b, a substantial increase in the PSC can be observed after the application of electrical pulses (1 V amplitude, 5 ms duration), indicating the formation of conducting Ag filaments in the mSiO2 layer (blue region). This PSC increment is enormous for the first pulse but attenuates after several pulse sequences. This suggests that the Ag ions can easily drift into the mSiO2 film within the first few pulses to form a conducting filament(s), thanks to the 3D porous structure. Simultaneously, the structure also facilitates the diffusion of the filament(s), leading to the slow saturation of the current where the drift and diffusion reach a balance. Once the electrical pulses are removed, the mSiO2 memristor experiences a sharp and spontaneous transition back to its original insulating state, implying the spontaneous dissolution of the Ag filaments into the mesoporous structure (pink region). This sharp transition represents a short relaxation time, and was found to be 600 μs based on the stretched-exponential based function (SEF) model (shown in Fig. S7†).45,46 It is worth mentioning that this relaxation time is significantly shorter than those reported from other bulk thin film-based memristors,17,19,21 thanks to the high Ag atom mobility offered by the unique 3D structure. Such shortened relaxation time is beneficial for enabling fast computing and processing. Fig. 5c further demonstrates the spike-rate-dependent plasticity (SRDP) behavior of the mSiO2 memristor when stimulated by pulse trains at different intervals. When multiple pulses are applied at short intervals (e.g. 5 ms, blue), the postsynaptic current will increase gradually, showing an accumulation effect of Ag ions in the mSiO2 film. On the other hand, if the interval between pulses increases (e.g. 20 ms, green), the current soon decays back to the original resting state as the redistribution of Ag ions is completed before the arrival of another pulse. Moreover, the synaptic weight in our artificial memristor synapse can also be modulated by the pulse amplitude and duration of the pulse. Spike-voltage dependent plasticity (SVDP) is demonstrated in Fig. 5d where continuous pulse trains with varying voltages (1 V, 2 V and 3 V) were applied on our memristor. It can be observed that the PSC increases remarkedly as the spike amplitude increases from 1 V to 3 V. This suggests that larger pulse voltages can trigger more Ag ions into the mSiO2 structure and result in higher device currents. Similarly, extending the duration of the pulse can also amplify the strength of the connection. This spike-duration-dependent plasticity (SDDP) behavior can be observed in Fig. 5e which shows a significantly increased response current when increasing the duration from 1 ms to 20 ms.
As a demonstration, we applied the mSiO2 memristor based RC system to perform a basic pattern recognition task to classify five different letters – “ADEPT”, short for our research project “Advanced Devices by ElectroPlaTing”. Prior to the recognition process, the pattern information needs to be firstly converted to temporal input signals. This can be achieved by treating the patterns (letters) as pixels with a size of 5 rows × 4 columns in either red (1) or white (0) colors. As an example, Fig. 6b presents the conversion of the letter “A” into a 5 × 4 pixel array. The converted array contains 5 temporal pulse sequences including signals of “0110”, “1001” and “1111” based on the spatial distribution of the red and white pixels. The pixel arrays and the associated pulse sequences for the letter “D”, “E”, “P” and “T” are listed in the ESI Fig. S8.† Each pixel-based letter was split into 5 temporal pulse sequences (3 V, 100 μs) by row to be delivered into the five selected mSiO2 memristors as shown in Fig. 6c. Here, our mSiO2 memristor acts as a physical reservoir to realize the mapping of the temporal input signal into a high dimensional space through its nonlinear dynamics and temporal memory properties offered by the drift and diffusion of Ag ions in its unique 3D connected mesoporous structure. Upon the arrival of the pulse sequences, the conductance of the mSiO2 memristor will change according to the temporal order from the pattern and the output current therefore contains the information of a spatially distributed pattern. Fig. 6d plots the recorded 5 output currents from the letter “A” where it is clear that different temporal inputs have led to different current states. The corresponding current responses from other letters are demonstrated in Fig. S8.† For the chosen 5 letters in this work, a total of six basic pulse sequences of different orders can be extracted. Fig. 6e plots the distribution of each pulse sequence from over 500 measurements based on several memristor devices. Distinct responses can be identified with sufficient margins to separate different sequences. Details of the current change for each pulse sequence are shown in Fig. S9.† These output currents are fed into the read-out function to realize pattern recognition function. The read-out function in this system is a 5 × 5 network with the 5 currents from the reservoir layer as the input and the 5 neurons representing the predicted letters as the output (shown in Fig. 6c). This network outputs the dot product of the 5 inputs and the weight associated with each output neuron, which represents the probability value of each letter. The letter with the highest probability is finally selected to be the predicted pattern. The network is trained via a supervised learning algorithm where the weights are adjusted to minimize the cross-entropy loss. More information on the network training process can be found in the Experimental section and the ESI.†
It is very important to notice that only 25 weights need to be trained in this network which is significantly smaller than that required in conventional neural networks.20 This significantly reduced network size allows the reduction of time and energy-consumption of the training process. Fig. 7a plots the training loss curve where the drastic reduction of loss can be observed within a few training epochs. This is accompanied by the increase of prediction accuracy as shown in Fig. 7b. After 20 epochs, a training accuracy of 100% was obtained, which supports the argument of a low training cost for the RC system. After the RC system was successfully trained, its performance was evaluated by recognizing letters from the test dataset which have never been seen by the system before. Fig. 7c shows the confusion matrix highlighting the experimentally obtained classification results from the RC system versus the prediction outputs (desired outputs). A recognition accuracy of 100% was achieved from the reservoir, indicating the outstanding performance of the trained mSiO2 diffusive memristor based-RC system.
Moreover, the robustness of our RC system in pattern recognition was further tested by classifying letters with added noises. A noise test dataset was generated by randomly inverting the color for different numbers of pixels in the letter to be recognized. Fig. 7d, e, and f illustrate the examples of letters that contain 1, 2 and 3 inverted pixels (noises) in their pattern, where the noisy pixels are highlighted in red squares. Recognition accuracies of 93%, 88%, and 75% were obtained from the reservoir by including 1, 2, and 3 noisy pixels, respectively. The corresponding confusion matrices are shown in Fig. 7g–i. By introducing noisy pixels to the original images, it was found that the letters can still be recognized by the RC system even when the output signal from the reservoir is disordered.
Dip-coating was used to coat silica films onto the TiN bottom contact. The TiN bottom contact with a thickness of 200 nm was deposited using the sputtering technique. The cleaned TiN substrates were vertically immersed in the prepared precursor solution and withdrawn at a rate of 100 mm min−1 in a 75% relative humidity chamber at 298 K. After storing in the humidity chamber (RH = 75%) for 72 h, the as-made films were aged at 120 °C for 10 h. The surfactant was removed by immersing it in the DCM solution for 4 hours and calcined at 350 °C for 5 h. Afterwards, these films were referred to as mesoporous silica films. A schematic showing the formation of mSiO2 thin films is shown in Fig. 1a.
Footnote |
† Electronic supplementary information (ESI) available. See DOI: https://doi.org/10.1039/d2nr05012a |
This journal is © The Royal Society of Chemistry 2022 |