Realizing reliable linearity and forming-free property in conductive bridging random access memory synapse by alloy electrode engineering

The linearity of conductance modulation of the artificial synapse severely restricts the recognition accuracy and the convergence rate in the learning of artificial neural networks. In this work, by alloy electrode engineering, a Ti–Ag device gained the forming-free property because Ag ions were promoted to migrate into the GeTeOx layer to form a thicker conductive filament. This facilitated a uniform change in conductance with the pulse number, and the alloy synapse achieved a significant improvement in linearity (350%), which demonstrated its enhancement in recognition accuracy. To further validate its potential as a comprehensive artificial synapse, the multi-essential synaptic behaviors, including spike-timing-dependent plasticity, spike-rate-dependent plasticity, paired-pulse facilitation, post-tetanic potentiation, and excitatory post-synaptic current, were achieved successfully. This work proposes a promising approach to enhance the performance of conductive bridging random access memory synaptic devices, which benefits the hardware implementation of neuromorphic systems.

The linearity of conductance modulation of the artificial synapse severely restricts the recognition accuracy and the convergence rate in the learning of artificial neural networks.In this work, by alloy electrode engineering, a Ti-Ag device gained the forming-free property because Ag ions were promoted to migrate into the GeTeO x layer to form a thicker conductive filament.This facilitated a uniform change in conductance with the pulse number, and the alloy synapse achieved a significant improvement in linearity (350%), which demonstrated its enhancement in recognition accuracy.To further validate its potential as a comprehensive artificial synapse, the multi-essential synaptic behaviors, including spike-timingdependent plasticity, spike-rate-dependent plasticity, paired-pulse facilitation, post-tetanic potentiation, and excitatory post-synaptic current, were achieved successfully.This work proposes a promising approach to enhance the performance of conductive bridging random access memory synaptic devices, which benefits the hardware implementation of neuromorphic systems.© 2024 The Author(s).[3][4][5][6][7][8][9][10][11] Conductive bridging random-access memory (CBRAM) is dominated by the formation and rupture of the conductive filament (CF) due to the oxidation and reduction of metallic ions.][14] However, a high voltage is required to be applied to the CBRAM device during the forming process, which comes at the cost of higher power consumption and complexity of the design. 14,15)In addition, the randomness of the growth process and the shape of the CF affect the accuracy of the conductance modulation, which restricts the application severely.Importantly, the linearity of the potentiation and depression is one of the essential parameters of a high-efficiently artificial synapse, and the learning accuracy significantly is degraded by the non-uniform modulation of conductance in a neural network. 16,17)[20][21][22][23] In this study, to overcome this issue, we propose a method to improve the non-volatile memory characteristics and synaptic performance through electrode alloying.Due to the ability of Ti to control oxygen ions, more Ag ions promote the growth of the thicker CF, resulting in low forming and switching voltages.To understand the physical mechanism, the ion migration and CF morphology were analyzed by the kinetic Monte Carlo method.Continuous conductance modulation was performed under DC and pulse modes, and the alloy synapse exhibited excellent linearity of potentiation and significant depression.Furthermore, fundamental synaptic behaviors including STDP, SRDP, PPF, PTP, and EPSC were realized respectively, indicating the feasibility of alloy synapses.This work demonstrated a potential approach for optimizing the properties of artificial synapses based on CBRAM, and advances the development of neuromorphic systems.
To obtain the Ti/GeTeO x /Ti-Ag (Ag) device, the Pt/Ti/SiO 2 /Si substrate was prepared.The Ag film was deposited by DC magnetron sputtering at a power of 30 W for 800 s, and the Ti-Ag film was deposited by DC magnetron co-sputtering with Ti and Ag targets at a power of 30 W for 600 s.Then, the GeTeO x film was deposited by RF sputtering with the GeTe target at a power of 40 W for 360 s.Finally, the Ti film was patterned by a mask aligner process to form the top electrode with a diameter of 100 μm.The vacuum environment of the chamber was less than 5.0 × 10 −5 Pa at RT, and the DC and RF sputtering systems above were both in Ar gas (40 rate sccm) at a working vacuum pressure of approximately 0.5 Pa.The cross-section was observed by scanning electron microscopy (Crossbeam 540).The chemical state and composition were examined by X-ray photoelectron spectroscopy (XPS).All electrical characteristics were performed using a Keysight B1500A semiconductor parameter analyzer.A DC bias was applied to the Ti top electrode and the Pt bottom electrode was grounded.
A typical Metal-Insulator-Metal structure of the Ti/GeTeO x /Ti-Ag/Pt device was demonstrated; a cross-section image is shown in Fig. 1(a) and a schematic is exhibited in the inset.The multiple-layer Ti-Ag, GeTeO x , and Ti thin films were 40, 90, and 90 nm, respectively, and performed a sharp separation.In addition, the results of the XPS material composition analysis of the Ti-Ag alloy and GeTeO x films are presented in Figs.1(b) and 1(c).The 2p 3/2 and 2p 1/2 peaks of Ti were located at 453.9 eV and 460.3 eV, and the 3d 5/2 and 3d 3/2 peaks of Ag were located at 374.2 eV and 368.2 eV, respectively.Compared with the peak area, the mole fraction of Ti:Ag was 20.06%:79.94%,indicating a wellprepared alloy electrode.In addition, the three peaks of Ge 3d were attributed to the existence of the Ge-O (32.2 eV), Ge-Te (30.0 eV), and Ge (29.4 eV) bonding energy.The two peaks of Te 3d were attributed to the existence of the Te-Ge (572.7 eV) and Te atom (572.9 eV) bonding energy, and the mole fraction of Ge:Te:O was 48.4%:35.06%:16.54%.These results imply that the alloy device was produced successfully.
To study the performance, the same forming process with a compliance current was applied, and the corresponding curves and its device-to-device variations of the Ti/GeTeO x/ Ti-Ag(Ag) device are shown in Fig. 2(a).Notably, the forming voltage of the Ti/GeTeO x /Ti-Ag device was much lower than that of the Ti/GeTeO x /Ag device, demonstrating a significant decrease by electrode alloying.When the voltage (−0.27V) was applied, the alloy device reached a low resistance, and the forming voltage was sufficiently low, which was similar to the set voltage.This meant the alloy device was able to be used directly without an extra forming process, and exhibited a forming-free property, which could reduce the cost of the higher power consumption and simplify the complexity of the circuit design.Subsequently, the typical bipolar resistive switching curves are depicted for comparison in Fig. 2(b).The devices changed from a highresistance state (HRS) to a low-resistance state (LRS) at the set voltage (V set ). Until the reset voltage (V reset ) was applied, the device returned to HRS.It is clear that the alloy device performed more coincident I-V curves, exhibiting outstanding uniformity.Significantly, the forming voltage was sufficiently low and similar to the set voltage.This meant an extra forming process was unnecessary, corresponding to the forming-free property, which might simplify the technical step.Specifically, the statistical distribution of switching voltages is shown in Fig. 2(c).The alloy device had a lower switching voltage with a narrower fluctuation.The coefficient of variation of V set of the two devices was 22.29% and 5.64%, while that of V reset was 6.67% and 10.45%, which meant the uniformity of V set was optimized greatly and changed tolerably of the V reset .In addition, the resistance distribution of the HRS and LRS is shown in Fig. 2(d).Both devices displayed a remarkable memory window with seldom degeneration for storing information.Furthermore, the device-to-device variations of the alloy device were investigated to demonstrate its reliability.As shown in Figs.2(e) and 2(f), the switching voltages presented high-level uniformity, and the resistance of LRS/HRS with a narrow distribution indicated reasonable repeatability.Consequently, the alloy device with high performance has great potential for neuromorphic computing as an artificial synapse.
In order to understand the mechanism, the physical models proposed are shown in Fig. 2(g).During the forming process, Ag atoms were oxidized to ions and migrated into the GeTeO x layer.Subsequently, these Ag ions were reduced to atoms and CFs formed completely.However, when the Ag metal was applied and in contact with the switching layer, the standard Gibbs free energy of oxide (ΔG) was low due to its highly chemical activity.The Ag ions were able to absorb many oxygen ions, thus reducing the amount of it migrating into the resistive layer.Due to the ability of more active Ti to control oxygen ions, more Ag ions migrated into the resistive layer when the Ag electrode was replaced by a Ti-Ag alloy electrode.Consequently, a smaller voltage was sufficient to form a complete conductive path, which caused a formingfree property.Meanwhile, more Ag atoms promoted the growth of the thicker CF (proved by kinetic Monte Carlo simulation in S1).During the set process, the device changed to LRS as the CF was connected by Ag atoms.As a reset voltage was applied, the Ag atoms were oxidized and caused the tip of the CF to rupture, and the device returned to HRS.Due to the uncontrollability and randomness of the Ag ion movement, the CF formed and ruptured randomly, resulting in poor uniformity and reliability.However, a large amount of Ag ions around the tip of the CF of the alloy device caused easy reconnection of the CFs.Therefore, the alloy device had a lower V set .Meanwhile, the thicker CF enhanced the switching control and reduced the variability, indicating a narrow distribution of V set .In addition, the tips of the CF proceed gradually rather than directly during its formation and rupture, which was beneficial to the linear modulated conductance.The synaptic plasticity characteristic was analogous to gradually changing the resistance by controlling the voltages.As shown in Figs.3(a) and 3(b), with the number of DC sweeps increasing, the current raised slightly in the negative voltage region, while the current decreased gradually in the positive voltage region.Significantly, the state of resistance was maintained after each sweep.This result indicated that the device simulated the synaptic behaviors preliminarily.Furthermore, the performance was investigated under a pulse operation consisting of 20 negative and 20 positive pulses.
As shown in Fig. 3(c), the normalized conductance gradually increased and the state was maintained after each negative pulse continuously, corresponding to long-term potentiation (LTP).As the positive pulses were applied, the normalized conductance decreased, which conformed to long-term depression (LTD) behavior.The device exhibited a continuous gradient characteristic of the conductance, which was similar to synaptic behavior.The normalized conductance varied remarkably uniformly with the pulse number increasing, indicating a high linearity of potentiation and depression.Importantly, by varying the pulse amplitude, width, and interval, the alloy device exhibited excellent linearity, which described the uniformity of the conductance changes (see S2).The pulse properties of the Ti/GeTeO x /Ag under an identical method are depicted for comparison and the same trend is observed, as shown in Fig. 3(d).According to calculations, 14) the linearity of the Ti-Ag alloy synapse increased by 350%.By mixed hardware-software studies of convolutional neural networks, the influence of linear conductance modulation on the recognition accuracy of the network was verified (see S3).Moreover, the trained model exhibited high recognition accuracy and tolerance to input image noise under different circumstances, which indicates the reliability of the results (see S4).The proposed model exhibited that the recognition accuracy increased with the increase in linearity.Thus, the alloy synaptic device with high linear conductance modulation was preferred for neuromorphic computing applications.Moreover, the various synaptic plasticity of the alloy device was investigated comprehensively to demonstrate the potential as an artificial synapse.The experiment of emulating the STDP learning rule was carried out firstly.As shown in Fig. 4(a), the pre-spike consisting of eight negative potentiation pulses and eight positive depression pulses was applied, and the postspike was designed as one pulse.The relative change was defined as ∆w = -100% , where Δω is the synaptic weight change value while I before and I after are the currents before and after the pre-spike and post-spike pairs.The STDP-like curve was obtained by changing the spike timing Δt (Δt = t post −t pre > 0), presented in Fig. 4(b).To investigate the effect of the spiking rate, a pair of 10 consecutive stimulus pulses with an amplitude of −0.33 V and a width of 100 ns, separated by different intervals, was applied.As shown in Fig. 4(c), the smaller the interval, the greater the change in current, indicating that the SRDP property in biological synapses was emulated well.Meanwhile, when the pre-synapse membrane was stimulated by an action potential, the synaptic transmission was enhanced temporarily.When a second identical stimulation was applied, the post-synapse produced a greater response, known as PPF.When the same stimulation was applied to the pre-synaptic membrane sequentially and repeatedly over a short time, the response of the post-synapse was markedly enhanced, known as PTP.To analyze more specifically, the changes in current following different numbers of successive pulses of stimulation are visually shown in Fig. 4(d).Statistically, the more pulses that are stimulated at the same interval in a short time, the greater the current.Consequently, this phenomenon was used to mimic biological plasticity, where the current change after the second pulse (I 2 −I 1 ) and the tenth consecutive pulse (I 10 −I 1 ) correspond to PPF and PTP.In addition, as the pulse stimulation was applied, the post-synaptic current at a read voltage of −0.05 V increased abruptly and then decreased gradually, corresponding to the EPSC behavior of biological synapses, as shown in Fig. 4(e).As shown in Fig. 4(f), to study the learning and forgetting process, a series of continuous pulses were applied.After 25 successive bias pulses, the synaptic weight reached 100%.Then, it underwent spontaneous decay for 100 s.To recover the pre-decay level of the synaptic weight, 15 successive bias pulses were sufficient.This indicated that receiving new information would go through the process of learning, forgetting, and relearning.Consequently, the alloy device has potential as an artificial synapse for the application of neuromorphic systems.
In summary, the proposed Ti-Ag synaptic device, based on alloy electrode engineering, demonstrated a forming-free property, low switching voltage, and outstanding uniformity and reliability.To mimic the synaptic behavior, the continuous variation characteristics of the conductance were investigated in DC and pulse modes, and the alloy synapse showed a significant improvement in the linearity during the LTP and LTD processes, which contributed to enhancing the recognition accuracy.In addition, the varying synaptic plasticity was investigated comprehensively, such as the STDP, SRDP, PPF, PTP, and EPSC synaptic behaviors, indicating the device's potential to be an excellent artificial synapse.This work provided a valuable method for optimizing the CBRAM synaptic performance.036505-5 © 2024 The Author(s).Published on behalf of The Japan Society of Applied Physics by IOP Publishing Ltd

Fig. 2 .
Fig. 2. Comparison of the resistive switching performance: (a) the I-V curve of forming process and its device-to-device variations of the Ti/GeTeO x/ Ti-Ag(Ag) device; (b) the I-V characteristics over 100 cycles; (c) cumulative probability distribution of the voltages; (d) the distribution of resistance at V read .The device-to-device variations of the alloy device: (e) the switching voltages, (f) the resistance.(g) The physical models of the set and reset process.

Fig. 3 .
Fig. 3.The gradual change of current with (a) negative sweep and (b) positive sweep.Normalized conductance pulse distribution for (c) Ti-Ag alloy device and (d) pure Ag device.

Fig. 4 .
Fig. 4. (a) The STDP realization schemes; (b) the STDP-like curve obtained; (c) measured currents for 10 pulse cycles with different pulse intervals; (d) mean changes ΔI of current during cycles of 10 pulses at different intervals; (e) EPSC characteristics with voltage pulse stimulation (−0.17 V, 300 ms); (f) the learning experience behaviors of the synaptic device.