Artificial Synapses Based on Atomic/Molecular Layer Deposited Bilayer-Structured Memristive Thin Films

This chapter deals with several kinds of ultrathin bilayer-structured memristors, such as Pt/Al 2 O 3 /HfO 2 /TiN, Pt/HfO 2 /HfO x /TiN, Pt/TiO 2 /Ti-based maleic acid (Ti-MA)/TaN, among which the asymmetric memristive functional layers were designed and prepared by atomic layer deposition (ALD) or molecular layer deposition (MLD) technique. These bilayer memristors exhibit a typical bipolar resistive switching characteristic, in accord with the space charge limited current model. Some important biologic synaptic functions have been achieved, including nonlinear transmission characteristics, spike-timing-dependent plasticity, short (cid:1) /long-term plasticity, paired-pulse facilitation, and conditioned reflex. The mechanism of bilayer memristive device has been proposed based on oxygen vacancies migration/diffusion model. Above all the ultrathin bilayer memristors fabricated by low temperature ALD/MLD are one competitive candidate for neuromorphic simulation and flexible electronic applications.


Introduction
The memristor concept was first proposed as the fourth fundamental passive circuit element by Chua in 1971 based on the completeness of the circuit theory, which indicates the relationship between magnetic flux and charge [1,2]. After thirty seven years, Strukov et al. eventually found the missing memristor in studying TiO 2 cross-arrays in 2008 [2]. This draws the extensive and intensive attention from the academia and the industry. Memristor is a two-terminal electrical device whose resistance can be tuned by changing the flux or charge through it. Memristor possesses a lot of advantages, e.g., simple device architecture, high energy efficiency, better compatibility with semiconductor industry, and high integration density.
A neural synapse, as the basic unit of learning and memory in the brain, plays a critical role in biological neural networks. Electronic synapses are utilized to emulate the bio-synapses' functions. Some researches on synapse simulation have been reported by adjusting synaptic weights so as to make an effective bio-inspired computing system [3][4][5][6]. Nevertheless, most work chose transistors and capacitors to realize artificial synapse, which produced high energy consumption at high integration density and limited the programming running. The new memristor has nonlinear transfer characteristics similar to the bio-synapse and is regarded as the closest to the synaptic device [4].
Although various materials and structures exhibit memristive behavior, almost all the memristor systems are based on the structural asymmetry [7,8]. For example, in the metal-insulator-metal (MIM) structure, the defects such as oxygen vacancy or active ions in the insulator layer can induce structural asymmetry under the action of the external field, or when one of the metal electrodes is active. Therefore, the asymmetric bilayer-structured memristors play a crucial role in constructing artificial neural networks for brain-inspired applications.
Atomic layer deposition (ALD) is a kind of commercial technology compatible with semiconductor processing. It shows unusual advantages in controllable fabrication of nano-laminate thin films due to its unique sequential self-limiting surface reaction mechanism at low growth temperature [9,10]. In early 2001 ALD has been known as candidate technology preferred for semiconductor industry along with metalorganic chemical vapor deposition (MOCVD) and plasmaenhanced CVD by the international technology roadmap for semiconductors (ITRS) [11]. ALD has become one of the most competitive deposition techniques for microelectronics and nanotechnology owing to sub-nanometer thickness control, large-area uniformity, excellent three-dimensional conformality, and good reproducibility. Thin films with low defect density can be prepared by ALD even at room temperature (RT) with plasma assistance [12]. Evidently, low temperature or RT ALD technology can greatly widen the flexible substrate choice range, showing exciting potentials in flexible electronic device fabrication. Molecular layer deposition (MLD) can be regarded as the subtype of ALD due to the molecular nature of the deposition process, suitable for growth of organic-inorganic hybrid materials [13].
In this section, we fabricated several synaptic devices of asymmetric bilayerstructured ultrathin memristors by atomic layer deposition (ALD) and molecular layer deposition (MLD), such as Pt/AlO x /HfO x /TiN, Pt/HfO 2 /HfO x /TiN, Pt/ TiO 2 /Ti-based maleic acid (Ti-MA)/TaN. Some biological synapse-like functions of longÀ/short-term plasticity (LTP and STP), spike-timing-dependent plasticity (STDP), and paired-pulse facilitation (PPF) have been achieved simultaneously. A memristive mechanism of an asymmetric bilayer-structured synaptic device has been proposed to explain synaptic plasticity based on the oxygen vacancy migration/diffusion model.

Fabrication processing
Asymmetric bilayer-structured ultrathin memristor based on Pt/A/B/TiN or TaN was fabricated on SiO 2 /Si substrates by thermal-ALD (TALD), MLD and plasma-enhanced ALD (PEALD), as illustrated in Figure 1a. Herein A and B act as asymmetric memristive functional layer, PEALD TiN or sputtered TaN as bottom electrode, sputtered Pt as top electrode with a spot size in diameter of 150 μm. Table 1 gives several typical bilayer ultrathin memristors and their architectures. The related deposition conditions have been listed in Table 2, including used metal precursors and reactants, source temperature and deposition temperature, and growth per cycle (GPC).  [15].

Electrical performances and synaptic functions
The electrical properties were measured under DC sweep and pulse modes using semiconductor parameter analyzer on probe station. The bottom electrode of memristors was set on ground and all the voltage signals were applied to the top electrode. The asymmetric bilayer ultrathin memristors were exploited to mimic some important synaptic functions such as long-term potentiation/depression, the transition from STP to LTP, PPF and STDP.

TiN/ZnO x /HfO x /Pt inorganic memristor
The I-V curves of Pt/HfO x /ZnO x /TiN inorganic memristor are plotted in Figure 1b under a modified DC double sweep. To mimic the functions of a nerve synapse, one multiple-state resistances should be obtained in bilayer memristor. A continuous set or reset process was performed by successive increasing the compliance from 0.1 to 1.0 mA at an interval of 0.1 mA or altering the reset voltage from 1.0 to 1.7 V at an interval of 0.05 V. 8 low resistance states (LRS) and 11 distinguishable high resistance states (HRS) are observed during consecutive set and reset process, respectively. Moreover the resistance can be continuously reduced or raised between multiple intermediate states without going back to the original state, which is key for electronic synapse [14]. The device conductivity decreases continuously with six easily recognized states after exerting sweep positive bias voltage from 0 to 1.4 V six times and the elevated conductivity with difficultly distinguishable ones after sweep negative pulse voltage from 0 to À0.6 V (Figure 1c and d), indicating the conductance change caused by consecutive potentiating or depressing signals. It can be attributed to the dynamic change of oxygen vacancy concentration and distribution in asymmetric bilayer structure of HfO x /ZnO x under various electrical signals [15].
A series of pulse signals were designed and applied to the memristor to test the important STDP rule in the Hebbian learning theory, as seen in the insets of I and III of Figure 2, including the VÀ/V+ = À1.0 V/1.0 V pulse pair signal as a presynaptic and postsynaptic spike with the 3 s interval time. Such design can prevent from the disturbance of excitatory postsynaptic current [16]. The time interval between the final presynaptic spike and the initial postsynaptic spike is defined as the relative time of Δt. The relative change of the synaptic weights (ΔW) is defined as: The initial postsynaptic or presynaptic current I 1 was used as the control value. After the spike pair was applied and over for 5 min, the measured presynaptic or postsynaptic current was I 2 .
The dependence of ΔW on Δt of Pt/HfO x /ZnO x /TiN in Figure 2 II and IV follows the STDP learning rule. While the presynaptic spike happens before the postsynaptic spike (Δt < 0), synaptic weights enhance, indicating long-term potentiation (LTPo); while the presynaptic spike appears after the postsynaptic spike, synaptic weights become small (Δt > 0), implying long-term depression (LTD). And the shorter the Δt between the two spikes, the larger the ΔW. The STDP data points of memristor in Figure 2 show evident statistical scatter, similar to the biological synapse.
In addition, Pt/HfO x /ZnO x /TiN device also exhibits the nonlinear transmission efficiency, and the transition from STP to LTP (not shown here) [15].

TiN/HfO x /AlO x /Pt inorganic memristor
The memristor device based on TiN/HfO x /AlO x /Pt can also emulate the biological synapse. Usually, the synapse operates under pulse signals rather than DC bias sweep voltage. The LTPo and LTD phenomena can be observed in Pt/AlO x /HfO x / TiN under 180 potentiating pulses (À0.5 V, 10 ms) and 180 depressing pulses (1 V, 10 ms), as shown in Figure 3a. The connection strength can be dynamically modulated by the consecutive external signals, determining the transfer efficiency between the electronic neurons.  (a) Change in the response current under the influence of consecutive potentiating or depressing pulses. After pulse stimulation, a 0.1 V 100 ns reading pulse was applied to read the response current. After À0.5 V potentiating pulses, the response current gradually increases (long-term potentiation), while the response current gradually decreases after 1 V depressing pulses (long-term depression). (b) Response of a memristor device to different pulse programs [17].
Further experiments have demonstrated that a pulse signal from amplitude of 1.0-1.5 V and pulse width of 50-100 ms leads to various current responses in Figure 3b. That is to say, the larger pulse amplitude, the longer pulse width, and the more pulse number will produce more significant response current change, which is analogous to long-term potentiation/depression of the human brain.
Synaptic plasticity can be divided into STP and LTP according to the timelines of enhanced synaptic connections. The repeated stimulation induced STP to LTP transition is illustrated in Figure 4. With increasing the rehearsal pulse number (N = 10, 40, 70, 100, 120), the resistance remaining becomes larger (Figure 4a). This procedure is similar to the Ebbinghaus forgetting curve related to human memory [18,19].
An exponential decay equation was employed to depict the relaxation process: where M(t), M 0 , and M e are the memory level at time t, t = 0, and at steady state after a long time, and τ of the relaxation time constant. The experimental and simulation results after 70 identical pulses are shown in Figure 4b, containing the dependence of τ on N in the inset. The decay rate is faster in the beginning and then becomes slower. The τ value increases from several seconds to 50 seconds with the training pulse number from 10 to 120, revealing a declining forgetting rate from $57% for N = 10 to $5% for N = 120. This confirms the transition from STP to LTP through repeated rehearsal and learning.
The STDP rule has been mimicked in TiN/HfO x /AlO x /Pt memristor, as indicated in Figure 5. The schematic of another training pulse signal with various amplitudes is shown in Figure 5a, different from the pulse design in Figure 2. A set of pulses were used as pre-synaptic/post-synaptic stimulation signals, respectively. Some different pulse signals designed at various spike timings (Δt) are designed and illustrated in Figure 5b. When the shortest Δt (10 ms) is inserted to the device, the largest ΔW of 50% for potentiation and À 80% for depression are realized.
The memristive mechanism of asymmetric TiN/HfO x /AlO x /Pt memristor has been deeply investigated with the aid of x-ray photoelectron spectroscopy (XPS) depth analyses, which will be discussed in the following Section 2.3.

TiN/HfO x /HfO x /Pt inorganic memristor
In the previous work on Pt/HfO x /ZnO x /TiN and TiN/HfO x /AlO x /Pt memristors, the asymmetric memristive functional layers of A and B are different materials. Next, we will focus on Pt/HfO 2 /HfO x /TiN bilayer-structured memristor, as illustrated in Figure 6a. 4 nm-thick non-stoichiometric HfO x films were prepared by  PEALD using the H 2 plasma and 2 nm-thick stoichiometric HfO 2 films by TALD using the H 2 O precursor, in basically consistent with the measured result by the cross-sectional high angle annular dark field (HAADF)-scanning transmission microscopy (STEM) in Figure 6b. The energy dispersive x-ray spectroscopy (EDS) elemental mapping images of Pt/HfO 2 /HfO x /TiN are shown in Figure 6c, revealing the stacking structure. In addition, XPS composition analyses show that the atomic ratio of Hf:O in the HfO 2 and HfO x layers is 1:2.04 and 1:1.84, respectively, indicating that stoichiometric HfO 2 and nonstoichiometric HfO x bilayer-structured memristors have been obtained [20]. Hence A and B herein represent HfO 2 and HfO x with various oxygen contents, respectively. This device unit based on TiN/ HfO x /HfO 2 /Pt memristor can also simulate the biological synapse learning rule of STDP, as indicated in Figure 6d. When the shortest spike timing of 10 ms is applied to the memristor device, the pulse train responses give rise to the largest ΔW value of 83% for potentiation and À 65% for depression, respectively [20]. These ΔW values for STDP are similar for Pt/HfO x /ZnO x /TiN, TiN/HfO x /AlO x /Pt and TiN/ HfO x /HfO 2 /Pt memristors.
The paired-pulse facilitation (PPF) is a phenomenon wherein the post-synaptic response induced by the spike increases when the time interval of the two spikes is very close [20]. PPF index can be defined as follows: G 1 and G 2 are the conductance values after the first and the second pulse, respectively. The time constants of τ 1 and τ 2 can be assigned to the fast and slow decaying terms, respectively.
Evidently Pt/HfO 2 /HfO x /TiN memristor displays the marked dependence of synaptic weight on pulse interval Δt by applying the pulse of À1.5 V and 2.5 V, respectively, as seen in Figure 7a and b. For shortest Δt of 400 ns, the PPF index increases to 135% under positive pulse and becomes À62% under negative pulse. For the negative pulse signals, the calculated τ 1 and τ 2 values are 357 ns and 2.47 ms, respectively; for the positive pulses, τ 1 and τ 2 are 1.48 ms and 6.79 ms, respectively. When the Δt decreases, the memory effect will be improved, which is ascribed to the fact that the smaller Δt between pulses produces less oxygen vacancies to drift back with more effective accumulation of the oxygen vacancies. Pt/HfO 2 /HfO x /TiN also mimics a classical conditioning under different pulse stimuli, as illustrated in Figure 8. In the famous experiment [21], a dog salivates (unconditioned response, UR) when watching the food (unconditioned stimulus, US) (Figure 8a), when it does not salivate (conditioned response, CR) on hearing the ring (conditioned stimulus, CS) alone (Figure 8b). Nonetheless, after some rehearsals, i.e. feeding the dog when ringing the bell (Figure 8c), the dog salivates even in only hearing the ring (Figure 8d). This elucidates that the dog has correlated the food with the ring. Furthermore, when taking away the food, the correlation between the food and ring gradually reduces and even disappears under only the conditioned stimuli (Figure 8e). The whole procedure can be emulated in the Pt/HfO 2 /HfO x /TiN by using +4 V and À 1.3 V stimuli with a single pulse duration of 5 ms.
Before rehearsing, the memristor has a low resistance state of 5 kΩ. The +4 V stimulus (US) causes a high resistance state of 3 MΩ (UR) (Figure 8f), when the À1.3 V stimulus (CS) only results in a low resistance state of 5 kΩ before training (Figure 8g). In Pavlov's experiments, the food and the ring exist simultaneously to reinforce the correlation between US and CS. In our experiments, the +2.7 V stimulus was exerted to the memristor, the same as the simultaneous stimuli of À1.3 V and + 4 V pulse signals. When two rehearsing sequences with +2.7 V pulse, the device becomes the high resistance output of 2 MΩ (CR) (Figure 8h). When removing the +4 V signal, the memristor continues to keep in a high resistance state under a series of À1.3 V stimuli alone and then returns to a low resistance state (Figure 8i), implying the setup and vanish of the classical conditional reflex.
The energy consumption is one important indicator for a practical electronic synaptic device in neuromorphic network. Pt/HfO 2 /HfO x /TiN memristor can be set in less than 100 ns and reset in less than 10 ns, indicating the rapid switching speed, as recorded in Figure 9a.
The current response curves versus the time after the applied programming signal during the set or reset operation are plotted in Figure 9b and c, respectively. The current rises after a waiting time of about 260 ns when a À 2 V/1 ms stimulus is applied, indicating the beginning of the set process (Figure 9b). The memristor resistance decreases from the initial high resistance state ($1 MΩ) to low resistance state ($800 Ω). Similarly, the current reduces after a waiting time of about 70 ns when a +3 V/1 ms signal is exerted, showing the occurrence of the reset process. The energy consumption per operation can be calculated to be 520 pJ for the set process and 1.05 nJ for the reset process by considering the pulse waveforms (time, response current, and pulse voltage), corresponding to the maximum energy consumption in one set or reset operation, as the memristor has been set in the lowest resistance state with the highest response current. Nevertheless, the actual operation of the electronic synapse is generally in the mediate resistance states ($80 kΩ). The response current of the memristor is inversely proportional to the resistance value of the synaptic device with a first-order approximation. So, the evaluated actual energy consumption per operation will decline in the range of around ten picojoules.
Finally, the impact of oxygen vacancy concentration in non-stoichiometric HfO x layers on resistive switching properties of Pt/HfO 2 /HfO x /TiN bilayer ultrathin memristor has been investigated. The memristor with 12.1% oxygen vacancy concentration in the HfO x layer exhibits comprehensively better performances such as the optimal pulse energy consumption, reset switching speed, and DC endurance and retention characteristics [20].

TaN/Ti-MA/TiO 2 /Pt organic: inorganic hybrid memristor
As mentioned above, we mainly elucidated the bio-synaptic functions of three asymmetric inorganic bilayer-structured ultrathin memristors. In this part, organic-inorganic hybrid bilayer memristors of TaN/Ti-MA/TiO 2 /Pt were prepared by low temperature MLD/ALD at 160°C. The synaptic plasticity has been explored deeply. Some superb synaptic functions, such as nonlinear transmission characteristics, STP/LTP, PPF, and STDP have been achieved in the hybrid memristors [22].
First the narrow-scan XPS and Fourier transform infrared (FTIR) spectroscopy were used to detect the chemical composition and organic group of Ti-based maleic acid (Ti-MA) hybrid film, as shown in  (Figure 10c). The FTIR spectrum of Ti-MA hybrid film (Figure 10d) displays the asymmetric and symmetric stretch of carboxylate groups at 1575 cm À1 and 1447 cm À1 . The splitting of 128 cm À1 indicates the bidentate bond mode between the Ti ion and carboxyl. As a result, Ti-MA inorganic-organic hybrid films have been fabricated successfully.
The resistive switching characteristics of the hybrid bilayer memristor of TaN/Ti-MA/TiO 2 /Pt have been examined for 100 times, as seen in in Figure 11a. The typical bipolar resistive switching behavior has been confirmed with narrow distribution of set voltage of À1.6 AE 0.2 V (red line) or reset voltage of 1 AE 0.1 V   (region 3, R 2 = 0.9904). All these prove the space charge limited current (SCLC) model in hybrid bilayer memristor [24], revealing the filament model of oxygen vacancy migration.
The PPF and STDP functions have also been characterized in hybrid memristor, as shown in Figures 12 and 13, respectively. A pair of pulses (À1 V, 400 ns) with different Δt were applied to the hybrid memristor (Figure 12a). The measured data can be well fit exponentially (Figure 12b). The PPF index has reached to 361% with the 400 ns pulse interval in hybrid memristor. When the pulse interval increases to 2400 ns, the PPF index dramatically tends to 3% [22]. Compared to inorganic bilayer memristor, the organic-inorganic hybrid bilayer device has much larger PPF index in the same pulse interval of 400 ns.  The STDP rule was emulated in hybrid memristor by using a pair of 0.8 V and À0.8 V pulses with 120 μs pulse width. The ΔW has a strong time correlation with maximum 35% increment at the Δt of À20 μs and À20% reduction at the Δt of 20 μs. These values are relatively smaller than the ΔW maximum value of 60-90% of inorganic memristors. Finally, the ΔW in hybrid device obeys the exponential association with the Δt, namely The measured data can be fitted well.
In addition, the conditioned reflex has been mimicked in hybrid film memristor, similar to the results of Pt/HfO 2 /HfO x /TiN memristor in Figure 8.
By comparison with inorganic bilayer memristors, it can be found that the organic-inorganic hybrid bilayer memristor has similar bio-synaptic functions with comparable switching speed and energy consumption. Moreover organic-inorganic hybrid materials may possess both the advantages of organic and inorganic components with excellent flexibility and tunability. Inorganic compounds have better electrical characteristics and thermal stability. Organic compounds own various functional groups, larger stretchability and low processing temperature. By means of the synergetic and complementary effects between organic and inorganic components, the comprehensive properties of hybrid memristive materials could be expected for significant improvement. The hybrid bilayer ultrathin memristor derived by low temperature MLD/ALD is one competitive candidate for flexible neuroscience applications.

Memristive mechanism
In Section 2.2, we focused on the electrical Performance and synaptic functions of several bilayer ultrathin memristors. In this section, the asymmetric memristive mechanism of the bilayer-structured memristors on TiN or TaN will be studied carefully. Taking Pt/AlO x /HfO x /TiN memristor as an example, the XPS depth profiles of asymmetric bilayer device units were obtained under various resistance states of the initial state, low resistance state (LRS), high resistance state (HRS), and medium resistance state [17]. XPS is a powerful surface analytical tool to determine the chemical valence and the oxygen vacancy contents in multilayer-structured metal oxide thin films [23,25]. Figure 14a-d records the high-resolution Al 2p, Hf 4f and O 1 s peaks in AlO x and HfO x layers for as-deposited Pt/AlO x /HfO x /TiN in the initial state. The Hf 4f spectra from the HfO x layer can be deconvoluted into four peaks (Figure 14b). The stronger peaks at $16.7 eV and 18.6 eV originate from the Hf 4+ in the HfO x layer, whereas the weaker ones with slightly lower energies of 15.6 eV and 17.9 eV come from the Hf (4Àx)+ in the low valence Hf sub-oxide. The content percentage of two Hf valence states in the HfO x layer can be roughly evaluated by calculating the area proportion of each peak, as shown in the inset of Figure 14b [26][27][28]. The percentage of Hf 4+ and Hf (4Àx)+ in the HfO x layer is around 89.7% and 9.4%, respectively. A similar analysis can be also carried out for the Al 2p spectra from AlO x layer (Figure 14a). Meanwhile the O 1 s spectra from the AlO x and HfO x layers can also be deconvoluted into two peaks. The stronger peaks at around 531.5 and 531.0 eV result from Al-O and Hf-O bonding in the AlO x and HfO x layers, respectively, whereas the weaker ones with a slightly higher energy of 532.1 eV in the O 1 s spectra are ascribed to the oxygen vacancies in the AlO x and HfO x layers according to the literature reports [26][27][28][29]. The calculated percentage of oxygen vacancies in the AlO x and HfO x layers is around 0.7% and 8.1%, respectively (Figure 14c and d).
Significantly, the oxygen vacancy content in the HfO x layer is much higher than that in the AlO x layer.
The XPS depth data of Pt/AlO x /HfO x /TiN memristor by Ar ion etching under various resistance states may provide some valuable information on the valence states and defects of metal oxide layers [24], for the initial state sample recorded in Figure 14e. The AlO x /HfO x bilayer structure could be recognized with an evident interfacial diffusion between AlO x /HfO x and HfO x /TiN (gray region). The depth distribution of the average oxygen vacancy concentration in the asymmetric Pt/AlO x /HfO x /TiN memristors under various resistance states of the initial state, LRS, HRS, and medium resistance state is illustrated in Figure 14f. Table 3   of 0 s, 90 s, 210 s, 390 s, and 510 s. Herein A, B, C, D, and E locate in the interface of the Pt/AlO x , AlO x layer, the interface of the AlO x /HfO x , HfO x layer, and the interface of HfO x /TiN, respectively. The oxygen vacancy distribution is inhomogeneous in the Pt/AlO x /HfO x /TiN memristor, and the oxygen vacancy concentration of the interfaces between AlO x /Pt (A), AlO x /HfO x (C), HfO x /TiN (E) is markedly higher than that of the adjacent AlO x (B) and HfO x (D) layers. Furthermore, the oxygen vacancy concentration in HfO x (D) is much higher than that in AlO x (B) layer.
In general, the resistive switching mechanism of metal oxide memristors is related to the connection and rupture of conductive filaments of oxygen vacancies. But the simple increase of oxygen vacancy concentration is not always effective. The non-uniform distribution of oxygen vacancies in memristors is the critical factor affecting the resistive switching behavior of memristive devices [30].
Based on the oxygen vacancy concentration and distribution in the Pt/AlO x / HfO x /TiN memristors under various resistance states in Figure 14f, we proposed a memristive mechanism of an asymmetric bilayer metal oxide synaptic device to explain synaptic plasticity, as illustrated in Figure 15.
There are much more random oxygen vacancies in the HfO x layer than in the AlO x layer for as-deposited Pt/AlO x /HfO x /TiN device. Meanwhile, the oxygen vacancy concentration in the interfaces of AlO x /HfO x and HfO x /TiN is evidently higher than the HfO x layer (Figure 15a). During the forming process, the disorderly distributed oxygen vacancies in the bilayer oxide layers and interfacial layers form conductive filaments under the external electrical field, similar to the soft breakdown of the capacitor. So the connection and disconnection of the conductive filaments lead to resistive switching. When inserting a À3 V forming voltage, the device turns from the initial state to LRS with suddenly resistance drop from 10 MΩ to 600 Ω, suggesting that the oxygen vacancies with positive charges (V O 2+ ) in the AlO x /HfO x interface, HfO x layer, and HfO x /TiN interface move to the AlO x layer and AlO x /Pt interface. Simultaneously, the oxygen vacancy concentration gradient help to the migration of the oxygen vacancies, forming localized conductive filaments of oxygen vacancies in the bilayer structured AlO x /HfO x device (Figure 15b).
After applying the +2.5 V reset voltage to the LRS device, the memristor transfers from LRS (600 Ω) to HRS (1 MΩ) (Figure 15c). During the reset process, the oxygen vacancies migrate from the AlO x /Pt interface and AlO x layer to the AlO x / HfO x interface and HfO x layer, leading to the rupture of oxygen vacancy conductive filaments in the AlO x layer. Besides, considering the thermophoresis/diffusiondriven oxygen migration [31,32], the middle position of the conductive filament in the AlO x layer first breaks up, causing a spatial gap, as indicated by the red arrow in Figure 15c. electron tunneling happens through the physical gap with the enhanced resistance. During the reset process, the oxygen vacancy concentration declines at the AlO x /Pt interface and AlO x layer and rises at the AlO layer, as proved by the relative variation of oxygen vacancy concentrations in Figure 14f and Table 3.
Because the synapse device usually operates under pulse mode to get more intermediate resistance states, a continuous pulse experiment was exploited to alter the device from a LRS (600 Ω) to a medium resistance state (50 kΩ) by imposing 40 pulses (+1.5 V, 10 ms) (Figure 15d). The XPS result (Figure 14f) indicates that the oxygen vacancy concentration curve of medium resistance state devices lies approximately between HRS and LRS in the AlO x layer and HfO x layer. The difference in oxygen vacancy concentration at the interface layers of AlO x /HfO x and HfO x /TiN among the medium resistance state, LRS, and HRS is slightly little. In consequence, during regular operations of synaptic memristors, the formation/ rupture of nanoscale conductive filaments tend to appear in the low k AlO x layer with lower electric field intensity [33]. Furthermore, the device conductance can be modulated by the oxygen vacancy drift under pulse electric field, producing a change in the concentration and distribution of oxygen vacancies at the interface of the metal/oxide and the interior. In the LRS of 600 Ω, the conductive filament is thick with the conductance of 22 G 0 , corresponding to a wide conductive filament with classical metallic properties. After 40 pulse stimuli, a medium resistance state of 50 kΩ is obtained with a conductance of 0.26 G 0 , where the conductive filament behaves as a quantum wire, producing a single-defect conducting path [31,33,34].
The oxygen vacancy migration/diffusion model can be used to explain the transition from STP to LTP in bilayer memristive device (Figure 4a). When imposing the +1.6 V pulse, the oxygen vacancies move from the AlO x layer to the HfO x layer Figure 15. Model of the formation and rupture of a conductive filament consisting of oxygen vacancies. After À3 V forming voltage, the device transfers from an initial resistance state (a) to a low resistance state (LRS) (b); after 2.5 V reset voltage, the device transfers from a low resistance state to a high resistance state (HRS) (c); after 40 continuous pulses (+1.5 V, 10 ms), the device transfers from LRS to a medium resistance state (d) [17].
with the reduced response current. When the voltage is removed, some oxygen vacancies may stay in a new steady position, however some oxygen vacancies may diffuse back to the old position owing to the gradient of oxygen concentration. This leads to the device conductance change with a reduced synaptic weight during the relaxation time. After applying repetitive pulse stimuli, the subsequent voltage forces the reversely diffused oxygen vacancies to move forward again so as to improve the migration efficiency until most oxygen vacancies attain new equilibrium positions. The remaining synaptic weight gradually increases with the increasing pulse number. This process is called repeated training and learning, corresponding to the transformation from STP to LTP [17].
The memristive mechanism from Pt/AlO x /HfO x /TiN device is also applicable to other bilayer-structured memristors such as Pt/HfO 2 /HfO x /TiN, Pt/TiO 2 /Ti-MA/TaN.

Conclusion
Our asymmetric bilayer-structured memristors fabricated by ALD/MLD and their main memristive features are summarized in Table 4, including set/reset voltage, ON/OFF ratio, and some important synaptic functions. Some similar work with asymmetric bilayer structure has also been listed in Table 4 for comparison. It can be seen that all memristors with asymmetric bilayer structure exhibit better resistive switching performance. Our memristors have relatively thinner functional layers, relatively smaller ON/OFF ratio and emulate more artificial synaptic functions such as LTPo, LTD, the transition from STP to LTP, PPF, STDP, and conditional reflex (CR). The memristive mechanism of our bilayer-structured ultrathin device has been proposed to explain the synaptic plasticity based on oxygen vacancies migration/diffusion model. The non-uniform distribution of oxygen vacancies in asymmetric bilayer memristors plays the crucial role in affecting the linkage/ rupture of conductive filaments.
In light of these promising results and the fabrication compatibility with semiconductor industry, the ALD/MLD-derived bi-layer ultrathin memristor devices have tremendous potential as billions of electronic synapses in next-generation artificial neural network and flexible electronics.