Self-Adaptive Spike Voltage-dependent Plasticity Emulated by a Metalloporphyrin-Based Memristor


 By threshold-regulated neural firing and synaptic weight updates in biological neuron–synapse combinations, neural systems can selectively and autonomously encode and process spatiotemporal information. Emulating such an exquisite biological process in electronic devices is a fundamental step toward realizing intelligent neuromorphic systems with self-adaptivity, energy-efficient in-situ edge/parallel computing, and probabilistic inference. Here we report a self-threshold design of prototype artificial axons based on metalloporphyrin, a molecular medium that allows dual electronic/ionic migration in hybrid heterojunction oxide memristors. Threshold behaviors in biological neurons are emulated by introducing metalloporphyrin into alumina-oxide memristors. We show that the memristor achieves smooth, gradual conductive transitions. As a unique feature of such a hybrid system, the endurable current-voltage characteristics of the memristor can be enhanced by altering the metal center to achieve the desired metal–oxygen bonding energy and oxygen migration dynamics. The spike voltage-dependent plasticity is recorded with a positive threshold voltage stemming from the interfacial counterbalance between the vacancy-induced Coulomb force and the external electric field. We further build memristive arrays that directly emulate the self-adaptive and signal-filtering function of the human visual system. These results suggest that the metalloporphyrin platform offers vast opportunities for implementing efficient neural-signal analysis in neuromorphic hardware.

Spike-time-dependent plasticity in neuron-synapse assemblies is essential for cloning human consciousness in artificial neural networks. The fundamental features of spike plasticity are historydependent spatiotemporal correlation 1,2 , probability inference 3 , and in-situ analog-digital mixed-signal processing 4-7 . To unlock the von Neumann bottleneck of neuromorphic computing, synaptic plasticity 5 must be faithfully replicated in artificial synapses. Spike-rate-dependent plasticity (SRDP), the fundamental mode of Hebbian plasticity in learning mechanisms, has thus far been emulated in various artificial synapses based on Chua's memristor concept 8,9 . Another paradigm is spike-timingdependent plasticity (STDP), which has been achieved by training oxide memristors on events with variable time intervals 10, 11 . 10 Despite vast progress, the physical processes in neuromorphic hardware significantly differ from those in actual synapses. These differences limit the fidelity and variety of desired synaptic functions.
Especially, spike plasticity in neuron-synapse links is based on stimulation-generated action potentials.
Physiologically, action potentials are generated by voltage-gated K + /Na + channels, which in turn modulate Ca 2+ channels that change the weights of synapses 12 . Emulating the threshold-regulated 15 plasticity, namely, the spike voltage-dependent plasticity (SVDP), is the inevitable direction of future brain-like computational systems [13][14][15] . For example, SVDP would benefit development of artificial synaptic devices for more realistic neuromorphic visual systems that automatically eliminate background information through an adaptive noise-filtering mechanism 16 . However, achieving such threshold-regulated behaviors by artificial ionic channels in the vertical mode, where a precise ionic 20 shuttle determines the smart selectivity of spiking events 17-19 , has been an insurmountable task.
To achieve threshold-regulated plasticity, the voltage-gated ionic migration dynamics with dominant activation energy must be created in multilayered devices. Oxygen ions in oxide memristive devices are an ideal ionic source, as their activation energy can be tuned by changing vacancy ratios 20 .
They also provide insights into hemoglobin in red blood cells, which transports oxygen along the 25 network of capillaries vasculatures into deep tissues of various organs 21,22 (Fig. 1a). Such regular processes rely on metalloporphyrin (MTPP), a molecule with an iron-coordination center mounted on a planar ring, which reversibly binds oxygen via the coordination bond 23 .
Inspired by these biological events, we reason that MTPPs with dual functionalities (electronic activity and ionic migration) could potentially be the molecular media for implementing adaptable 30 ionic dynamics in memristors. Herein, we report a hybrid heterojunction of MTPP/oxide that facilitates the external field sensitivity of ionic responsiveness and enables artificial synapses with a non-zero threshold voltage property. We show that this unipolar plasticity fundamentally differs from SRDP and STDP, and allows a potentiation voltage at 10 V and a depression voltage at 4 V. By programing different positive-voltage spikes at fixed frequencies and time intervals (Fig. 1b), we achieved neuromorphic systems with a self-adaptive signal-filtering functionality resembling that of human visual recognition systems. 5 The proof-of-concept memristive devices are configured as ITO/ZnTPP (~25 nm)/AlOx (~7 nm)/Al (Fig. 1c, d). The non-stoichiometric AlOx layer with oxygen vacancies is the reservoir of oxygen ions (O 2− ), while the ZnTPP layer adjusts the ionic oxygen migration. As indicated by the energy diagram ( Supplementary Fig. 1), the device preferentially transports holes. In this device configuration, the ZnTPP layer is expected to act as a p-type active molecular framework that to −10 μA. The negative differential resistance peaked at −6.7 V during the first RESET sweep.
The I-V characteristics were further investigated by repeating low-voltage sweep cycles of 0 ↔ 4 V/0 ↔ −4 V (Fig. 2c). In the positive and negative directions, hysteresis loops appeared with ultimate currents decreasing approximately from 3.8-3.0 μA and from −2.00 to −0.96 μA, respectively, indicating that the low voltage caused a continuous transition into a higher resistance state. Therefore, 25 the potentiation and depression in ZnTPP/AlOx memristors can be selectively operated at a unipolar voltage (e.g., 10 V for potentiation and 4 V for depression; see Fig. 2d, e). This voltage thresholdregulated plasticity can be defined as SVDP. Unlike other forms of synaptic plasticity, such as SRDP and STDP, SVDP can be regarded as a relatively independent function of signal processing under a fixed timing interval and rate frequency. In this sense, it resembles the action potential of a neuron and 30 synaptic plasticity under the mechanism of voltage-gated ionic channels. The voltage-threshold feature of the ZnTPP/AlOx device might also realize advanced computing features such as parallel implementation, adaptivity, and probability inference. In addition, the smooth I-V curves were highly repeatable ( Supplementary Fig. 1), suggesting that hybrid ZnTPP/AlOx is a suitable model for advanced neuromorphic computing.
To confirm the SVDP behaviors, we carried out additional measurements in biasing modes (Fig.   2f). The stimulus-response exhibited an inverse trend with a critical value of ~6-7 V. Bias stimuli above (10, 12, 13, or 14 V) and below (1, 2, 3, or 5 V) the critical value induced potentiation and 5 depression, respectively. Hence, the devices can selectively perform potentiation and depression of input signals, resembling those of action potentials at neurons. These processes are reliable and reproducible, suggesting the robustness of this function in artificial neural hardware.
To shed light on the underlying mechanism of plasticity in the ZnTPP/AlOx devices, we performed a series of structural and stoichiometric characterizations with theoretical simulations and 10 device evaluations. The current and hysteresis areas were monotonically increasing functions of the device area (5 × 10 3 µm 2 to 5 × 10 5 µm 2 , Supplementary Fig. 3) and scanning speed (0.01-5 V/s, Supplementary Fig. 4). The current increased from ~74 to 1,500 μA in the high resistance state and from ~454 to 4,048 μA in the low resistance state, while the hysteresis area increased from ~54.9 to 2.0 VμA/cm 2 . These results suggest that device behaviors with the SVDP feature are dominated by 15 homogeneous ionic migration rather than by local filamentary conduction 26 . Conductive filaments can also be excluded by the cross-sectional TEM images and X-ray spectroscopy (STEM-EDX) analyses  30 Zn 2+ and free O 2− in the ZnTPP matrix, respectively 27,28 . XPS analyses also revealed a blue shift (100 meV) in the Zn 2p spectrum after electrical stimulation (Fig. 3c), arising from the formation of coordination bonds between Zn 2+ and O 2− . Coordination bonding with fixed Zn 2+ sites can finely regulate the ion distribution and migration, which are random in general memristive media [29][30][31] .
To further investigate cation-anion interactions, we examined the metal-atom effect of MTPP (M = Zn/Ni/Co/Fe/2H) on the electric properties of the memristors (Supplementary Fig. 8). The metal effects of MTPPs are detailed in Tables S1 and S2. The initial conductance, conductance saturation, 5 and change range strongly depended on the type of MTPP. As the different MTPPs possessed different M-O bonding energies, the results confirm the close correlation between O 2− movement and metal atoms. The ionic migration can thus be described as a hopping process through the MTPP molecular matrix, which consists of reversible M-O binding and dissociation of ionic oxygen from an MTPP, followed by association with another MTPP. This process also repeats in an iterative cycle (Fig. 3d). ZnTPP/AlOx hetero-interface 36 ( Fig. 3e; Supplementary Fig. 9). In contrast, when MTPP/oxide devices are operated below the critical field, a limited number of O 2− ions are freed from the edge of the slipping plane; moreover, within the main release area, the migration speed is quickly reduced by back-scanning owing to the dramatically reduced acceleration. Here, their drift is prevented by the 5 insurmountable Coulomb force between Al 3+ and O 2− (Fig. 3e). We concluded that Coulomb binding energy is the primary regulator of the ionic migration behavior under external energization.
To verify that our devices provide the building blocks for neuromorphic systems, we emulated voltage threshold plasticity rather than SRDP and STDP 37 . To demonstrate the superiority of the MTPP/oxide devices as building blocks for neuromorphic hardware, we conducted a validation study using three 8 × 8 ZnTPP memristor arrays. In our design, the yellow and violet arrays of the word "Hi" represent the input optical image and the memorized electrical pattern, respectively (Fig. 4a).
Assuming sufficient conversion of the photonic signals into electrical signals, the graphic word "Hi" 25 with and without background interference was repeatedly input to the memristor array, and the adaptive filtering capability of the array was evaluated. In the first commonly emulated case without background, only the memristor pixels corresponding to the word were input, along with stimulations (Starget = 10/5 V, W = 100 ms, T1 = 100 ms, T2 = 300 ms). In Case 2, a certain number of images was inputted, and the graphic word 'Hi' in case 2 was memorized and consolidated by a continuous 30 increase of conductance Gaverage from 30.7 to 66.7 μS (Fig. 4b; Supplementary Fig. 13). Figure 4c presents two contrasting cases that affirm the superiority of SVDP in the signal-filtering process.
Analogous to the biological visual nervous system, the array retained a residual impression of signals that were presented earlier, mimicking the persistence of vision. Case 3 was another common emulation case, in which the noise filtering relied only on spontaneous decay. After 30 inputs, the impression level of the word was gradually enhanced from 31.5 to 59.6 μS (Starget = 10/5 V, W = 100 ms, T1 = 100 ms, T2 = 300 ms), whereas the noise pixels without filtering operation spontaneously relaxed to 18.4 μS. Finally, in Case 4, the target "Hi" pixels were operated with 30 identical inputs, 5 but the noise pixels were applied with pulse stimuli (Snoise= 6/5 V, W = 100 ms, T1 = 100 ms, T2 = 300 ms, N = 30) (Supplementary Fig. 13). This case more closely represents the real biological situation than the other cases. The changing conductance values in each pixel are shown in Supplementary   Figures 13 and 14. As expected, the impression of the word "Hi" in Case 4 was homoplastically changed from 32.1 to 58.0 μS, whereas the surrounding background information was considerably 10 weakened to ~7.4 μS. Consequently, the signal-to-noise contrast ratio of Case 4 filtered by SVDP was approximately 8, which was much higher than in Case 3 without the low-voltage training operation.
In      Visualizations of the memorizing processes in an ideal case. c, Noise filtering with and without the 5 SVDP mode. Each pixel represents a single memristor.