The DOI matches its Crossref record; 45 of the 47 references it cites were found in a registry or an index; 13 of the 22 sources its claims cite were checked for support.
Experimental demonstration of highly reliable dynamic memristor for artificial neuron and neuromorphic computing
Open accessAbstract
Neuromorphic computing, a computing paradigm inspired by the human brain, enables energy-efficient and fast artificial neural networks. To process information, neuromorphic computing directly mimics the operation of biological neurons in a human brain. To effectively imitate biological neurons with electrical devices, memristor-based artificial neurons attract attention because of their simple structure, energy efficiency, and excellent scalability. However, memristor's non-reliability issues have been one of the main obstacles for the development of memristor-based artificial neurons and neuromorphic computings. Here, we show a memristor 1R cross-bar array without transistor devices for individual memristor access with low variation, 100% yield, large dynamic range, and fast speed for artificial neuron and neuromorphic computing. Based on the developed memristor, we experimentally demonstrate a memristor-based neuron with leaky-integrate and fire property with excellent reliability. Furthermore, we develop a neuro-memristive computing system based on the short-term memory effect of the developed memristor for efficient processing of sequential data. Our neuro-memristive computing system successfully trains and generates bio-medical sequential data (antimicrobial peptides) while using a small number of training parameters. Our results open up the possibility of memristor-based artificial neurons and neuromorphic computing systems, which are essential for energy-efficient edge computing devices.
Overview
CoChat Summary
TL;DR
The paper reports a transistor-free 1R gradual TiOx memristor cross-bar array with high reliability and uses it to build both a leaky-integrate-and-fire artificial neuron and a neuro-memristive system for sequential-data processing. It claims successful generation of antimicrobial peptide (AMP) candidates with only 50,520 trainable parameters and strong device uniformity, yield, and endurance.
Key findings
- The gradual TiOx memristor showed 1.39% temporal variation during 125 consecutive DC cycles without current compliance, 3.87% spatial variation across 30 randomly selected devices, a rectifying ratio of 10^4, and an on/off ratio larger than 10^3.
- The device achieved endurance of more than 10^6 cycles without degradation, and the array-level decaying time constants across 400 memristors were about 40 ms, supporting uniform short-term memory behavior.
- A memristor-based LIF neuron was demonstrated: with a 47 kΩ resistor and 10 nF capacitor, the neuron fired after 2 presynaptic pulses in one example; with a 94 kΩ resistor it needed 6 consecutive spikes, and with 23.5 kΩ it fired within 2 spikes.
- In AMP sequence generation, the system used 1554 AMP sequences for training and a 20 × 20 memristor array; generated sequences from the mixed W/I-ratio case were predicted as AMP 84.11% of the time, compared with 74.07% and 79.21% for the short-term and long-term cases.
Methods
- Fabricated a 20 × 20 transistor-free Pt/gradual TiOx/Ti cross-bar array using lithography, e-beam deposition, anodizing, and etching; electrical testing used a Keithley 4200A-SCS for I–V curves and a DAQ/preamplifier setup for pulse experiments.
- For the artificial neuron, combined the volatile gradual TiOx memristor with a parallel capacitor and serial resistor to emulate LIF behavior; tested frequency dependence, spike amplitude/width dependence, and device-to-device/cycle-to-cycle uniformity.
- For sequential data, trained a neuro-memristive simulator on 1554 AMP sequences from DADP, APD, and ADAM, using four W/I ratios (5/95, 10/90, 15/85, 20/80) and a readout MLP of 80 × 500 × 20; the system was trained for 50 epochs.
Limitations
- The paper itself states the device has short-term memory and “cannot serve as a long-term memory,” so its function is limited to leaky-integrator style processing rather than persistent storage.
- The text makes evident that the work is primarily a demonstration on a specific AMP sequence task and on device/circuit-level experiments; broader real-world benchmark evaluation is not stated in the paper.
- The paper notes that the AMP study uses a simulator and readout training rather than a fully integrated end-to-end hardware-only system; this is evident from the text.
Model-written summary. Check it against the paper before relying on it.
Integrity signals
Registry
Cites
Statistics
Figures

![Fig. 2 Effects of gradual oxygen concentration in TiOx switching layer. a A cross-sectional TEM image of the gradual TiOx memristor. b TOF-SIMS depth profile results of anodized TiOx. c TOF-SIMS depth profile results of sputtered TiO2. d A schematic representation of the switching mechanism of the gradual TiOx memristor. e Size dependency of the gradual TiOx memristor. The memristors with four different sizes (5 × 5, 10 × 10, 20 × 20, and 50 × 50 [µm2]) are fabricated on the same wafer. The current is measured at 1 V at a low resistance state (LRS) during the linear voltage sweep from 0 V to 4 V. f An activation energy (EA) calculated from the time constant under various temperatures. The time constant decreased when the temperature increased, and the activation energy calculated from the change of the decay time along the temperature implies that the resistive switching and short-term dynamics originate from the oxygen anions in TiOx.](/paper-figures/867f4f9fd5c7459b8c3031572557959d/experimental-demonstration-of-highly-reliable-dynamic-memristor-for-artificial-figure-2.jpg)

![Fig. 4 Binary and sequence input processing of the gradual TiOx memristor. a Response of 24 gradual TiOx memristors to the 4-bit inputs. The memristors are randomly selected from a 20×20 cross-bar array. Seven different binary inputs are used (brown: [1 1 1 1], yellow: [1 1 0 1], red: [0 0 0 1], sky blue: [1 1 1 0], blue: [0 0 1 0], green: [1 0 0 0], and black: [0 0 0 0]). For the input “1”, the set pulse (4.3 V, 50 µs) is used while the ground is maintained for “0”. The read pulse amplitude and width are 1.5 V and 200 µs, respectively. The gradual TiOx memristor successfully separates different binary inputs at the last time step (4), and all devices show similar responses without severe device-to-device variations. b Experimental results of memristors’ output separating two sequences having different orders. Memristors for past elements have smaller output currents than memristors for recent elements. c Schematic of the way how a sequence input (D D D C C B D C C D B A D B A C D C B A) is applied to memristors. The sequence is transformed into a pulse train and applied to the assigned memristors in order. The unit length of a single pulse (or ground) is 100 µs. The set pulse amplitude is 4 V, but only a portion (duty cycle) of the 100 µs is elevated to 4 V; the rest of the 100 µs is grounded. The read pulse (1.5 V, 100 µs) is applied between every set pulse and rest. d–f Memristors’ output through the whole input sequence in a short-term, middle-term, and long-term memristors. The final results of each duty cycle case show the different processing abilities of memristors with different duty cycles. The memristors process sequences with a longer time domain along with an increase in the width/interval (W/I) ratio. The output of the long-term memristors only contains frequency information of each alphabet, while the order information is clearly shown in the short-term memristors. The sequence with the red box below the figure represents the range of the sequence that might affect the final output. All of the results in Fig. 4 are measured from a 20 × 20 cross-bar array.](/paper-figures/40d25403c64f4f04ab8df1f0d4f9689b/experimental-demonstration-of-highly-reliable-dynamic-memristor-for-artificial-figure-4.jpg)

Verification
Claims the paper makes
4 Valid · 10 Not valid · 5 Not checked · 94 Own results and interpretations
The cited sources support only part of this claim — read each source’s verdict below.
In the paper
“This is because H2O molecules, used during the anodizing process, significantly reduce the activation energy19.”
§ Experimental demonstration of highly reliable dynamic memristor for artificial neuron and neuromorphic computing > Results > Insulator thickness modulation in gradual TiOx A cross-sectional transmission electron microscopy (TEM) image of a gradual TiOx memristor is shown in Fig. 2a. A thin gradual TiOx layer (about 30 nm) is sandwiched by the top electrode (TE) and the bottom electrode (BE), showing a simple structure of the developed memristor. As shown in Fig. 2a, the anodized TiOx layer does not have any porous features, even though the anodizing technique is widely used to form nanoporous TiO2 by anodizing the Ti film in a fluoride-rich electrolyte. The time of flight secondary ion mass spectrometer (TOF-SIMS) results reveal the difference between the anodized TiOx layer and the radio frequency (RF)-sputtered TiO2 layer. As shown in Fig. 2b, for an anodized TiOx layer, the oxygen concentration and TiO2 concentration gradually decrease from the top (Pt/TiOx interface) to the bottom (TiOx/Ti interface), while the Ti concentration increases. However, these concentration gradients are not observed for the sputtered TiO2 layer (see Fig. 2c). This is because the anodizing process is a top-down oxidation process; therefore, the upper Ti layer undergoes longer oxidation and becomes an oxygen-rich insulating TiO2 layer, while the bottom Ti layer experiences a shorter oxidation time and forms an oxygen-deficient metallic TiOx layer. This oxygen gradient effectively controls the thickness of the insulating region and results in resistive switching with a large on/off ratio of the anodizing TiOx layer, while the sputtered TiO2 layer cannot effectively change the thickness of the insulating region due to uniform oxygen concentration and shows a low on/off current ratio and a non-uniform switching characteristic (see Supplementary Fig. 5). To be more specific, the oxygen vacancy gradient enables resistive switching with a self-decaying property (see Fig. 2d). When a positive bias is applied to TE while grounding BE, oxygen anions shift from TiOx near BE to TiOx near TE. TiOx near BE loses its oxygen and the metallic TiOx region increases. Therefore, the effective thickness of insulating TiOx decreases, the reduced effective insulating thickness increases the overall device’s conductivity, and the device becomes a low resistance state (LRS). However, when a negative bias is applied to TE, oxygen anions in oxygen-rich TiOx near TE shift to TiOx near BE. The effective insulating thickness expands again, and the device becomes a high resistance state (HRS) (see Supplementary Fig. 6). The LRS to HRS transition (reset) occurs without external stimulus due to the diffusion of oxygen anions, which results in short-term (volatile) memory effects with about 40 ms time constant. To demonstrate that the device switches its resistance state through the oxygen anion shift in the overall TiOx bulk, the currents from memristors of different sizes are compared, as shown in Fig. 2e. During the linear voltage sweep, the current of the memristor at 1 V (LRS) is read. The results show that the current increases as the size increases. When the memristor operates through a conductive filament, the memristor does not show size dependency, as the locally formed filament conducts almost all of the current regardless of the device’s size. However, when the memristor switches its resistance through the overall bulk, the size affects the current. The size dependency in the results verifies that the gradual TiOx memristor operates by modulating the oxygen anion distribution in the overall TiOx bulk. In Supplementary Fig. 7, the device current reduces as the device size decreases due to the non-filamentary nature of the device. Moreover, it does not significantly change other properties such as decaying speed and potentiation ratio, which means that the device is scalable without performance degradation. Furthermore, by calculating activation energy (EA) from the decaying time constants under several elevated temperatures from 323 K to 423 K, we have proven that the oxygen anion movement across the whole TiOx bulk dominates resistive switching (see Fig. 2f). The calculated activation energy (EA) is about 0.21 eV, which is similar to the activation energy of the ionized oxygen vacancy in polycrystalline H2O-rich TiO219. The calculated activation energy of the gradual TiOx memristor is much smaller than the oxygen diffusion activation energy (1.05 eV) in rutile TiO220. This is because H2O molecules, used during the anodizing process, significantly reduce the activation energy19. Moreover, in the anodized TiOx layer, it is observed that the upper region of the TiOx has a fully amorphous structure while the bottom region has a polycrystalline structure, because polycrystalline metallic Ti concentration is much higher in the bottom region of the TiOx layer (see Supplementary Figs. 8, 9). The oxygen activation energy is low in the case of the amorphous phase21, thus the gradual TiOx layer has much smaller oxygen anion activation energy compared to the rutile TiO2 layer. This small activation energy allows the oxygen anions to easily move along the bias and induces resistive switching with a large on/off ratio and short-term memory effect. All of these results demonstrate that resistance switching in gradual TiOx is induced by the movement of oxygen anions across the whole TiOx bulk, unlike other filamentary-type memristors.
What the sources say
- Partly supportedcites [19] Hossein-Babaei, F. & Alaei-Sheini, N. Electronic conduction in Ti/Poly-TiO2/Ti structures. Sci. Rep. 6, 1–11 (2016). sourceExperimental — accuracy not yet measured
The Results and Discussion section reports that H2O-assisted oxygen-vacancy motion on rutile surfaces has a lower activation energy. But the Methods > Fabrication section says the TiO2 layers were made by thermal oxidation, not anodizing; the paper examines adsorbed water and atmospheric humidity, not H2O used during an anodizing process. (LLM, conf=0.89)
plainstates a cause- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Partly supported
Every source the paper cites for this claim was checked and supports it.
In the paper
“The calculated activation energy (EA) is about 0.21 eV, which is similar to the activation energy of the ionized oxygen vacancy in polycrystalline H2O-rich TiO219.”
§ Experimental demonstration of highly reliable dynamic memristor for artificial neuron and neuromorphic computing > Results > Insulator thickness modulation in gradual TiOx A cross-sectional transmission electron microscopy (TEM) image of a gradual TiOx memristor is shown in Fig. 2a. A thin gradual TiOx layer (about 30 nm) is sandwiched by the top electrode (TE) and the bottom electrode (BE), showing a simple structure of the developed memristor. As shown in Fig. 2a, the anodized TiOx layer does not have any porous features, even though the anodizing technique is widely used to form nanoporous TiO2 by anodizing the Ti film in a fluoride-rich electrolyte. The time of flight secondary ion mass spectrometer (TOF-SIMS) results reveal the difference between the anodized TiOx layer and the radio frequency (RF)-sputtered TiO2 layer. As shown in Fig. 2b, for an anodized TiOx layer, the oxygen concentration and TiO2 concentration gradually decrease from the top (Pt/TiOx interface) to the bottom (TiOx/Ti interface), while the Ti concentration increases. However, these concentration gradients are not observed for the sputtered TiO2 layer (see Fig. 2c). This is because the anodizing process is a top-down oxidation process; therefore, the upper Ti layer undergoes longer oxidation and becomes an oxygen-rich insulating TiO2 layer, while the bottom Ti layer experiences a shorter oxidation time and forms an oxygen-deficient metallic TiOx layer. This oxygen gradient effectively controls the thickness of the insulating region and results in resistive switching with a large on/off ratio of the anodizing TiOx layer, while the sputtered TiO2 layer cannot effectively change the thickness of the insulating region due to uniform oxygen concentration and shows a low on/off current ratio and a non-uniform switching characteristic (see Supplementary Fig. 5). To be more specific, the oxygen vacancy gradient enables resistive switching with a self-decaying property (see Fig. 2d). When a positive bias is applied to TE while grounding BE, oxygen anions shift from TiOx near BE to TiOx near TE. TiOx near BE loses its oxygen and the metallic TiOx region increases. Therefore, the effective thickness of insulating TiOx decreases, the reduced effective insulating thickness increases the overall device’s conductivity, and the device becomes a low resistance state (LRS). However, when a negative bias is applied to TE, oxygen anions in oxygen-rich TiOx near TE shift to TiOx near BE. The effective insulating thickness expands again, and the device becomes a high resistance state (HRS) (see Supplementary Fig. 6). The LRS to HRS transition (reset) occurs without external stimulus due to the diffusion of oxygen anions, which results in short-term (volatile) memory effects with about 40 ms time constant. To demonstrate that the device switches its resistance state through the oxygen anion shift in the overall TiOx bulk, the currents from memristors of different sizes are compared, as shown in Fig. 2e. During the linear voltage sweep, the current of the memristor at 1 V (LRS) is read. The results show that the current increases as the size increases. When the memristor operates through a conductive filament, the memristor does not show size dependency, as the locally formed filament conducts almost all of the current regardless of the device’s size. However, when the memristor switches its resistance through the overall bulk, the size affects the current. The size dependency in the results verifies that the gradual TiOx memristor operates by modulating the oxygen anion distribution in the overall TiOx bulk. In Supplementary Fig. 7, the device current reduces as the device size decreases due to the non-filamentary nature of the device. Moreover, it does not significantly change other properties such as decaying speed and potentiation ratio, which means that the device is scalable without performance degradation. Furthermore, by calculating activation energy (EA) from the decaying time constants under several elevated temperatures from 323 K to 423 K, we have proven that the oxygen anion movement across the whole TiOx bulk dominates resistive switching (see Fig. 2f). The calculated activation energy (EA) is about 0.21 eV, which is similar to the activation energy of the ionized oxygen vacancy in polycrystalline H2O-rich TiO219. The calculated activation energy of the gradual TiOx memristor is much smaller than the oxygen diffusion activation energy (1.05 eV) in rutile TiO220. This is because H2O molecules, used during the anodizing process, significantly reduce the activation energy19. Moreover, in the anodized TiOx layer, it is observed that the upper region of the TiOx has a fully amorphous structure while the bottom region has a polycrystalline structure, because polycrystalline metallic Ti concentration is much higher in the bottom region of the TiOx layer (see Supplementary Figs. 8, 9). The oxygen activation energy is low in the case of the amorphous phase21, thus the gradual TiOx layer has much smaller oxygen anion activation energy compared to the rutile TiO2 layer. This small activation energy allows the oxygen anions to easily move along the bias and induces resistive switching with a large on/off ratio and short-term memory effect. All of these results demonstrate that resistance switching in gradual TiOx is induced by the movement of oxygen anions across the whole TiOx bulk, unlike other filamentary-type memristors.
What the sources say
- Supportedcites [19] Hossein-Babaei, F. & Alaei-Sheini, N. Electronic conduction in Ti/Poly-TiO2/Ti structures. Sci. Rep. 6, 1–11 (2016). source
The Results and Discussion section describes polycrystalline TiO2 with grain boundaries and reports an activation-energy range of 0.2–0.5 eV for H2O-assisted oxygen-vacancy motion on rutile facets. The citing paper’s calculated 0.21 eV is near the lower end of that range; the source does not itself report 0.21 eV. (LLM, conf=0.91)
plainstates an association- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Supported
The cited sources support only part of this claim — read each source’s verdict below.
In the paper
“These forming-free and compliance current-free properties of gradual TiOx memristors make the device more suitable for neuromorphic hardware in edge devices with simple circuit designs18.”
§ Experimental demonstration of highly reliable dynamic memristor for artificial neuron and neuromorphic computing > Results > Characteristics of gradual TiOx-based memristor The memristive device with gradual TiOx layer is fabricated via an anodizing process (see “Methods” section), as shown in Fig. 1a. The device shows only a 1.39% temporal variation (σ/μ) during 125 consecutive DC cycles without current compliance (see Fig. 1b). The I–V curve also shows that the device has a high rectifying ratio of 104 and high on/off ratio larger than 103. This high rectifying ratio (current at Vread/current at −Vread) prevents the selector-less cross-bar array from the sneak path current problems (see Supplementary Fig. 1), and the high on/off ratio (>103) reduces the peripheral circuit burden to discriminate the conductance change, which are the major bottlenecks in conventional memristor cross-bar arrays. Because of the forming-free nature of the device, where the forming process usually damages the devices’ performance17, the device in the memristor cross-bar array achieves high endurance of more than 106 cycles without any degradation of the electrical performance (see Fig. 1c and Supplementary Fig. 2). These forming-free and compliance current-free properties of gradual TiOx memristors make the device more suitable for neuromorphic hardware in edge devices with simple circuit designs18. The gradual TiOx memristor shows exceptionally uniform resistive switching with only a 3.87% spatial variation (σ/μ) from 30 randomly selected devices, as shown in Fig. 1d. The memristor device also exhibits a low temporal variation of about 1.67% (σ/μ) during 50 cycles of 200 consecutive set pulses (4.5 V, 10 μs) and read pulses (1.5 V, 100 μs) (see Fig. 1e). These results show the high uniformity of the developed device, which has been a significant bottleneck for various memristor devices. The uniform 40 ms time constant (τ) during the self-decaying process is also achieved with a low temporal variation of 3.86% (σ/μ), which is one of the key points for a reliable artificial neuron and neuro-memristive computing system. Finally, all 400 devices from a 20 × 20 cross-bar array (see Supplementary Fig. 3) are measured to show spatial uniformity in terms of time constant and conductance change as a function of the number of applied pulses. The uniformity of the decaying time constants from a cross-bar array is shown in Fig. 1f. The time constants measured from the 400 memristors are ~40 ms, which is similar to the time constant calculated from the stand-alone device shown in Fig. 1e. This means that all the devices in an array operate uniformly, and due to the self-rectifying behavior, the read process is not disturbed by the sneak path problem. Even though the device cannot serve as a long-term memory due to the short-term characteristics, the short-term memory characteristic with excellent uniformity makes the developed memristor suitable for artificial neuron and neuro-memristive computing. As shown in Fig. 1g, the device changes its conductance from 55 nS to 1200 nS (on average) with small spatial variations by applying 16 write pulses. To show the potentiation of the memristors with only 16 pulses, a longer set pulse (4.5 V, 100 µs) is used, and the memristor conductance is measured by read pulses (2 V, 100 µs). The conductance map as a function of the device’s location in the optical image of the array shows uniform distribution conductance change, as shown in Fig. 1h. The conductance changes during the 16 pulses are represented in Supplementary Fig. 4.
What the sources say
- Partly supportedcites [18] Ng, S., John, R. A., Yang, J. & Mathews, N. Forming-less compliance-free multistate memristors as synaptic connections for brain-inspired computing. ACS Appl. Electron. Mater. 2, 817–826 (2020). sourceExperimental — accuracy not yet measured
The Conclusion (page unavailable) reports gradual, forming-free, compliance-current-free switching suitable for brain-inspired hardware; the Abstract says avoiding control transistors reduces peripheral-circuit complexity. But the device studied uses SnOx (tin oxide), not TiOx, and the cited passages do not establish suitability specifically for edge devices. (LLM, conf=0.90)
plainstates a cause- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Partly supported
The cited sources that could be checked couldn’t confirm or dispute this claim.
In the paper
“A real AMP and the generated sequence from our neuro-memristive computing system both show the similar structure and composition of hydrophobicity and hydrophilicity, which is one of the key points of the AMP.”
§ Experimental demonstration of highly reliable dynamic memristor for artificial neuron and neuromorphic computing > Results > Training and generation of neuro-memristive computing for sequential data AMPs are natural antibiotics that exist in living organisms and act as antimicrobials by bursting the epidermis of bacteria or eukaryotes (see Fig. 5a). AMPs have an important role in the immune system of several living organisms, and many studies have been conducted regarding their use for development of new drugs for various diseases such as SARS and cancer30,31. Since AMPs are composed of one short peptide, they have the advantage of being able to mass-produce if their amino-acid sequences are known. The amphipathicity in which hydrophilic groups and hydrophobic groups in a peptide coexist in balance makes the AMP burst the bacteria. The AMPs are aggregated and inserted into the bacterial epidermis (surface of the bacteria). After that, because the bacterial epidermis is hydrophobic while the environment is hydrophilic, the hydrophobic side of the AMP is attached to the bacterial epidermis. The hydrophilic side of AMPs are faced with each other, and then the aggregated AMPs finally form a pore structure across the bacterial epidermis. Because of the pore consisting of AMPs, the cytoplasm of the bacteria flows outside and the bacteria is killed. This cell death process induced by AMPs is called bacterial lysis. However, the AMP should have a special structure, which is formed through complex, high-dimensional basis amino-acid grammar32, as well as be amphipathic to interact with the bacterial epidermis and burst it. Therefore, it rarely makes the AMP by simply concatenating hydrophobic and hydrophilic short amino-acid sequences. In previous studies, it has been verified that LSTM-based neural networks could learn AMP’s amino-acid grammar and create new potential AMP sequences32,33. However, LSTMs have too many trainable parameters and require too much complex computing to insert the whole network into an edge processor. Instead of LSTM, we propose a neuro-memristive computing system based on the gradual TiOx memristor to generate new, potential AMP sequences. To simulate the neuro-memristive computing based on the measured data from memristor arrays, we design an entire system on a circuit level (see Supplementary Fig. 19). The designed system’s operation is conceptually represented in Supplementary Fig. 20. Based on the design, the simulation of the neuro-memristive computing having four different W/I ratios (5/95, 10/90, 15/85, and 20/80) was tested for AMP generation. First, the input sequence is transformed into a 20-dimensional (or 20-channel) vector having the same time step to the original sequence data. Each memristor in the cross-bar array processes the spike train from one channel (one amino-acid) among the 20 channels in the transformed input data (see Fig. 5b). In our system design, the system consists of four groups of memristors where each group has different W/I ratios to process the given input data differently (see Supplementary Figs. 19, 20, and “Methods”). To train the neuro-memristive computing system to generate new AMPs, the training data is pre-processed first. The target data to train the weights in the readout function is generated from the training data. For example, for the input sequence in Fig. 5b, the target data is made by removing the first amino-acid “G” and adding an end cursor “@” at the end of the sequence. Thus, the target data becomes “IGKFLHSAKKFGKAFVGEIMNS@”. The end cursor is used to inform the sequence end. The first amino-acid, which is “G”, enters the memristor cross-bar array. Every column line represents each amino-acid used for AMP, and each row line represents a group of memristors having the same W/I ratio, as shown in Supplementary Figs. 19, 20. When the input amino-acid is “G”, the column line for “G” is grounded and the other columns are floating. Then, the voltage pulses with various W/I ratios are transported from the row line and potentiate the memristors in column “G”. After the potentiation of memristors for “G”, a read stage is performed by applying the read voltage to the row lines while the column lines are grounded. The memristor cross-bar array output has 1 × 80 size because each group of memristors makes a 1 × 20 output and there are four groups. The cross-bar array output is an input of the readout function, and the readout function trains the weights based on the discrepancy between the readout function output and the corresponding target, which is the next amino-acid “I” in the given example. By repeating this process until the end cursor enters the system, the neuro-memristive computing system can learn the amino-acid grammar for the AMPs. To compare the effect of using various W/I ratios, the four sequence sets were compared (a training set, generation sets from a short-term (low W/I ratio), long-term (high W/I ratio), and a mixed case(various W/I ratios)). The short-term, long-term, and the mixed case have the same readout function network size and the same number of trainable parameters to fairly compare the effect of mixed W/I ratios for data processing without the effect of network size. The result was evaluated by the generated output sequences’ several features (probability of being AMP, global charge, molecular weights, and amino-acid distribution) using an AMP prediction tool34 and a Python package for AMP analysis35. Figure 5c–f show the comparison among the training set and generation sets from short-term, long-term, and mixed case. The AMP prediction results in Fig. 5c show that 74.07% and 79.21% of sequences in a generation set from the short-term and long-term case, respectively, are predicted as AMP. Then, 84.11% of sequences from mixed case are predicted as AMP. All probability values of AMP in each set are represented in Supplementary Fig. 21. The average global charge of peptides representing the overall charge of amino-acid residues in each sequence36 of the generation set from the mixed case is 2.32 pKa while the sequences from the short-term and long-term cases show 3.89 pKa and 3.11 pKa, respectively (see Fig. 5d). The average global charge value from the mixed case is similar to the training set (2.28 pKa), which means the neuro-memristive computing system with mixed W/I ratios can learn the complex pattern of the given data set. The molecular weights and the amino-acid distribution of the generated sequences from the mixed case are also similar to the training set, as shown in Fig. 5e, f. The different dynamic responses of memristors for different W/I ratios help the neuro-memristive computing system learn the complex amino-acid grammar and how to construct amphipathic peptides by transforming the input data into the higher dimension basis and making the training of the readout function effective. Finally, the 3D peptide structures of an AMP sequence in the training set and a generated sequence with the highest probability of being AMP are shown in Fig. 5g, h. A real AMP and the generated sequence from our neuro-memristive computing system both show the similar structure and composition of hydrophobicity and hydrophilicity, which is one of the key points of the AMP. In summary, our neuro-memristive computing system composed of a gradual TiOx memristor cross-bar array demonstrates the potential of memristor-based computing for sequence type data by successfully generating new AMPs with a small number of parameters (only 50,520 parameters). All of these results reveal that the gradual TiOx memristor-based neuro-memristive computing system has great benefits regarding sequence type input data.
What the sources say
- Could not verifycites [41] Alland, C. et al. RPBS: a web resource for structural bioinformatics. Nucleic Acids Res. 33, 44–49 (2005). sourceExperimental — accuracy not yet measured
The AUTOMAT section discusses sequence-similarity searches, and the References mention hydrophobic cluster analysis, but neither addresses the claimed comparison between a real AMP and a neuro-memristive-generated sequence. Searches for antimicrobial peptides, the generated sequence, and hydrophobicity/hydrophilicity found no passage in our copy supporting that comparison. (LLM, conf=0.78)
- Could not verifycites [42] Néron, B. et al. Mobyle: a new full web bioinformatics framework. Bioinformatics 25, 3005–3011 (2009). sourceExperimental — accuracy not yet measured
The System Overview (§3.5; page unavailable) describes a protein-modelling workflow using a sunflower genome sequence, not a comparison between an antimicrobial peptide and a generated sequence. Searches for neuro-memristive computing, antimicrobial peptides, hydrophobicity, and hydrophilicity found no passage addressing the claimed similarity. (LLM, conf=0.78)
plainstates an associationSingle case- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Could not verify
The cited sources that could be checked couldn’t confirm or dispute this claim.
In the paper
“The average global charge of peptides representing the overall charge of amino-acid residues in each sequence36 of the generation set from the mixed case is 2.32 pKa while the sequences from the short-term and long-term cases show 3.89 pKa and 3.11 pKa, respectively (see Fig. 5d).”
§ Experimental demonstration of highly reliable dynamic memristor for artificial neuron and neuromorphic computing > Results > Training and generation of neuro-memristive computing for sequential data AMPs are natural antibiotics that exist in living organisms and act as antimicrobials by bursting the epidermis of bacteria or eukaryotes (see Fig. 5a). AMPs have an important role in the immune system of several living organisms, and many studies have been conducted regarding their use for development of new drugs for various diseases such as SARS and cancer30,31. Since AMPs are composed of one short peptide, they have the advantage of being able to mass-produce if their amino-acid sequences are known. The amphipathicity in which hydrophilic groups and hydrophobic groups in a peptide coexist in balance makes the AMP burst the bacteria. The AMPs are aggregated and inserted into the bacterial epidermis (surface of the bacteria). After that, because the bacterial epidermis is hydrophobic while the environment is hydrophilic, the hydrophobic side of the AMP is attached to the bacterial epidermis. The hydrophilic side of AMPs are faced with each other, and then the aggregated AMPs finally form a pore structure across the bacterial epidermis. Because of the pore consisting of AMPs, the cytoplasm of the bacteria flows outside and the bacteria is killed. This cell death process induced by AMPs is called bacterial lysis. However, the AMP should have a special structure, which is formed through complex, high-dimensional basis amino-acid grammar32, as well as be amphipathic to interact with the bacterial epidermis and burst it. Therefore, it rarely makes the AMP by simply concatenating hydrophobic and hydrophilic short amino-acid sequences. In previous studies, it has been verified that LSTM-based neural networks could learn AMP’s amino-acid grammar and create new potential AMP sequences32,33. However, LSTMs have too many trainable parameters and require too much complex computing to insert the whole network into an edge processor. Instead of LSTM, we propose a neuro-memristive computing system based on the gradual TiOx memristor to generate new, potential AMP sequences. To simulate the neuro-memristive computing based on the measured data from memristor arrays, we design an entire system on a circuit level (see Supplementary Fig. 19). The designed system’s operation is conceptually represented in Supplementary Fig. 20. Based on the design, the simulation of the neuro-memristive computing having four different W/I ratios (5/95, 10/90, 15/85, and 20/80) was tested for AMP generation. First, the input sequence is transformed into a 20-dimensional (or 20-channel) vector having the same time step to the original sequence data. Each memristor in the cross-bar array processes the spike train from one channel (one amino-acid) among the 20 channels in the transformed input data (see Fig. 5b). In our system design, the system consists of four groups of memristors where each group has different W/I ratios to process the given input data differently (see Supplementary Figs. 19, 20, and “Methods”). To train the neuro-memristive computing system to generate new AMPs, the training data is pre-processed first. The target data to train the weights in the readout function is generated from the training data. For example, for the input sequence in Fig. 5b, the target data is made by removing the first amino-acid “G” and adding an end cursor “@” at the end of the sequence. Thus, the target data becomes “IGKFLHSAKKFGKAFVGEIMNS@”. The end cursor is used to inform the sequence end. The first amino-acid, which is “G”, enters the memristor cross-bar array. Every column line represents each amino-acid used for AMP, and each row line represents a group of memristors having the same W/I ratio, as shown in Supplementary Figs. 19, 20. When the input amino-acid is “G”, the column line for “G” is grounded and the other columns are floating. Then, the voltage pulses with various W/I ratios are transported from the row line and potentiate the memristors in column “G”. After the potentiation of memristors for “G”, a read stage is performed by applying the read voltage to the row lines while the column lines are grounded. The memristor cross-bar array output has 1 × 80 size because each group of memristors makes a 1 × 20 output and there are four groups. The cross-bar array output is an input of the readout function, and the readout function trains the weights based on the discrepancy between the readout function output and the corresponding target, which is the next amino-acid “I” in the given example. By repeating this process until the end cursor enters the system, the neuro-memristive computing system can learn the amino-acid grammar for the AMPs. To compare the effect of using various W/I ratios, the four sequence sets were compared (a training set, generation sets from a short-term (low W/I ratio), long-term (high W/I ratio), and a mixed case(various W/I ratios)). The short-term, long-term, and the mixed case have the same readout function network size and the same number of trainable parameters to fairly compare the effect of mixed W/I ratios for data processing without the effect of network size. The result was evaluated by the generated output sequences’ several features (probability of being AMP, global charge, molecular weights, and amino-acid distribution) using an AMP prediction tool34 and a Python package for AMP analysis35. Figure 5c–f show the comparison among the training set and generation sets from short-term, long-term, and mixed case. The AMP prediction results in Fig. 5c show that 74.07% and 79.21% of sequences in a generation set from the short-term and long-term case, respectively, are predicted as AMP. Then, 84.11% of sequences from mixed case are predicted as AMP. All probability values of AMP in each set are represented in Supplementary Fig. 21. The average global charge of peptides representing the overall charge of amino-acid residues in each sequence36 of the generation set from the mixed case is 2.32 pKa while the sequences from the short-term and long-term cases show 3.89 pKa and 3.11 pKa, respectively (see Fig. 5d). The average global charge value from the mixed case is similar to the training set (2.28 pKa), which means the neuro-memristive computing system with mixed W/I ratios can learn the complex pattern of the given data set. The molecular weights and the amino-acid distribution of the generated sequences from the mixed case are also similar to the training set, as shown in Fig. 5e, f. The different dynamic responses of memristors for different W/I ratios help the neuro-memristive computing system learn the complex amino-acid grammar and how to construct amphipathic peptides by transforming the input data into the higher dimension basis and making the training of the readout function effective. Finally, the 3D peptide structures of an AMP sequence in the training set and a generated sequence with the highest probability of being AMP are shown in Fig. 5g, h. A real AMP and the generated sequence from our neuro-memristive computing system both show the similar structure and composition of hydrophobicity and hydrophilicity, which is one of the key points of the AMP. In summary, our neuro-memristive computing system composed of a gradual TiOx memristor cross-bar array demonstrates the potential of memristor-based computing for sequence type data by successfully generating new AMPs with a small number of parameters (only 50,520 parameters). All of these results reveal that the gradual TiOx memristor-based neuro-memristive computing system has great benefits regarding sequence type input data.
What the sources say
- Could not verifycites [36] Phoenix, D. A. & Harris, F. The hydrophobic moment and its use in the classification of amphiphilic structures (review). Mol. Membr. Biol. 19, 1–10 (2002). sourceExperimental — accuracy not yet measured
The review’s Surface active α-helices section discusses the distribution of charged residues, but it does not report an average global charge for a mixed-case peptide generation set. Searches for “2.32 pKa” and the short-term and long-term charge values found no matching result in our copy. (LLM, conf=0.76)
plain- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Could not verify
67 more claims · 6 with their cited sources checked
Create a free account to see all 113 claims and every source verdict.
- It may state two separate claims as one. 31
- The quote only partly states this claim. 2
- Probably not a substantial claim. 10
Claims other papers make about this one
Each is checked against this paper’s own text: does it say what the citing paper says it does?
The source backs part of the claim, not all of it — or backs it for a different population, condition or comparison.
In the citing paper
“Based on the desirable characteristics of the memristor crossbar array, various studies have demonstrated that the energy efficiency, hardware size, and inference speed can be effectively improved compared with the conventional CMOS-based processors 12,17,18”
From Linear conductance update improvement of CMOS-compatible second-order memristors for fast and energy-efficient training of a neural network using a memristor crossbar array (2023) · cites this paper as [18]
Experimental — accuracy not yet measuredThe source backs part of the claim, not all of it — or backs it for a different population, condition or comparison.
In the citing paper
“For non-identical pulse methods, an extra read operation is first required to check the device's current conductance state.”
From Linear conductance update improvement of CMOS-compatible second-order memristors for fast and energy-efficient training of a neural network using a memristor crossbar array (2023) · cites this paper as [18]
Experimental — accuracy not yet measuredThe source backs part of the claim, not all of it — or backs it for a different population, condition or comparison.
In the citing paper
“After reading the device's conductance, a proper pulse for the current device conductance must be selected to decide the amplitude or length of the pulse.”
From Linear conductance update improvement of CMOS-compatible second-order memristors for fast and energy-efficient training of a neural network using a memristor crossbar array (2023) · cites this paper as [18]
Experimental — accuracy not yet measuredThe check wasn’t confident either way, so it gives no verdict — never disputed.
In the citing paper
“The read operation and the pulse selecting process, however, consume significant energy and time because an energy-hungry integrator and analog-to-digital converter (ADC) circuits are used for the read operation, and complex peripheral circuits are required for the selecting process.6,18,46”
From Linear conductance update improvement of CMOS-compatible second-order memristors for fast and energy-efficient training of a neural network using a memristor crossbar array (2023) · cites this paper as [18]
Experimental — accuracy not yet measured
Citations
Cites
UCSF Chimera—A visualization system for exploratory research and analysis
Pettersen, E. F., Goddard, T. D., Huang, C. C., Couch, G. S., Greenblatt, D. M., Meng, E. C., & Ferrin, T. E. (2004) · Journal of Computational Chemistry · doi:10.1002/jcc.20084 · Cited by 45,828 · cited 1×
“We visualize this PDB data into the 3D model by using UCSF Chimera 43 .”
MethodHighly accurate protein structure prediction with AlphaFold
Jumper, J., Evans, R., Pritzel, A., Green, T., Figurnov, M., Ronneberger, O., Tunyasuvunakool, K., Bates, R., Žídek, A., Potapenko, A., et al. (2021) · Nature · doi:10.1038/s41586-021-03819-2 · Cited by 43,764 · cited 1×
“A rti fi cial neural networks (ANNs) have shown their effectiveness in various fi elds such as autonomous cars, fi nding new drugs, designing circuits, or predicting protein structures 1 – 4 .”
BackgroundReservoir computing approaches to recurrent neural network training
Lukoševičius, M., & Jaeger, H. (2009) · Computer Science Review · doi:10.1016/j.cosrev.2009.03.005 · Cited by 2,956 · cited 1×
“Based on the sequential/temporal data processing ability of short-term memristors, a simple neural network, which is called a read-out function, can easily learn the input data with a very small number of parameters compared to conventional long short-term memory (LSTM) model 24 – 26 .”
BackgroundAPD3: the antimicrobial peptide database as a tool for research and education
Wang, G., Li, X., & Wang, Z. (2016) · Nucleic Acids Research · doi:10.1093/nar/gkv1278 · Cited by 2,388 · cited 1×
“The training set consists of 1554 AMP sequences, which are collected from three publicly opened databases: the database of Anuran defense peptides (DADP) 38 , the antimicrobial peptide database (APD) 39 , and the a database for antimicrobial peptides (ADAM) 40 .”
MethodMemristors with diffusive dynamics as synaptic emulators for neuromorphic computing
Wang, Z., Joshi, S., Savel’ev, S. E., Jiang, H., Midya, R., Lin, P., Hu, M., Ge, N., Strachan, J. P., Li, Z., et al. (2017) · Nature Materials · doi:10.1038/nmat4756 · Cited by 2,358 · cited 1×
“Among them, diffusive memristors (SiO x N y :Ag) or Mott memristors (NbO 2 or VO 2 ) have been extensively studied as arti fi cial neuron devices 7,14,15 .”
BackgroundLee, M.-J., Lee, C. B., Lee, D., Lee, S. R., Chang, M., Hur, J. H., Kim, Y.-B., Kim, C.-J., Seo, D. H., Seo, S., et al. (2011) · Nature Materials · doi:10.1038/nmat3070 · Cited by 2,228
Neuromorphic Silicon Neuron Circuits
Indiveri, G., Linares-Barranco, B., Hamilton, T. J., Schaik, A. V., Etienne-Cummings, R., Delbruck, T., Liu, S.-C., Dudek, P., Häfliger, P., Renaud, S., et al. (2011) · Frontiers in Neuroscience · doi:10.3389/fnins.2011.00073 · Cited by 1,821 · cited 1×
“The conventional CMOS-based neuro-morphic hardware uses complex CMOS circuitry to imitate the neuronal activity 6 .”
MethodFully memristive neural networks for pattern classification with unsupervised learning
Wang, Z., Joshi, S., Savel’ev, S., Song, W., Midya, R., Li, Y., Rao, M., Yan, P., Asapu, S., Zhuo, Y., et al. (2018) · Nature Electronics · doi:10.1038/s41928-018-0023-2 · Cited by 1,106 · cited 3×
“Previous studies have developed several memristive arti fi cial neurons and proved that memristors are proper devices for ef fi cient arti fi cial neurons 5,7,10,11 .”
BackgroundMethodDeep Reinforcement Learning framework for Autonomous Driving
Sallab, A. E., Abdou, M., Perot, E., & Yogamani, S. (2017) · Electronic Imaging · doi:10.2352/issn.2470-1173.2017.19.avm-023 · Cited by 1,045 · cited 1×
“Moreover, the artificial neuron’s responses with different combinations of the circuit components are measured; (1) different capacitance values and (2) different capacitance and resistance values while keeping the same R × C value (see Supplementary Figs.”
BackgroundReservoir computing using dynamic memristors for temporal information processing
Du, C., Cai, F., Zidan, M. A., Ma, W., Lee, S. H., & Lu, W. D. (2017) · Nature Communications · doi:10.1038/s41467-017-02337-y · Cited by 1,015 · cited 1×
“Similar approaches to process sequential/temporal data with volatile memristors 12,13,27 – 29 have been extensively studied, but there have been several limitations.”
Method
Cited by
Memristor‐Based Neuromorphic Chips
Duan, X., Cao, Z., Gao, K., Yan, W., Sun, S., Zhou, G., Wu, Z., Ren, F., & Sun, B. (2024) · Advanced Materials · doi:10.1002/adma.202310704 · Cited by 432
All-ferroelectric implementation of reservoir computing
Chen, Z., Li, W., Fan, Z., Dong, S., Chen, Y., Qin, M., Zeng, M., Lu, X., Zhou, G., Gao, X., et al. (2023) · Nature Communications · doi:10.1038/s41467-023-39371-y · Cited by 183
“Such power consumption is at least 3 times lower than those of the state-ofthe-art filamentary memristors used for RC hardware systems [7-9].”
Essential Characteristics of Memristors for Neuromorphic Computing
Chen, W., Song, L., Wang, S., Zhang, Z., Wang, G., Hu, G., & Gao, S. (2022) · Advanced Electronic Materials · doi:10.1002/aelm.202200833 · Cited by 153
Seok, H., Lee, D., Son, S., Choi, H., Kim, G., & Kim, T. (2024) · Advanced Electronic Materials · doi:10.1002/aelm.202300839 · Cited by 155
Short-term synaptic plasticity in emerging devices for neuromorphic computing
Li, C., Zhang, X., Chen, P., Zhou, K., Yu, J., Wu, G., Xiang, D., Jiang, H., Wang, M., & Liu, Q. (2023) · iScience · doi:10.1016/j.isci.2023.106315 · Cited by 146
“The time constant of the interface-based devices generally ranges from 1ms to 100 ms.(55,56)”
Background
Metadata
- Authors
- Park, S.-O., Jeong, H., Park, J., Bae, J., & Choi, S.
- Affiliations
- Korea Advanced Institute of Science and Technology (KR)
- Corresponding author
- Choi, Shinhyun
- Venue
- Nature Communications
- Published
- 2022-06-03
- Volume
- 13
- Issue
- 1
- Pages
- 2888-2888
- ISSN
- 2041-1723 (ISSN-L)
- PMC
- PMC9166790
- OpenAlex
- W4281628293
- Cited by
- 331
- References
- 47
- Citations by year
- 2026: 61 · 2025: 105 · 2024: 100 · 2023: 55 · 2022: 10
- FWCI
- 18.48
- Type
- journal article
- Language
- en
- Open access
- gold
- doi.org
- Licence
- cc-by
- creativecommons.org/licenses/by/4.0 (tdm)
- creativecommons.org/licenses/by/4.0 (vor)
- Locations
- Nature Communications (publishedVersion, cc-by)
- PubMed Central (publishedVersion, cc-by)
- PubMed (publishedVersion)
- RePEc: Research Papers in Economics (submittedVersion)
- DOAJ (DOAJ: Directory of Open Access Journals) (submittedVersion)
- Funders
- Samsung (SRFC-IT2101-04)
- Samsung Research Funding & Incubation Center of Samsung Electronics
- Topics
- Physical Sciences › Engineering › Electrical and Electronic Engineering › Advanced Memory and Neural Computing
- Physical Sciences › Computer Science › Artificial Intelligence › Neural Networks and Applications
- Physical Sciences › Computer Science › Artificial Intelligence › Neural Networks and Reservoir Computing
- Keywords
- memristors, artificial neurons, neuromorphic computing, device reliability, leaky integrate-and-fire neurons, sequential data processing, antimicrobial peptides, edge computing, energy-efficient computing
- MeSH terms
- Brain, Humans, Neurons, Reproducibility of Results, Neural Networks, Computer
About these checks
Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.
The claim verdicts this report counts are experimental: their accuracy is not yet measured.
A language model drew the claims from the paper on 2026-09-30, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).
A second model labelled each claim on 2026-09-30 (model jev-1.13-20260917, definition revision gold-guideline-v0.2).
Cited sources are checked with a language model’s help (gpt-6-sol).
Statistics and arithmetic are recomputed, not judged.
Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.
What counts as a claim ›How we measure confidence ›How we check this ›
Are you an author, or do you see a problem with this page? Request a review
How to cite
Cite CoChat for the checks on this page, and the paper for its own findings.
- This page
- CoChat. (2026). Citation check: Experimental demonstration of highly reliable dynamic memristor for artificial neuron and neuromorphic computing. https://app.cochat.ai/papers/10.1038/s41467-022-30539-6
- The paper
- Park, S.-O., Jeong, H., Park, J., Bae, J., & Choi, S. (2022). Experimental demonstration of highly reliable dynamic memristor for artificial neuron and neuromorphic computing. Nature Communications. https://doi.org/10.1038/s41467-022-30539-6
Data sources
Bibliographic data: Semantic Scholar, OpenAlex, Crossref.
Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).
Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.