Student or educator? Sign up with your school email and get 6 months of Pro free.

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 access

Abstract

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

Crossref · 10.1038/s41467-022-30539-6

Cites

45 of 45 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Figures

Fig. 1 The structure of gradual TiOx memristor and its electrical characteristics. a Schematic of Pt/gradual TiOx/Ti memristor structure showing gradual oxygen concentration in the TiOx layer. b Consecutive 125 I–V sweep curves of the device showing very small temporal variation (1.39%) during the cycles. c Endurance test showing 5 × 106 consecutive set/reset pulse without degradation. d Spatial variation (3.87%) calculated through DC voltage sweep in the 30 stand-alone devices. e Fifty consecutive cycles of the pulsed response of a single stand-alone device. A single cycle is composed of 200 set pulse responses (4.5 V, 10 µs) and following self-decaying during 200 ms with a 40 ms decaying time constant. f A 2D map for the decay time constant of 400 devices in a 20 × 20 cross-bar array. g Analog conductance change in a 20 × 20 cross-bar array with 0, 4, 8, 12, and 16 set pulses (4.5 V, 100 µs). h A 3D representation of the conductance change along with the set pulses.
Fig. 1 The structure of gradual TiOx memristor and its electrical characteristics. a Schematic of Pt/gradual TiOx/Ti memristor structure showing gradual oxygen concentration in the TiOx layer. b Consecutive 125 I–V sweep curves of the device showing very small temporal variation (1.39%) during the cycles. c Endurance test showing 5 × 106 consecutive set/reset pulse without degradation. d Spatial variation (3.87%) calculated through DC voltage sweep in the 30 stand-alone devices. e Fifty consecutive cycles of the pulsed response of a single stand-alone device. A single cycle is composed of 200 set pulse responses (4.5 V, 10 µs) and following self-decaying during 200 ms with a 40 ms decaying time constant. f A 2D map for the decay time constant of 400 devices in a 20 × 20 cross-bar array. g Analog conductance change in a 20 × 20 cross-bar array with 0, 4, 8, 12, and 16 set pulses (4.5 V, 100 µs). h A 3D representation of the conductance change along with the set pulses.CC BY · PubMed Central
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.
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.CC BY · PubMed Central
Fig. 3 Gradual TiOx memristor-based LIF neuron. a An illustration of a biological neuron and the analogous memristor-based artificial LIF neuron. b The LIF neuron operation of the developed memristor-based artificial neuron. Applied spike train condition: 4 V, 200 μs width, and 50 μs interval. c The schematic of the circuitry used for the gradual TiOx memristor-based artificial neuron. d–f The output of the artificial neuron along with different presynaptic spike frequencies (3200 Hz (d) and 1600 Hz (e)), and the illustration of the experimental setup. During the test, the pulse amplitude and width are fixed (4 V, 200 μs). g–i The output of the artificial neuron along with the different synaptic strengths (different resistors). When the synaptic strength is weak (large resistance (g)), the artificial neuron fires within six spikes, while it requires fewer spikes when the synaptic strength is strong (small resistance (h)). j, k The outputs of the 22 different artificial neurons demonstrate low variation of the gradual TiOx memristor-based artificial neuron.
Fig. 3 Gradual TiOx memristor-based LIF neuron. a An illustration of a biological neuron and the analogous memristor-based artificial LIF neuron. b The LIF neuron operation of the developed memristor-based artificial neuron. Applied spike train condition: 4 V, 200 μs width, and 50 μs interval. c The schematic of the circuitry used for the gradual TiOx memristor-based artificial neuron. d–f The output of the artificial neuron along with different presynaptic spike frequencies (3200 Hz (d) and 1600 Hz (e)), and the illustration of the experimental setup. During the test, the pulse amplitude and width are fixed (4 V, 200 μs). g–i The output of the artificial neuron along with the different synaptic strengths (different resistors). When the synaptic strength is weak (large resistance (g)), the artificial neuron fires within six spikes, while it requires fewer spikes when the synaptic strength is strong (small resistance (h)). j, k The outputs of the 22 different artificial neurons demonstrate low variation of the gradual TiOx memristor-based artificial neuron.CC BY · PubMed Central
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.
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.CC BY · PubMed Central
Fig. 5 Schematic of the antimicrobial operation mechanism of AMP and the evaluation of generated AMP sequences. a AMPs bursting the bacteria epidermis by pore formation. b An illustration of the transformation of an input sequence into a 20-dimensional vector (or 20 channels) with 24-time steps, where 20 is the number of the unique amino-acids in the input sequence with the added end cursor “@”, and 24 is the length of the sequence after adding the end cursor. c The percentage of AMP sequences in the generation set from short-term (low W/I ratio), long-term (high W/I ratio), and mixed case (various W/I ratios), respectively. d Average global charge from each sequence set. The mixed case shows a similar global charge value to the training set, while the others do not. Inset: the total global charge value of the entire sequences in each set. e Average molecular weight from each sequence set. The mixed case generates sequences that have similar weights to the training set. Inset: the total molecular weights of all of the sequences. f Amino-acid distribution of each sequence set. g The 3D image of an AMP (protein data bank ID: 2mag44) in the training set. h A predicted 3D image of the generated AMP from the mixed case neuro-memristive computing system.
Fig. 5 Schematic of the antimicrobial operation mechanism of AMP and the evaluation of generated AMP sequences. a AMPs bursting the bacteria epidermis by pore formation. b An illustration of the transformation of an input sequence into a 20-dimensional vector (or 20 channels) with 24-time steps, where 20 is the number of the unique amino-acids in the input sequence with the added end cursor “@”, and 24 is the length of the sequence after adding the end cursor. c The percentage of AMP sequences in the generation set from short-term (low W/I ratio), long-term (high W/I ratio), and mixed case (various W/I ratios), respectively. d Average global charge from each sequence set. The mixed case shows a similar global charge value to the training set, while the others do not. Inset: the total global charge value of the entire sequences in each set. e Average molecular weight from each sequence set. The mixed case generates sequences that have similar weights to the training set. Inset: the total molecular weights of all of the sequences. f Amino-acid distribution of each sequence set. g The 3D image of an AMP (protein data bank ID: 2mag44) in the training set. h A predicted 3D image of the generated AMP from the mixed case neuro-memristive computing system.CC BY · PubMed Central

Verification

Claims the paper makes

4 Valid · 10 Not valid · 5 Not checked · 94 Own results and interpretations

67 more claims · 6 with their cited sources checked

Create a free account to see all 113 claims and every source verdict.

See all claims — free account
  • It may state two separate claims as one. 31
  • The quote only partly states this claim. 2
  • Probably not a substantial claim. 10

Create a free account to see these claims.

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?

Citations

Cites

Show all 47 references

Cited by

See all 331

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)
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
Licence
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.