CoChat

Retention Secured Nonlinear and Self-Rectifying Analog Charge Trap Memristor for Energy-Efficient Neuromorphic Hardware

Open access

The DOI matches its Crossref record; 44 of the 45 references it cites were found in a registry or an index; 3 of the 34 sources its claims cite were checked for support.

Abstract

A memristive crossbar array (MCA) is an ideal platform for emerging memory and neuromorphic hardware due to its high bitwise density capability. A charge trap memristor (CTM) is an attractive candidate for the memristor cell of the MCA, because the embodied rectifying characteristic frees it from the sneak current issue. Although the potential of the CTM devices has been suggested, their practical viability needs to be further proved. Here, a Pt/Ta 2 O 5 /Nb 2 O 5‐ x /Al 2 O 3‐ y /Ti CTM stack exhibiting high retention and array‐level uniformity is proposed, allowing a highly reliable selector‐less MCA. It shows high self‐rectifying and nonlinear current‐voltage characteristics below 1 µA of programming current with a continuous analog switching behavior. Also, its retention is longer than 10 5 s at 150 °C, suggesting the device is highly stable for non‐volatile analog applications. A plausible band diagram model is proposed based on the electronic spectroscopy results and conduction mechanism analysis. The self‐rectifying and nonlinear characteristics allow reducing the on‐chip training energy consumption by 71% for the MNIST dataset training task with an optimized programming scheme.

Overview

CoChat Summary

## TL;DR

The paper proposes a Pt/Ta2O5/Nb2O5-x/Al2O3-y/Ti charge trap memristor that combines self-rectification, nonlinear analog switching, and high retention for selector-less crossbar arrays. In a 32×32 array and MNIST simulation, the device enabled an estimated 71% energy reduction under an optimized programming scheme and about 91% recognition accuracy.

## Key findings

  • The PTNAT device shows strong rectifying and nonlinear behavior, with a maximum on/off ratio of ≈8 × 10^3 at 4.5 V and a maximum rectifying current ratio of ≈5.7 × 10^4 at 7.5 V.
  • The device operates with analog switching below 1 µA programming current, is electroforming-free, and shows endurance up to 10^5 cycles.
  • Retention is reported as longer than 10^5 s at 150 °C in the abstract, and the conclusion states >2 × 10^5 s at 150 °C; the body also reports stable LRS and HRS at 150 °C (read at 2 V).
  • In a 32 × 32 crossbar array, 80 randomly selected cells showed high array-level uniformity, and the proposed biasing scheme reduced on-chip training energy from 0.66 J to 0.19 J, a 71% reduction; compared with the conventional half-voltage/serial scheme, energy fell from 0.51 J to 0.31 J (39% reduction) with parallel programming.
  • MNIST training on a fully connected network (784-256-10), trained for 10 epochs on 60,000 samples, achieved an estimated recognition accuracy of ≈91%.

## Methods

  • Device design and fabrication: a Pt/Ta2O5/Nb2O5-x/Al2O3-y/Ti stack was fabricated on SiO2/Si; the device area was 25 µm² and the crossbar line width was 5 µm.
  • Characterization: single-device and array electrical measurements were performed with a probe station/Keithley 4200A-SCS and an ArC ONE platform; spectroscopy used XPS, TEM, REELS, and UPS to analyze composition and band structure.
  • Analysis and simulation: conduction was fit with temperature-dependent and state-dependent Schottky emission analysis; neuromorphic energy and MNIST accuracy were evaluated using a software-based emulator of a fully connected DLP network with bi-directional weight updates.

## Limitations

  • The paper notes the device is “not ideal yet”; the estimated ≈91% MNIST accuracy is lower than the ideal ≈97% possible in the same network, due to non-ideal linearity and conductance variation.
  • The array testing platform’s current sensing limit was about ≈10^-8 A, so lower current levels were not detectable.
  • Evident from the text: some retention results are reported at different temperatures and times (125 °C for 1000 s in one measurement, 150 °C for >10^5 s or >2 × 10^5 s elsewhere), so the exact headline retention figure is not fully consistent across sections.

Model-written summary. Check it against the paper before relying on it.

Integrity signals

Registry

Crossref · 10.1002/advs.202205654

Cites

44 of 44 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Figures

Figure 1 Pt/Ta2O5/Nb2O5‐ x /Al2O3‐ y /Ti (PTNAT) cell characteristics. a) The cross‐section transmission electron microscopy (TEM) image of the device and fast‐Fourier transform (FFT) image. b) The resistance switching I–V curves of the device with 1 µA compliance current (I CC) measured at different positive voltage sweeps (6–10 V) and −10 V fixed voltage sweeps. c) I–V curves with continuously increasing voltage sweeps (6–10 V) without reset. d) The on/off ratio and the rectifying ratio as a function of the applied voltages. e) The cumulative probabilities of the current levels of LRS in the positive voltage region, which were obtained from the first voltage sweeps in 20 cells. f) The endurance properties up to 105 cycles read at 4 V at room temperature. g) The average and standard deviation of both LTP and LTD for 70 pulses during ten cycles. h) 10 cycles of LTP and LTD operation of the device. i) The retention characteristics of 3‐bit intermediate states at 125 °C at 2 V.
Figure 1 Pt/Ta2O5/Nb2O5‐ x /Al2O3‐ y /Ti (PTNAT) cell characteristics. a) The cross‐section transmission electron microscopy (TEM) image of the device and fast‐Fourier transform (FFT) image. b) The resistance switching I–V curves of the device with 1 µA compliance current (I CC) measured at different positive voltage sweeps (6–10 V) and −10 V fixed voltage sweeps. c) I–V curves with continuously increasing voltage sweeps (6–10 V) without reset. d) The on/off ratio and the rectifying ratio as a function of the applied voltages. e) The cumulative probabilities of the current levels of LRS in the positive voltage region, which were obtained from the first voltage sweeps in 20 cells. f) The endurance properties up to 105 cycles read at 4 V at room temperature. g) The average and standard deviation of both LTP and LTD for 70 pulses during ten cycles. h) 10 cycles of LTP and LTD operation of the device. i) The retention characteristics of 3‐bit intermediate states at 125 °C at 2 V.CC BY · PubMed Central
Figure 2 Retention tendency comparison with 2‐layer and 3‐layer devices. The resistance switching I–V curves of the a) Pt/Nb2O5‐ x /Al2O3‐ y /Ti (PNAT), b) Pt/Ta2O5/Nb2O5‐ x /Ti (PTNT), and c) Pt/Ta2O5/Nb2O5‐ x /Al2O3‐ y /Ti (PTNAT) device. Each I–V curve was measured with a 10 V positive voltage sweep and a −10 V negative voltage sweep with a 10−6 A compliance current. The retention characteristics of d) PNAT (4 V at room temperature (RT)), e) PTNT (3 V at RT), and f) PTNAT (2 V at 150 °C). Schematic energy band diagram models of g) PNAT, h) PTNT, and i) PTNAT devices.
Figure 2 Retention tendency comparison with 2‐layer and 3‐layer devices. The resistance switching I–V curves of the a) Pt/Nb2O5‐ x /Al2O3‐ y /Ti (PNAT), b) Pt/Ta2O5/Nb2O5‐ x /Ti (PTNT), and c) Pt/Ta2O5/Nb2O5‐ x /Al2O3‐ y /Ti (PTNAT) device. Each I–V curve was measured with a 10 V positive voltage sweep and a −10 V negative voltage sweep with a 10−6 A compliance current. The retention characteristics of d) PNAT (4 V at room temperature (RT)), e) PTNT (3 V at RT), and f) PTNAT (2 V at 150 °C). Schematic energy band diagram models of g) PNAT, h) PTNT, and i) PTNAT devices.CC BY · PubMed Central
Figure 3 Conduction mechanism fitting of PTNAT device. a) I–V curves of the LRS measured at temperatures ranging from 40 to 70 °C. b) Schottky emission form (ln(J/T 2) vs E 1/2) plot for the voltages ranging from 3.5 to 6.5 V in LRS with the calculated Ta2O5 layer partial electric field. c) Analog LRS states Schottky emission fitting in +4.5 V read voltage and room temperature condition after +6 to +9 V different sweep voltages. d) High‐frequency permittivity (ε op) and Schottky barrier height (Φ B) extracted at each temperature condition I–V curves. e) ε op and Φ B values extracted at each analog state. Φ B decreased with a more programmed analog state (higher conductance state).
Figure 3 Conduction mechanism fitting of PTNAT device. a) I–V curves of the LRS measured at temperatures ranging from 40 to 70 °C. b) Schottky emission form (ln(J/T 2) vs E 1/2) plot for the voltages ranging from 3.5 to 6.5 V in LRS with the calculated Ta2O5 layer partial electric field. c) Analog LRS states Schottky emission fitting in +4.5 V read voltage and room temperature condition after +6 to +9 V different sweep voltages. d) High‐frequency permittivity (ε op) and Schottky barrier height (Φ B) extracted at each temperature condition I–V curves. e) ε op and Φ B values extracted at each analog state. Φ B decreased with a more programmed analog state (higher conductance state).CC BY · PubMed Central
Figure 4 PTNAT XPS depth profile results and resistance switching mechanism of the device. Depth profiling results of a) Ta 4f, b) Nb 3d, and c) Al 2p core levels of the Ta2O5(10 nm)/Nb2O5‐ x (28 nm)/Al2O3‐ y (8 nm)/Ti sample. The left panels show the raw XPS depth profile data. The right panels enlarge the square region. d) The schematic energy band diagram of the device for zero bias condition. Shallow and deep trap levels exist in the non‐stoichiometric Nb2O5‐ x layer. e) The illustration of the resistance switching process with charge trapping and de‐trapping. In the HRS, deep traps are empty, and conduction is not fluent with the high band offset at the Ta2O5 interface (I). When a positive set bias is applied on the Pt electrode, the electrons can be trapped in the Nb2O5‐ x deep trap sites (II). With the induced negative space charge, the partial electric field across the Ta2O5 layer increased (III), and LRS is achieved by effective Schottky barrier height lowering (IV). When a negative bias is applied (V, VI), the trapped electrons can be released to the Ti electrode, and the HRS is obtained.
Figure 4 PTNAT XPS depth profile results and resistance switching mechanism of the device. Depth profiling results of a) Ta 4f, b) Nb 3d, and c) Al 2p core levels of the Ta2O5(10 nm)/Nb2O5‐ x (28 nm)/Al2O3‐ y (8 nm)/Ti sample. The left panels show the raw XPS depth profile data. The right panels enlarge the square region. d) The schematic energy band diagram of the device for zero bias condition. Shallow and deep trap levels exist in the non‐stoichiometric Nb2O5‐ x layer. e) The illustration of the resistance switching process with charge trapping and de‐trapping. In the HRS, deep traps are empty, and conduction is not fluent with the high band offset at the Ta2O5 interface (I). When a positive set bias is applied on the Pt electrode, the electrons can be trapped in the Nb2O5‐ x deep trap sites (II). With the induced negative space charge, the partial electric field across the Ta2O5 layer increased (III), and LRS is achieved by effective Schottky barrier height lowering (IV). When a negative bias is applied (V, VI), the trapped electrons can be released to the Ti electrode, and the HRS is obtained.CC BY · PubMed Central
Figure 5 32 × 32 MCA integration and energy‐efficient neuromorphic computing demonstration. a) A top‐view SEM image of the 32 × 32 MCA. b) I–V curves of randomly chosen 80 cells with the half voltage scheme. The inset shows the half voltage array measurement scheme. c) The cumulative probabilities of the current levels at the LRS (read voltages of 1.3, 4, and 6 V), and the HRS (read voltages of 4 and 6 V). d) The serial (left) and parallel (right) programming schemes for weight update. e) Suggested potentiation and depression scheme for parallel programming. f) The double‐layer perceptron neural network used for the MNIST dataset training simulation. It consists of 784 input neurons, 256 hidden neurons, and 10 output neurons. g) The MNIST data recognition rate by training epoch considering the experimental synaptic characteristics of the CTM device. The final accuracy was about 91%. h) Estimated energy required for training by serial and parallel programming in half voltage scheme and those in the suggested scheme.
Figure 5 32 × 32 MCA integration and energy‐efficient neuromorphic computing demonstration. a) A top‐view SEM image of the 32 × 32 MCA. b) I–V curves of randomly chosen 80 cells with the half voltage scheme. The inset shows the half voltage array measurement scheme. c) The cumulative probabilities of the current levels at the LRS (read voltages of 1.3, 4, and 6 V), and the HRS (read voltages of 4 and 6 V). d) The serial (left) and parallel (right) programming schemes for weight update. e) Suggested potentiation and depression scheme for parallel programming. f) The double‐layer perceptron neural network used for the MNIST dataset training simulation. It consists of 784 input neurons, 256 hidden neurons, and 10 output neurons. g) The MNIST data recognition rate by training epoch considering the experimental synaptic characteristics of the CTM device. The final accuracy was about 91%. h) Estimated energy required for training by serial and parallel programming in half voltage scheme and those in the suggested scheme.CC BY · PubMed Central

Verification

Claims the paper makes

2 Valid · 1 Not valid · 18 Not checked · 90 Own results and interpretations

61 more claims

Create a free account to see all 111 claims and every source verdict.

See all claims — free account
  • It may state two separate claims as one. 40
  • The quote only partly states this claim. 2
  • Probably not a substantial claim. 7

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?

1 more claim about this paper

Create a free account to see every claim other papers make about this one.

See all claims — free account

Citations

Cites

Show all 44 references

Research works printed but not linked: 1

  1. Modelli A., 2004, 2004 International Conference on Integrated Circuit and Design Technology, 211

Cited by

Metadata

Authors
Kim, G., Son, S., Song, H., Jeon, J. B., Lee, J., Cheong, W. H., Choi, S., & Kim, K. M.
Affiliations
  • Korea Advanced Institute of Science and Technology (KR)
  • Samsung (South Korea) (KR)
Corresponding author
Kim, Kyung Min
Venue
Advanced Science
Publisher
Wiley
Published
2022-11-27
Volume
10
Issue
3
Pages
e2205654-e2205654
ISSN
2198-3844 (ISSN-L)
OpenAlex
W4310277646
Cited by
72
References
45
Citations by year
2026: 13 · 2025: 22 · 2024: 24 · 2023: 12 · 2022: 1
FWCI
4.04
Type
journal article
Language
en
Open access
Locations
  • Advanced Science (publishedVersion, cc-by)
  • PubMed Central (publishedVersion, cc-by)
  • PubMed (publishedVersion)
  • DOAJ (DOAJ: Directory of Open Access Journals) (submittedVersion, cc-by-sa)
Funders
National Research Foundation of Korea (2020M3F3A2A01081585, 2020M3F3A2A01081916, 2022M3I7A4085484, 2022M3F3A2A01085466)
Topics
  • Physical Sciences › Engineering › Electrical and Electronic Engineering › Advanced Memory and Neural Computing
  • Physical Sciences › Engineering › Electrical and Electronic Engineering › Ferroelectric and Negative Capacitance Devices
  • Physical Sciences › Engineering › Electrical and Electronic Engineering › Semiconductor materials and devices
Keywords
memristor crossbar, nonlinear current-voltage characteristics, sneak current suppression, retention stability, conduction mechanism, neuromorphic hardware, energy-efficient training, MNIST classification

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 2026-10-10, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).

A second model labelled each claim on 2026-10-10 (model jev-1.13-20260917, definition revision gold-guideline-v0.2).

Cited sources are checked with a language model’s help (qwen3.8-omni-flash and 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

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.