Source-linked AI summary
Variability in Resistive Memories
Juan B. Roldán, Enrique Miranda, David Maldonado, Alexey N. Mikhaylov, Nikolay V. Agudov, Alexander A. Dubkov, Maria N. Koryazhkina, Mireia B. González, Marco A. Villena, Samuel Poblador, Mercedes Saludes-Tapia, Rodrigo Picos, Francisco Jiménez-Molinos, Stavros G. Stavrinides, Emili Salvador, Francisco J. Alonso, Francesca Campabadal, Bernardo Spagnolo, Mario Lanza, Leon O. Chua
TL;DR
Cycle-to-cycle variability is an inherent challenge in resistive memories and must be represented accurately for circuit-oriented use. This review surveys experimental characterization and stochastic, compact, behavioral, quantum, and mesoscopic modeling approaches. It concludes that variability appears under ramped and pulsed operation, with model parameters and formulations capturing device-specific behavior, while fabrication and modeling assumptions constrain applicability.
Problem
The review addresses inherent cycle-to-cycle variability in resistive memories and the need to assess it experimentally and in models used for circuit design.
Method
The review compares experimental characterization with physically based, stochastic, compact, quantum, behavioral, and mesoscopic approaches to modeling variability.
Results
The review reports that variability is observed under ramped and pulsed signals and that potentiation is faster than depression in the studied devices, requiring transition-specific model parameters.
Takeaways & Limitations
Accurate variability treatment requires models that reflect the physical origin and operating conditions of resistive switching rather than treating variability as a generic current-noise source.
Takeaways & Limitations
The mesoscopic quantum model is a phenomenological framework rather than a complete simulation tool, and two-dimensional materials remain limited by nonoptimized fabrication processes.
Abstract
from arXiv · showhide
Resistive memories are outstanding electron devices that have displayed a large potential in a plethora of applications such as nonvolatile data storage, neuromorphic computing, hardware cryptography, etc. Their fabrication control and performance have been notably improved in the last few years to cope with the requirements of massive industrial production. However, the most important hurdle to progress in their development is the so-called cycle-to-cycle variability, which is inherently rooted in the resistive switching mechanism behind the operational principle of these devices. In order to achieve the whole picture, variability must be assessed from different viewpoints going from the experimental characterization to the adequation of modeling and simulation techniques. Herein, special emphasis is put on the modeling part because the accurate representation of the phenomenon is critical for circuit designers. In this respect, a number of approaches are used to the date: stochastic, behavioral, mesoscopic..., each of them covering particular aspects of the electron and ion transport mechanisms occurring within the switching material. These subjects are dealt with in this review, with the aim of presenting the most recent advancements in the treatment of variability in resistive memories.
1. Introduction
Redox-based resistive memories offer useful nonvolatile, low-power, CMOS-compatible platforms for storage, security, and neuromorphic computing, but their stochastic switching creates cycle-to-cycle variability. This review addresses that variability experimentally and through physical, stochastic, behavioral, and mesoscopic modeling approaches.
- Redox-based devices combine endurance, retention, low-power consumption, CMOS compatibility, and three-dimensional integration.
- Their stochastic operation supports physical unclonable functions and random-number generators for cryptographic integrated circuits.
- Low power and analog vector-matrix multiplication make these devices relevant to compute-in-memory neuromorphic hardware.
- Redox-based memories comprise electrochemical metallization cells and valence-change memories, with both exhibiting stochastic set and reset variations.
- The review focuses on variability from experimental and modeling perspectives, covering transition-metal oxides, two-dimensional materials, and physically based models.
2.1. Oxide-Based Devices
Oxide-based filamentary VCMs switch through redox-driven formation and partial disruption of conductive filaments, producing useful bipolar resistance states but substantial cycle-to-cycle variability. Measurements show that variability depends on the resistance state and switching parameter, while successive cycles retain significant correlations.
- Filamentary VCM switching arises from redox processes and partial conductive-filament formation or disruption in the oxide.
- HfO2-based devices provide reproducible resistance states, endurance, and a broad application basis including memory, logic, and hardware security.
- Cycle-To-Cycle Variability: 3000 voltage-ramp cycles show greater cycle-to-cycle spread in HRS current than LRS current, with higher variability also for Iset than Ireset and Vset than Vreset.
- Cycle-To-Cycle Variability: Successive-cycle correlations exceed 0.9 for Ion and Ireset, while Ioff remains correlated with ρ ≥ 0.63.
- Cycle-To-Cycle Variability: 105 pulsed cycles preserve correlations of ρ ≥ 0.81 for Ion and ρ ≥ 0.75 for Ioff, while the Ion/Ioff ratio slightly decreases during cycling.
- Variability Sources: Variability reflects random filament path formation, possible nucleation in weak regions, endurance limits, retention concerns, resistance-state instability, and random telegraph noise.
2.2. Devices Based on Two-Dimensional Materials
Two-dimensional materials are explored as a route to control resistive-switching variability, while device architectures, electrode design, and fabrication methods address different variability sources. Their benefits are balanced by defect- and scale-related limitations, with CVD improving homogeneity relative to exfoliation.
- Material strategies: Stack engineering and electrode design aim to control conductive-filament formation, ion migration, interfacial barriers, and cycling uniformity.An Ag–Cu alloy top electrode was used in amorphous-silicon memristors to improve uniformity in the cycling response.
- Material strategies: 2D materials can confine resistive switching through crystallographic features such as grain boundaries, while offering diverse electronic properties.The switching process is suggested to occur preferentially along grain boundaries and other lower-resistance regions.
- Device architectures: Transistor-assisted cells regulate switching and suppress sneak-path effects: 1T1R limits the voltage range, while 1T2R stores information as a resistance ratio.The resistance-ratio approach uses two serially connected memristors and can decrease cycle-to-cycle variability effects.
- Variability sources: Exfoliated 2D materials introduce thickness variations, contaminants, wrinkles, gaps, and boundaries that can affect cycle-to-cycle and device-to-device variability.Mechanical transfer prevents large-area fabrication and accurate thickness control, making homogeneous defect distributions difficult to maintain.
- Variability sources: Defects can either reduce variability by localizing switching in lower-resistance regions or increase it when excessive or uncontrolled.The same defect classes may stabilize switching when they confine the process, but can produce the opposite effect when present excessively.
- Fabrication approaches: Chemical-vapor deposition reduces defects associated with exfoliation, improving cycle-to-cycle stability and increasing device-to-device homogeneity through larger-area growth.The cited comparison links CVD with fewer undesired defects, better cycling stability, and more homogeneous devices.
2.3. Good Practices for the Experimental Data Analysis Focused on Variability
Reliable variability analysis requires statistically grounded treatment of measured I–V curves, switching voltages, and device populations. Representative curves and compact plots are useful, but can mislead without distributional validation and sufficiently broad sampling.
- I–V curve representation: Statistical analysis should replace unsupported claims that a manually selected I–V curve is representative of the device response.A probability-density plot identifies the most frequent electrical response rather than merely the geometric center of a curve cloud.
- I–V curve representation: The probability-density-based curve can reveal a typical memory window that differs substantially from the manually selected or median reference.Using the pink or green reference suggests roughly 7 orders of magnitude, whereas the most probable window is 2 orders smaller.
- I–V curve representation: The mean I–V curve is unreliable when current dispersion is high, particularly in the reset region spanning 9 orders of magnitude between HRS and LRS.Although the mean estimates the set response reasonably in this example, it lies far from relevant reset-response regions.
- Switching-voltage analysis: Set and reset voltage analysis requires systematic extraction methods and distribution parameters to justify whether experimental data follow a Weibull distribution.Consistent extraction avoids bias, while reporting distribution parameters supports statistically defensible variability claims.
- Variability visualization: Histograms, cycle-to-cycle progression, and boxplots offer complementary variability views, but Gaussian fits require statistical justification and boxplots do not support inference.Boxplots compactly display information and are recommended for device-to-device studies involving many devices.
- Variability visualization: Device-to-device variability claims require a statistically significant device population, with boxplots able to show both device-to-device and cycle-to-cycle variability compactly.One cited example studied 48 identical devices.
3. Variability Physical Models
Physical variability models represent cycle-to-cycle stochasticity in resistive memories for circuit-oriented analysis, spanning compact, mesoscopic, and quantum-transport descriptions. These approaches reproduce several experimental distributions and trends, while retaining important scope and fitting limitations.
- Stanford-based models: Stanford-model simulations include a cycle-to-cycle variability implementation for resistance, switching-voltage, and current-related parameters.The model is intended to support variability-aware electronic-design-automation tools for integrated-circuit design.
- Stanford-based models: The Stanford model reproduces experimental I–V variation and parameter spreads, but cannot fully model set- and reset-voltage variations.The simulated variation is comparable to experiment in most aspects, while some voltage variability remains unrepresented.
- Stanford-based models: Increasing δ0g raises current variation, whereas increasing Tcrit lowers current variation by changing gap deviation and the threshold-temperature condition.These parameters connect variability trends to the simulated filament gap and temperature-dependent switching behavior.
- Stanford-based models: Series resistance is required in one VCM simulation to reproduce the experimental set and reset I–V curve shapes, including a 22.3 Ω resistance case.Changing other model parameters could not produce the required curve-shape changes.
- Quantum and mesoscopic approaches: The mesoscopic framework treats LRS conductance as generally less variable than HRS because HRS transport depends sensitively on a small atomic constriction and its barrier.The approach uses quantum-point-contact concepts but is presented as a phenomenological framework rather than a complete simulation tool.
- Quantum and mesoscopic approaches: In the quantum approach, barrier-height variability affects HRS strongly but is largely irrelevant to LRS, while normally distributed barriers produce lognormally distributed current amplitudes.Departures from normal barrier-variation distributions may indicate jumps or strong correlations in the stochastic process.
4. Variability Stochastic Models
The review organizes stochastic modeling of resistive switching around dynamical, microstructural, thermodynamical, and stochastic approaches, emphasizing physically consistent noise and compact simulation models. These models represent variability through fluctuating state variables, noise sources, and conductive-filament dynamics, while balancing physical accuracy against computational complexity.
- Modeling approaches: Four modeling approaches—dynamical, microstructural, thermodynamical, and stochastic—are distinguished for representing memristive-system variability.Dynamical models omit explicit stochasticity; microstructural models increase physical detail but require more parameters and computation.
- Thermal noise: Thermodynamical and stochastic models incorporate fluctuations through noise sources whose intensity can depend on temperature and satisfy fluctuation–dissipation constraints.Thermal noise drives random transitions between potential wells, while colored or non-Gaussian noise requires modified Langevin equations.
- Stochastic formulations: Stochastic models combine electrical relations with Langevin dynamics to represent random forces and fluctuations in resistive switching.The Langevin formulation is related to the Fokker–Planck description and can model diffusion of defects or ions.
- Model outcomes: Distributed and lumped stochastic models can describe bipolar resistive switching and predict stochastic resonance and transient bimodality induced by external noise.Under certain assumptions, the models reduce to Brownian diffusion equations that require limited computational resources and may admit analytical solutions.
- Thermodynamic consistency: The stochastic formulations recover equilibrium Gibbs or Boltzmann distributions, supporting the thermodynamic consistency of the modeled dynamics.The steady-state solutions of the relevant Fokker–Planck and Langevin equations are identified with these equilibrium distributions.
- Noise implementation: Memristor models use multiplicative noise when resistance depends on the state variable, whereas constant resistance reduces the noise to an additive Nyquist source.For resistive memories, the stochastic variables describe diffusing ions or defects whose concentration or conductive-filament length changes the memristance.
- Compact modeling: Compact stochastic models connect physical variability with circuit simulation, but simplified versions rely on assumptions such as ideal periodic microstructure and low thermal-noise intensity.The compact model is linked to the Stanford model while using a noise formulation consistent with fluctuation–dissipation relations.
5. Variability Behavioral Models
Behavioral models represent cycle-to-cycle variability by varying model parameters, accounting for correlations, hysteresis-driven propagation, voltage-sweep effects, and temporal memory in switching cycles.
- General memristor model formalism: Monte Carlo simulations vary memristor-model parameters to generate cycle-to-cycle variability in HfO2-based RRAM devices across repeated resistance-state transitions.The framework uses a charge- and flux-domain representation and evaluates devices over several tens of cycles.
- General memristor model formalism: The studied devices switch from HRS to LRS faster than from LRS to HRS, so model parameters capture asymmetric conductive-filament creation and dissolution dynamics.The equations still reproduce measurements satisfactorily despite the different transition speeds.
- Obtained Results and Discussion: Parameter n shows transition-dependent correlations, requiring separate models for potentiation-to-depression and depression-to-potentiation transitions.A linear relation is reported for one transition direction, while the opposite transition is modeled differently.
- Dynamic memdiode model: Hysteresis propagates variability across curve regions, while assigning variability to too many parameters can produce over-random simulated I–V characteristics.For example, Ioff affects both HRS current and set transition, whereas Ion affects both LRS current and reset transition.
- Dynamic memdiode model: Incomplete reset and insufficient voltage span can produce different initial and final HRS-current distributions, with I–V loops stabilizing only after additional cycles.At 0.9 V, incomplete reset determines the final HRS current in the reported example.
- Time-Series Modeling of Cycle-To-Cycle Variability: Time-series analysis models cycle-to-cycle memory using ARIMA(0,1,1) for Vset and ARIMA(0,1,2) for Vreset, enabling prediction of parameter evolution.The approach uses prior cycle values and residual information after transforming nonstationary series into stationary differential series.
6. Discussion
The discussion emphasizes that cycle-to-cycle variability must be incorporated into resistive-memory models and EDA tools, while no standard model yet covers all technologies and behaviors.
- Predictable behavior across large groups of memory cells remains difficult, making variability incorporation into EDA simulation tools important for device design.
- Behavioral models simplify equations and accelerate simulations, but they obscure the physical meaning of model parameters.
- Voltage-dependent randomness, cycle-to-cycle correlations, multivariate parameter dispersion, and complex noise processes remain open modeling issues.
- No standard model is yet available across the reviewed platforms and modeling approaches.The review identifies a substantial remaining need for modeling development.
7. Conclusions
The review examines resistive-memory variability experimentally and through modeling, including devices based on conventional transition-metal oxides and two-dimensional materials.
- Resistive memories combine CMOS compatibility and broad application potential with inherent cycle-to-cycle variability.
- Variability appears under both ramped and pulsed voltage characterization, with its operational influence depending on fabrication materials and device physics.
- Current variations were analyzed over long resistive-switching series in HfO2 devices and h-BN-based devices.
- The review develops modeling approaches for resistive-switching variability and discusses representative methods for analyzing process variations.
Keywords
The paper concerns resistive-memory variability, memristors, modeling, resistive switching, and two-dimensional materials.
- The keywords identify two-dimensional materials, experimental characterization, memristors, modeling, resistive memory, resistive switching, and variability.