Simulations of an analog ReRAM cell accurately reproduced the measured nonlinear current-voltage behaviour seen during a slow voltage sweep, while a separate analysis found corresponding equilibrium states under pulsed and quasi-static inputs. The cross-domain match was limited to the same memory-state window, and the equilibrium finding came from the model.
The study evaluated whether equilibrium stabilization around the device’s symmetry point was consistent under pulsed and quasi-static responses. The symmetry point is the point around which the simulated conductance response oscillated. The researchers used electrical-characterization data from a 200 nm × 200 nm TiN/TaOx/HfOx/TiN ReRAM cell.
A model tested against electrical measurements
The advanced model included external parasitic resistances, corrected ion-migration dynamics and changing electrical-conduction parameters linked to defect density.
For the quasi-static tests, the simulated voltage rose in a staircase, with 10 millivolt steps and 100 milliseconds at each step, giving a global sweep rate of 0.1 volts per second. Under that protocol, the model accurately reproduced the nonlinear experimental current-voltage characteristic using hopping transport.
A separate single-pulse SET test examined a change from an initial read resistance of 8 kilohms to a final resistance of 2 kilohms, allowing a tolerance of plus or minus 300 ohms. Across the voltage and time conditions tested, the simulated and experimental switching kinetics agreed. The reported switching-kinetics nonlinearity was about 175 millivolts per decade.
Projection of the simulated sweep estimated a migration onset speed of about 1.1 nanometres per second and an average displacement of about 6 nanometres for both SET and RESET transitions. These are model-derived estimates; the supplied analysis does not report microscopic experimental confirmation.
The same framework reproduced analog pulsing
The researchers tested open-loop analog programming in simulation with 200 positive pulses and 200 negative pulses, each lasting 300 nanoseconds, followed by 300 alternating pulses around the symmetry point. The model closely reproduced the resulting bidirectional conductance modulation and the oscillation around that point.
A stochastic version of the model was used to represent resistance fluctuations. It matched the measured fluctuation trend with a root mean squared error of 2.16 percent. Each plotted batch contained 50 consecutive read-resistance points drawn from a sequence of 700 pulses. That error describes the model’s agreement with the measured trend; it is not a formal uncertainty range across devices.
What the simulated equilibrium represents
In the quasi-static simulation, the SET and RESET state routes intersected at one equilibrium point where the state-variable fluxes were equal. In pulsed operation, the simulated symmetry point was reached independently of the starting condition for a given input, a behaviour described as a fading-memory regime. For the same memory-state window, the equilibrium defect state on the quasi-static routes matched the defect state at the pulsed symmetry point.
The modeled results associated larger programming voltages and shorter pulses with lower switching energy and a higher maximum temperature in the switching region. The supplied evidence treats both quantities as simulation outputs rather than measurements in a finished device.
Training results differed between simulated symmetry settings
For a Tiki-Taka training simulation, the researchers max-normalized simulated conductance traces and fitted them to soft-bound device models. They then used those fitted models as presets in a three-layer fully connected network trained on MNIST.
The training simulations showed lower accuracy when the simulated symmetry point was away from the centre of the conductance window. The most extreme shifted-symmetry setting did not converge, while the centred setting had the highest accuracy and fastest convergence. In that centred setting, the positive pulse was 1.47 volts and the negative pulse was minus 1.45 volts.
These results describe differences between simulated switching symmetries in an algorithm experiment. They do not demonstrate Tiki-Taka training on physical ReRAM hardware, and the equilibrium correspondence remains a simulation finding rather than an experimentally established microscopic mechanism.
Evidence remains tied to the tested model and device
The evidence concerns the specified TiN/TaOx/HfOx/TiN architecture and an unspecified number of devices and independent replicates. The supplied analysis reports no formal inferential tests, confidence intervals or replicate-level uncertainty. The variability model also omitted explicit trap-capture and emission time constants, while the fitted training model did not include inter-device variability.
The supplied analysis identifies several next tests: spectroscopy to check the modeled defect redistribution, broader studies across device, material and fabrication variability, and experiments to determine whether pulse-amplitude tuning produces the reported training behaviour in physical hardware. Those questions remain open beyond the simulations reported here.
The supplied document identifies arXiv:2608.25767v1, dated 26 August 2026. The authors declared no competing interests. Funding came through PHASTRAC, SNSF ALMOND and NEUROTEC, with support from European, Swiss and German agencies.
Paper data and sources
Original title: Study of Resistive Switching Dynamics and Memory States Equilibria in Analog Filamentary Conductive-Metal-Oxide/HfOx ReRAM via Compact Modeling
Authors: Matteo Galetta, Donato Francesco Falcone, Victoria Clerico et al.
Journal/Repository: Adv. Electron. Mater. 12, no. 7 (2026): e00373
Status: Peer-reviewed
First online: 2026-08-26
DOI: 10.1002/aelm.202500373
Original paper · Full text