{"id":"f3934698-5c92-4dd1-b66f-13c47ad9a7db","arxiv_id":"2412.05307","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A metal-oxide memristor driven by a sub-threshold sine wave plus Gaussian noise shows maximal signal-to-noise ratio and stabilized resistive switching at an optimal noise intensity, a signature of stochastic resonance.","lead":"Adding noise to a weak periodic signal makes a metal-oxide memristor switch more regularly, and the signal-to-noise ratio of its response peaks at an optimal noise level. This is experimental evidence that stochastic resonance, the constructive role of noise, can appear in these memory devices.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The SNR curve in Fig. 12 is the sole quantitative evidence for the central SR claim, yet the paper never specifies how Rm(t) is extracted or how SNR is computed, and the curve has no error bars; without this, the peak at θζ≈10–12 mV²s could be a measurement artifact.","rationale":"The reader's verdict is already CONDITIONAL, and my concern does not move it; it sharpens it. I identify the experimental SNR extraction in Section 4.3 as the single most load-bearing concern rather than the model equivalence highlighted in the reader's weakest_assumption. The central claim is an observed SNR peak, so if that peak is not robustly measurable, the claim fails regardless of the model. The reader did mention the SNR curve's missing error bars and definition in the rationale, but did not place it in the weakest_assumption slot; hence partial agreement. A concrete protocol test as described would settle whether the peak is real or an artifact. If the peak survives the test, the SR observation stands, and the model equivalence becomes a secondary mechanistic question; if it does not, the paper's headline conclusion would need to be rejected. The reader's identified concern about the effective-temperature mapping (Eq. 3.5) is also legitimate, but it is less decisive for the existence of SR and more about the interpretation.","tokens_in":16098,"tokens_out":9948,"duration_ms":95874,"concrete_test":"Obtain the raw logged V(t) and I(t) series for each noise level and implement a fixed protocol: Rm(t) = (V_applied(t) − I(t)·100Ω)/I(t) with explicit zero-crossing handling; compute PSD via Welch's method with a Hann window over at least 200 periods at 10 Hz; define SNR as the ratio of the 10 Hz peak power to the median power in the 8–12 Hz band excluding ±0.5 Hz around the peak, after subtracting the measurement noise floor recorded with the signal off; repeat on at least three devices and report bootstrap 95% confidence intervals. The SR claim is supported only if the peak at θζ ≈ 10–12 mV²s is reproduced in all devices and the lower confidence bound of the peak exceeds the values at neighboring intensities.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 4.3 reports the central evidence: 'If we consider the SNR measured for the main harmonic at 10 Hz in the spectrum of Rm(t) relative to the level of noise, we obtain the dependence of SNR on external intensity shown in Fig. 12.' No definition of Rm(t) is given, despite the device being measured as current through a 100 Ω load in series. The SNR formula, frequency resolution, windowing, number of averaged cycles, and noise-floor subtraction are unspecified, and Fig. 12 has no error bars (single device, single series). Because the non-monotonic peak is the only quantitative support for the headline SR claim, any of these unspecified choices—e.g., spectral leakage from the periodic switching, division by near-zero current at zero crossings, or noise-floor contamination—could produce or remove the peak. The reader's weakest_assumption focuses on the model's effective-temperature equivalence (Eq. 3.5), which is also a concern, but the SNR extraction is more load-bearing: even a perfectly valid model cannot rescue a claim whose sole quantitative evidence is an under-specified, unreplicated measurement.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper reports an experimental and theoretical study of stochastic resonance (SR) in a ZrO2(Y)/Ta2O5-based metal-oxide memristive device. The authors apply a sub-threshold 10 Hz sinusoidal voltage with superimposed white Gaussian noise and characterize the response through I-V hysteresis loops, time-series statistics of switching parameters, and the spectral signal-to-noise ratio (SNR) of the memristance at the driving frequency. They report that the I-V hysteresis loop area first grows and then shrinks with noise intensity, the ratio of high- to low-resistance-state currents is non-monotonic with a maximum near θζ = 14.43 mV²s, resistive-state variability decreases, and the SNR at 10 Hz has a maximum near θζ ≈ 10–12 mV²s. These observations are interpreted with a coarse-grained stochastic memristor model in which external voltage noise is folded into an effective thermal-noise intensity, and the switching time relative to the driving period is used to explain the non-monotonic hysteresis and SNR behavior.","tokens_in":16383,"tokens_out":4274,"duration_ms":39856,"significance":"If the main claims hold, the paper provides a useful experimental demonstration that external noise can stabilize and regularize resistive switching in a realistic metal-oxide memristor, going beyond earlier theoretical and preliminary experimental studies. The paper's strengths include the use of multiple complementary observables (I-V loop area, resistance ratio, state variability, SNR, autocorrelation functions) and the connection to a stochastic model developed and calibrated in previous work. However, the quantitative evidence for the headline SR claim rests on a single unreplicated SNR curve whose extraction procedure is not documented, and the model-based interpretation relies on an effective-temperature equivalence that is asserted rather than validated for the specific device and frequency range. These gaps currently prevent the paper from fully establishing the SR mechanism.","major_comments":[{"comment":"The SNR curve in Fig. 12 is the central quantitative evidence for the SR claim, but the manuscript does not specify how the memristance Rm(t) is extracted from the measured current through the 100 Ω series resistor, nor does it give the SNR formula, the FFT windowing or frequency resolution, the number of averaged cycles, or the noise-floor subtraction procedure. The phrase \"relative to the level of noise\" is too vague to reproduce the measurement. In addition, Fig. 12 has no error bars and appears to be from a single device and a single measurement series. Without this information, the non-monotonic peak near θζ ≈ 10–12 mV²s could be produced or destroyed by spectral leakage from the switching transients, by division by near-zero spectral components, or by noise-floor contamination. Please provide a complete extraction protocol and include error bars or replicate measurements.","section":"§4.3, Fig. 12"},{"comment":"The model's central assumption is that external white Gaussian voltage noise can be absorbed into the effective thermal noise intensity via Eq. (3.5), so that increasing the external noise is equivalent to raising the device temperature. This equivalence requires that high-frequency voltage fluctuations affect defect hopping in exactly the same way as thermal fluctuations, which is not obvious for a device with Joule-heating inertia and ion-migration delays. Indeed, §4.2 states that for noise intensities below 34.16 mV²s the voltage spikes \"do not alter much the CFs since the thermal inertia and the ion migration delays connected to the RS processes avoid a great modification in the CF nature.\" That statement is in tension with the effective-temperature picture used to explain the observed SR-like behavior. The authors should justify the frequency and amplitude range over which Eq. (3.5) applies to this device, or explicitly delimit the model's interpretative role.","section":"§3, Eqs. (3.3)–(3.5)"},{"comment":"The comparison between measured and simulated I-V characteristics in Fig. 5 is only qualitative. The model parameters are taken from the prior calibration in Ref. [34], and no quantitative metric (e.g., hysteresis-loop area versus θζ, or switching-voltage distributions) is compared between model and experiment. Moreover, the simulated curves are ensemble-averaged while the experimental curves are single-realization measurements, so the two are not directly comparable in a statistical sense. Since the model is used to explain the observed regularities through the switching-time argument in §4.1, the paper should either provide a quantitative model-experiment comparison or clearly present the model as an illustrative rather than validated mechanism.","section":"§4.1, Fig. 5"}],"minor_comments":[{"comment":"The paragraph describing the classical SR experiment and the measurement setup appears twice verbatim in Section 2; one copy should be removed.","section":"§2"},{"comment":"The equations in Section 3 are heavily garbled in the manuscript text (for example, Eq. (3.1) and the definitions of symbols are unreadable), which makes it difficult to verify the derivation of Eqs. (3.3)–(3.5). The final version should contain correctly rendered mathematics.","section":"§3"},{"comment":"The title contains a spacing typo (\"de vice\") and the abstract text says \"Qme series\" instead of \"time series\"; these should be corrected.","section":"Title and Abstract"},{"comment":"The text states that the sampling rate of 6.553 kHz corresponds to a correlation time of 1.526×10⁻⁴ s; this is the sampling interval, not a correlation time, and the wording should be fixed.","section":"§2"}],"recommendation":"major_revision","confidential_remarks":"The duplicated paragraph and garbled equations suggest the manuscript was not carefully proofread. The main technical gap is the under-specified SNR extraction in Fig. 12; this is a fixable issue but it is load-bearing for the central SR claim. The effective-temperature equivalence in Eq. (3.5) also needs a concrete justification or a clear limitation statement."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper reports non-monotonic SNR versus external noise intensity in a ZrO2(Y)/Ta2O5 memristor under sub-threshold sinusoidal drive plus white Gaussian noise. The experimental dataset is new, and the observation is plausible. I like the complementary evidence: the I-V hysteresis area goes up then down, the resistive-state variability drops at intermediate noise, and the ARIMA models show changing correlation structure. The comparison with the stochastic memristor model is a nice addition, even if only qualitative.\n\nThe load-bearing problem is the SNR curve in Fig. 12. The paper never specifies how Rm(t) is reconstructed from the current through the 100 Ω series resistor, nor how SNR is computed (bin width, windowing, averaging, noise-floor subtraction), and the curve has no error bars. Since that curve is the only quantitative evidence for the headline SR claim, the under-specification is not cosmetic. The peak at θζ ≈ 10–12 mV²s could be produced or moved by spectral leakage, by how the noise floor is referenced, or by the way Rm(t) is defined. Without replicates, we have no idea whether the peak is reproducible.\n\nThe model-based interpretation also rests on an assumption: external voltage noise is folded into an effective thermal noise (Eq. 3.5). That is reasonable within the model's coarse-grained picture, but the parameters come from prior calibration, and the I-V comparison is qualitative. Still, the central claim does not depend on the model; it depends on the SNR measurement.\n\nWhat the paper does well: it uses time-series statistics carefully, acknowledges that high noise degrades switching, and positions the result honestly against prior theoretical and experimental SR work. The writing is clear, and the authors mostly avoid overclaiming—except for the SNR curve itself.\n\nFor peer review: yes, this deserves a serious referee. It is a plausible within-subfield result that, if confirmed, would be useful for RRAM variability mitigation. But the authors need to provide a full measurement protocol, error bars from multiple devices or repeated runs, and ideally a model fit to the SNR curve. I would send it out, expecting major revision.\n\nFor a reading group, I'd bring it up as an example of a promising result whose quantitative evidence needs scrutiny. I wouldn't cite it in its current form.","headline":"Plausible but under-specified experimental claim of stochastic resonance in a memristor; the SNR curve needs a clear protocol and error bars before the peak can be trusted.","tokens_in":16981,"tokens_out":2002,"would_cite":false,"duration_ms":20147,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Optimal noise peaks a memristor's response and steadies its switching.","keywords":["memristor","stochastic resonance","resistive switching","yttria-stabilized zirconia","tantalum oxide","signal-to-noise ratio","noise-induced switching","coarse-grained stochastic model"],"falsifier":"Measure the SNR-versus-noise curve with band-limited noise of the same variance but different spectral content, or at several driving frequencies, and compare with the model's effective-temperature prediction. If the peak position moves when only the noise spectrum changes, or if the SNR peak disappears when the drive frequency is changed while the model says it should persist, the equivalence between external noise and thermal noise is falsified; a quantitative fit of the simulated I–V loops to the measured ones would also settle whether the mechanism is the only one at work.","tokens_in":15921,"feed_emoji":"📈","tokens_out":8796,"duration_ms":68722,"temperature":0.7,"pith_summary":"This paper tries to show that random noise, usually a nuisance in a metal-oxide memristor (a two-terminal resistor whose resistance remembers the voltage it has seen), can be put to work. The authors drive a $\\mathrm{ZrO_2(Y)/Ta_2O_5}$ memristor with a 10 Hz sine wave too weak to switch it, add white Gaussian voltage noise of increasing intensity, and find that the device's signal-to-noise ratio (SNR) at the driving frequency first falls, then rises to a clear maximum near $\\theta_\\zeta \\approx 10$–$12$ $\\mathrm{mV}^2\\mathrm{s}$, and then falls again — the fingerprint of stochastic resonance. At that optimal noise level the resistive-switching hysteresis loop is largest, the ratio of high- to low-resistance state currents is maximized, and cycle-to-cycle variability of the resistive states is sharply reduced. If correct, this means noise is not merely a stability problem for memristive devices but a tunable resource: a deliberately chosen noise level can stabilize switching under a sub-threshold drive.","feed_headline":"Optimal noise peaks a memristor's response and steadies switching","feed_subtitle":"At a specific noise level a memristor's switching becomes more stable and its response peaks: stochastic resonance.","key_machinery":"The argument runs on a coarse-grained stochastic memristor model in which oxygen vacancies hop between trapping sites under a periodic potential tilted by the applied voltage. The hopping is described by a Langevin equation for a diffusing particle; the external white Gaussian voltage noise $\\zeta(t)$ is added to the deterministic drive $V_0(t)$, and because both noise sources enter the same linear way, they merge into one effective noise intensity $\\theta_\\nu = (q/\\varepsilon L)^2 \\theta_\\zeta + \\theta_\\xi$. That folding turns the external noise into an apparent temperature increase, which lowers the Kramers switching time $\\tau = \\tau_0 \\exp(E_a/\\theta_\\nu)$ and makes the threshold-like resistance profile switchable. The model's Fokker–Planck equation for the defect concentration, together with a thresholded resistivity function and an effective drift coefficient $\\sinh(BV_0/2\\theta_\\nu)$, produces simulated I–V loops that track the experimental trend of growing-then-shrinking hysteresis.","core_discovery":"The central claim is that a metal-oxide memristive device exhibits classic stochastic resonance when white Gaussian noise is superimposed on a sub-threshold sinusoidal driving voltage. Measured as the signal-to-noise ratio of the memristance at the 10 Hz driving harmonic, the response is non-monotonic in the external noise intensity: the SNR decreases for very weak noise, then grows to a maximum near $\\theta_\\zeta \\approx 10$–$12$ $\\mathrm{mV}^2\\mathrm{s}$, and then decreases again at higher intensities. In the same range, the area of the I–V hysteresis loop first increases and later shrinks, the completeness of SET/RESET switching improves, and the variation of the resistive states after switching drops and saturates. The authors interpret all of this with a stochastic memristor model in which the external voltage noise adds to the thermal noise intensity, so that increasing external noise is equivalent to raising the effective temperature; the observed regularities are presented as evidence that noise plays a constructive role in nonequilibrium memristive systems.","pith_inferences":["If the effective-temperature equivalence holds quantitatively, the optimal noise intensity should shift with the driving frequency and with ambient temperature; measuring that shift would be a direct test of the mechanism.","The same strategy may transfer to other filamentary metal-oxide memristors, but the optimal noise level likely scales with the material's activation energy, so device-specific tuning would be required.","Used deliberately, noise could serve as a free control parameter in neuromorphic circuits, where tunable stochasticity in switching is often desirable; the paper does not explore this application.","Because high-frequency noise spikes are assumed not to modify the filament directly, the mechanism implies a bandwidth limit: noise with spectral content faster than the thermal and ionic response times should stop behaving like an effective temperature."],"forward_implications":["At the optimal noise intensity the device can be switched with a drive amplitude below the usual threshold, so noise can lower the voltage budget for resistive switching.","The signal-to-noise ratio peak gives a measurable operating point: adding noise around $\\theta_\\zeta \\approx 10$–$12$ $\\mathrm{mV}^2\\mathrm{s}$ maximizes the memristance response at the driving frequency.","Cycle-to-cycle variability of the resistive states shrinks with increasing noise up to saturation, implying noise can be used to stabilize device statistics rather than only disturb them.","The non-monotonic hysteresis-loop area provides an experimentally accessible signature of stochastic resonance that does not require spectral analysis.","Time-series models fitted to the reset voltage at different noise intensities allow forecasting of switching-voltage series and quantify how noise changes the memory of past cycles."],"supporting_citations":[{"why":"Supplies the coarse-grained stochastic memristor model (Langevin equation, Fokker–Planck equation, Kramers time) that this paper extends to external noise.","marker":"[34]"},{"why":"First theoretical proposal that noise can enhance the memristive response and predicts the growing amplitude of memristance oscillations with noise.","marker":"[49]"},{"why":"Earlier experiment superimposing white Gaussian noise on a rectangular switching signal and reporting noise-stabilized switching in a memristive device.","marker":"[50]"},{"why":"Prior study of noise-induced resistive switching in the same device stack, giving the basis for the time-series and variability analysis used here.","marker":"[53]"},{"why":"Classical review that defines stochastic resonance and the SNR-versus-noise criterion used to identify the phenomenon.","marker":"[45]"},{"why":"Introduces the time series statistical analysis methodology used to extract switching parameters and correlations.","marker":"[54]"},{"why":"Describes the multilayer ZrO2(Y)/Ta2O5 stack with stabilized resistive switching used to fabricate the devices.","marker":"[27]"}],"fun_headline_variants":["Noise's sweet spot stabilizes memristor switching and boosts response","Stochastic resonance found in metal-oxide memristor: optimal noise helps","Optimal noise sharpens memristor response and stabilizes switching","Memristor stochastic resonance: noise improves switching at right level","Noise tune-up: memristor switching stabilizes at optimal noise"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole mechanistic story rests on treating added external voltage noise as equivalent to raising the device temperature, since the model only sees the combined noise intensity; if the real device reacts to fast noise spikes through Joule-heating inertia or ion-migration delays that the model ignores, the claimed cause of the SNR peak would not follow from the data.","fun_headline_variants_meta":{"raw":{"variants":["Noise's sweet spot stabilizes memristor switching and boosts response","Stochastic resonance found in metal-oxide memristor: optimal noise helps","Optimal noise sharpens memristor response and stabilizes switching","Memristor stochastic resonance: noise improves switching at right level","Noise tune-up: memristor switching stabilizes at optimal noise"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000208,"raw_usage":{"total_tokens":1383,"prompt_tokens":905,"completion_tokens":478,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":521,"completion_tokens_details":{"reasoning_tokens":384}},"tokens_in":521,"tokens_out":478,"duration_ms":4288,"temperature":1.0,"reasoning_tokens":384,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T11:29:03.927585+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Measure the SNR-versus-noise curve with band-limited noise of the same variance but different spectral content, or at several driving frequencies, and compare with the model's effective-temperature prediction. If the peak position moves when only the noise spectrum changes, or if the SNR peak disappears when the drive frequency is changed while the model says it should persist, the equivalence between external noise and thermal noise is falsified; a quantitative fit of the simulated I–V loops to the measured ones would also settle whether the mechanism is the only one at work.","supporting_citations":[{"cited_title":"Nonstationary distri butions and relaxation times in a stochastic model of memristor","cited_arxiv_id":null,"evidence_quote":"Supplies the coarse-grained stochastic memristor model (Langevin equation, Fokker–Planck equation, Kramers time) that this paper extends to external noise."},{"cited_title":"Stochastic memory: Memory enhancement due to noise","cited_arxiv_id":null,"evidence_quote":"First theoretical proposal that noise can enhance the memristive response and predicts the growing amplitude of memristance oscillations with noise."},{"cited_title":"On the benefici al role of noise in resistive switching","cited_arxiv_id":null,"evidence_quote":"Earlier experiment superimposing white Gaussian noise on a rectangular switching signal and reporting noise-stabilized switching in a memristive device."},{"cited_title":"Time series statis- tical analysis: A powerfu l tool to evaluate the variability of resistive swi tching memories","cited_arxiv_id":null,"evidence_quote":"Introduces the time series statistical analysis methodology used to extract switching parameters and correlations."},{"cited_title":"Mul- tilayer Metal Oxide Memr istive Device with Stabilized Resistive Switching","cited_arxiv_id":null,"evidence_quote":"Describes the multilayer ZrO2(Y)/Ta2O5 stack with stabilized resistive switching used to fabricate the devices."}],"review_version":1}