{"id":"e98a11ed-2306-4830-aa76-4c7be8fb0fc1","arxiv_id":"2411.13872","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"Reset voltage variability in h-BN memristors depends on the extraction method, and the reset voltage series is autocorrelated and can be modeled with an ARIMA process.","lead":"This paper compares three ways of measuring the reset voltage of h-BN memristors and fits statistical and circuit models to their cycle-to-cycle variability. It matters because variability is the main obstacle to using these devices in memory and encryption circuits.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"V_reset2's defining condition dQ/dφ=0 is I/V≈0, so it detects the post-rupture current-collapse voltage, not the reset transition; the CV2=28% statistic is therefore not measuring the same quantity as CV1 and CV3.","rationale":"The paper's central contribution is a set of reset-voltage extraction techniques and their variability statistics; among these, the charge-flux method is presented as a new, noise-robust way to obtain V_reset. The derivation in Section IV shows that the defining condition dQ/dφ=0 is equivalent to I=0, which is reached only after the conductive filament has ruptured and the current has collapsed. Consequently, V_reset2 is not an estimator of the reset threshold but of the completion point of the reset transition. Since the reported CV2 is the headline statistic for this method, the comparison with CV1 and CV3 loses its meaning: the three numbers characterize different aspects of the I-V loop. This threatens the abstract's claim that 'different numerical techniques to extract the reset voltage' were compared. It does not necessarily invalidate the autocorrelation finding or the circuit-breaker simulation, which may rely on V_reset1 or other data, but it weakens a main result and requires re-analysis. The proposed test—checking the actual current at V_reset2 and re-extracting with a slope-based criterion—would settle whether the low CV2 is physical or definitional. Other limitations (unreported ARIMA coefficients, no raw data) are real but secondary; the physical definition of the extracted voltage is the more fundamental issue.","tokens_in":10881,"tokens_out":5535,"duration_ms":55771,"concrete_test":"Take the raw I-V traces for the full cycle series; for each cycle compute the ratio r = I(V_reset2)/I_LRS_start, where I_LRS_start is the current at the beginning of the reset sweep. If the median r is <0.05, V_reset2 is a post-rupture point. Then re-extract V_reset2 using the voltage at which |dI/dV| is maximum (or where dQ/dφ is minimum) on the reset branch; recompute CV2 and compare to CV1 and CV3. If CV2 shifts by more than a few percentage points or the ordering changes, the reported comparison is an artifact of the chosen null-current definition.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section IV defines V_reset2 at the first point where dQ/dφ is null in the charge-flux plane (Fig. 2(c)-(e)). From Eqs. (1)-(2), dQ/dφ = (dQ/dt)/(dφ/dt) = i(t)/v(t), so the null condition is i(t)=0 for v(t)≠0. The paper itself states this occurs when 'the device current drops off to negligible values.' In a bipolar reset sweep, the current is initially at LRS and only falls to ~0 after the conductive filament has ruptured; hence V_reset2 is the voltage at which the device has already completed the reset transition, not the voltage at which reset is triggered. The first method (maximum LRS current) and third method (minimum dI/dV) both target the onset or steepest part of the transition. Comparing CV2=28% with CV1=40% and CV3=32% is therefore not a comparison of three estimators of the same physical quantity. Moreover, the null-current point is sensitive to the measurement noise floor and to the voltage-sweep stop value, which can truncate the upper tail of V_reset2 and artificially reduce its dispersion. The central claim that the charge-flux method 'minimizes noise' may be true for locating the end of reset, but it has not been shown to be a valid reset-voltage extraction.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper studies cycle-to-cycle variability in Au/Ti/h-BN/Au memristive devices. The authors extract the reset voltage by three methods: (i) the voltage at maximum LRS current, (ii) the voltage at which the charge derivative with respect to flux is null in the charge-flux (Q–φ) domain, and (iii) the voltage at which the current derivative is minimum. They report coefficients of variation CV1=40%, CV2=28%, CV3=32% and cumulative distribution functions for the extracted parameters. They fit the Q–φ curves with a compact model (Eq. (3)) and analyze the cycle-to-cycle series of V_reset with an ARIMA model, concluding that the reset voltage is autocorrelated while the set voltage is not. Finally, they use a 20×20 circuit-breaker network simulator with thermal switching to reproduce one set/reset I-V cycle and illustrate conductive-filament formation and rupture with network snapshots.","tokens_in":11194,"tokens_out":7934,"duration_ms":75510,"significance":"The paper addresses an underexplored topic: quantitative variability assessment and modeling of h-BN based memristors. Its strengths include the use of raw experimental I-V data for the main variability statistics, the explicit comparison of three extraction criteria, and the integration of charge-flux, time-series, and percolation-network analyses. If the concerns about the reset-voltage definitions and the ARIMA validation are resolved, the work would provide a useful methodology reference for a community that has few such parameter-extraction studies in 2D-material memristors. The reported CV values and the claim that reset voltages carry a serial correlation are the type of falsifiable quantitative statements that are valuable. At present, however, the load-bearing statistics CV2 and the ARIMA claim are not sufficiently supported.","major_comments":[{"comment":"The definition of V_reset2 as the first point where dQ/dφ = 0 is not a reset-onset estimate. Since Q = ∫ i dt and φ = ∫ v dt, dQ/dφ = i(t)/v(t); the null condition is reached when the device current has already fallen to near zero, i.e., after the conductive filament has ruptured. The paper itself states that this point corresponds to 'the device current drops off to negligible values.' Therefore V_reset2 is a reset-completion (current-collapse) voltage, while V_reset1 (maximum LRS current) and V_reset3 (minimum current derivative) target the onset or steepest resistive drop. The comparison CV1=40%, CV2=28%, CV3=32% is thus not a comparison of three estimators of the same physical quantity. The null-current point is also sensitive to the measurement noise floor and to the stop voltage of the sweep, which can truncate the upper tail of V_reset2 and lower its dispersion. Please either provide evidence that the null-current point coincides with the onset of filament rupture, or redefine V_reset2 explicitly as a reset-completion voltage and avoid a direct CV comparison with the other two methods.","section":"Section IV, Fig. 2, Eqs. (1)-(2)"},{"comment":"The ARIMA analysis is not sufficiently validated. The manuscript does not report the estimated coefficients, standard errors, or residual diagnostics (e.g., Ljung-Box tests) for Eq. (4), and Fig. 5 appears to plot the model's fitted values against the same V_reset series used to estimate the model. On this evidence, the claim that the reset voltage series has a predictive 'memory' structure is not established; a fitted ARIMA model can track an estimation sample even if it has no out-of-sample skill. Please add a proper model-identification table, residual diagnostics, and an out-of-sample or cross-validated evaluation. In addition, the statement that V_set has no autocorrelation and therefore 'no time series model can be extracted' is too strong; a white-noise series is still described by a constant-mean model.","section":"Section IV, Eq. (4), Fig. 5"}],"minor_comments":[{"comment":"Several typographical errors should be corrected, including 'derivate' instead of 'derivative'; the numerical differentiation scheme used to compute dI/dV should also be specified (e.g., smoothing window or filter).","section":"Section IV, Fig. 2"},{"comment":"Equation (3) is cited as a compact expression, but its explicit mathematical form does not appear in the manuscript text; please ensure the equation is rendered.","section":"Eq. (3)"},{"comment":"The CDF figures do not state the number of cycles and whether the data come from a single device or multiple devices; this information is necessary to interpret the reported CV values.","section":"Fig. 1(i) and Fig. 3"},{"comment":"The caption of Fig. 5 does not specify which of the three reset-voltage extraction methods is plotted; please state whether the series corresponds to V_reset1, V_reset2, or V_reset3.","section":"Fig. 5"},{"comment":"Only one I-V cycle is compared between the circuit-breaker simulation and experiment; a quantitative error measure (e.g., normalized RMS error) and, ideally, a small ensemble of simulated cycles would make the 'reasonably good' fit claim testable.","section":"Fig. 6-8"},{"comment":"The statement that the charge-flux transformation 'minimizes the effects of electric noise' is plausible but unquantified; a noise-injection or repeatability test would support this claim.","section":"Section IV"}],"recommendation":"major_revision","confidential_remarks":"The stress-test concern about V_reset2 is valid and should be answered before acceptance. The paper has merit and can be revised; I would not reject it outright. The ARIMA validation also needs to be strengthened, but both issues are addressable within the scope of the manuscript."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear colleague,\n\nQuick take: this is a solid, competent characterization study that applies a known toolkit to h-BN memristors and reports genuinely new numbers (CV1=40%, CV2=28%, CV3=32%; ARIMA memory in V_reset but not V_set; a circuit-breaker reproduction of a full set/reset cycle). The work is useful for reliability engineers working with 2D-material devices. The novelty is not in the methods—charge-flux modeling (ref 40), time-series analysis (refs 15, 23), and the CB simulator (ref 31) are all previously published, largely by the same group—but in the application and the specific results.\n\nThe main soft spot is the V_reset2 extraction. The paper defines it as the point where dQ/dphi is null. Since Q and phi are time integrals of current and voltage, dQ/dphi = i(t)/v(t). So the null condition is i(t)≈0, which in a bipolar reset sweep occurs after the conductive filament has already ruptured and the current has collapsed. The paper itself says this is where 'the device current drops off to negligible values.' That makes V_reset2 a post-rupture quantity, not the reset transition onset. Comparing its CV (28%) with CV1 (40%, max-LRS-current) and CV3 (32%, min dI/dV) is not a comparison of three estimators of the same physical parameter. It may still be a useful statistic, but the claim that the charge-flux method gives a lower-variability reset voltage needs re-framing or justification that the null-current point is the physically relevant reset event. Also, that point is sensitive to the noise floor and the sweep stop voltage, which can truncate the distribution and artificially reduce CV.\n\nOther issues are minor. The ARIMA coefficients are not reported, and the model is fit and evaluated on the same series; residual diagnostics would help. The circuit-breaker parameters are tuned to reproduce the same I-V curve shown, so it is a fit, not a prediction. No raw data or code are included. None of this falsifies the central results—the autocorrelation in V_reset is supported by the ACF/PACF plots, and the CV ordering is plausible—but the paper would be stronger with artifact release and more detail.\n\nWho's this for: people working on memristor variability, especially in h-BN or other 2D dielectrics. It deserves a serious referee: the V_reset2 issue is fixable with reframing, and the dataset is valuable. I'd engage with it and ask for revisions rather than reject.","headline":"Useful h-BN variability data, but the charge-flux V_reset2 measures post-rupture current collapse, not reset onset, so the three-way CV comparison is partially apples-to-oranges.","tokens_in":11783,"tokens_out":2229,"would_cite":false,"duration_ms":21204,"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":"The reset voltage of Au/Ti/h-BN/Au memristors is autocorrelated across cycles, so an ARIMA model can predict each cycle's value from the previous one.","keywords":["hexagonal boron nitride","memristive devices","resistive switching","cycle-to-cycle variability","reset voltage extraction","charge-flux domain","time series analysis","circuit breaker simulation"],"falsifier":"Measure, over the same set of switching cycles, the voltage at the onset of the abrupt current decrease during reset and the voltage where $dQ/d\\phi$ first becomes null; if the null point consistently occurs after the current has already dropped by more than half of its low-resistance value, then $V_{reset2}$ is a post-rupture marker and the coefficient-of-variance comparison across the three methods does not describe the same transition.","tokens_in":10659,"feed_emoji":"⚡","tokens_out":8086,"duration_ms":62088,"temperature":0.7,"pith_summary":"This paper tries to establish how much the reset voltage of hexagonal-boron-nitride memristors varies from cycle to cycle and whether that variability is structured or purely random. Working with Au/Ti/h-BN/Au devices, it shows that the measured coefficient of variance depends on the extraction method: 40% when reset is read at the maximum current, 28% when read at the null of the charge-flux derivative, and 32% when read at the minimum of the current derivative. It further shows that the reset-voltage series is autocorrelated while the set-voltage series is not, so a time-series (ARIMA) model can calculate each reset voltage from previous cycles' values. If correct, variability in these devices is not irreducible noise: the reset process carries a memory across cycles that compact models and circuit simulators can exploit, and the set process can be treated as independent from cycle to cycle.","feed_headline":"h-BN reset voltage is autocorrelated across cycles","feed_subtitle":"Three extraction methods report 28–40% cycle-to-cycle variability, and an ARIMA model uses past reset values to predict the next.","key_machinery":"The argument is carried by three extraction methods for $V_{reset}$; by the charge-flux transformation $Q(t)=\\int_0^t i(t')\\,dt'$, $\\phi(t)=\\int_0^t v(t')\\,dt'$, in which the reset point is the first null of $dQ/d\\phi$; by the autoregressive integrated moving-average (ARIMA) time-series model that relates $V_{reset,t}$ to $V_{reset,t-1}$ and the previous residual; and by a circuit-breaker simulator whose breakers switch between $R_{on}=300\\,\\Omega$ and $R_{off}=10^8\\,\\Omega$ with threshold voltages $V_{on}=0.17$ V and $V_{off}=0.182$ V. The charge-flux domain removes measurement noise by integration and makes the $Q$-$\\phi$ curves fit with only three parameters, while the ARIMA model converts the observed autocorrelation into a predictive rule and the circuit-breaker network links the electrical curves to the spatial formation and rupture of conductive nanofilaments.","core_discovery":"The central claim is that, in Au/Ti/h-BN/Au memristive devices, the cycle-to-cycle dispersion of the reset voltage is a real, method-dependent quantity and that the reset process is memory-correlated while the set process is not. Concretely, the coefficient of variance of $V_{reset}$ is 40% when read at the maximum of the LRS current, 28% when read at the first point where $dQ/d\\phi=0$ in the charge-flux plane, and 32% when read at the minimum of the current derivative. The set voltage shows no autocorrelation, whereas the reset voltage does, and the paper builds an ARIMA model of the form $V_{reset,t} = a\\,V_{reset,t-1} + b\\,\\epsilon_{t-1} + c$ that uses the previous cycle's reset voltage and previous modeling error to compute the current one. A two-level circuit-breaker network with quantum-point-contact conduction reproduces the measured set and reset I-V curves, locating the abrupt reset in the rupture of one or several rows of the percolation path. The paper also fits the experimental $Q$-$\\phi$ curves with a three-parameter compact expression, so the whole switching cycle can be represented statistically for circuit simulation.","pith_inferences":["The paper does not discuss it, but if reset voltages are autocorrelated, h-BN-based true random number generators that sample reset events should test for and remove this serial dependence before treating the values as entropy.","The contrast between autocorrelated reset and independent set suggests that filament rupture retains memory of the previous filament geometry while filament formation does not; this could be tested by comparing ARIMA coefficients across h-BN thicknesses and electrode metals.","A practical extension would be to track the fitted $n$ parameter of the charge-flux model over device lifetime, since the paper shows it correlates with $Q_{reset}$ and $\\phi_{reset}$, to see whether it drifts with cycling or device degradation.","The charge-flux extraction method could be applied to other noisy two-dimensional-material memristors to test whether the observed ordering of coefficients of variance (maximum-current method highest, charge-flux method lowest) is a general property or specific to this stack."],"forward_implications":["Cycle-to-cycle reset variability in h-BN memristors can be captured by a one-lag ARIMA model, so statistical compact models should include a reset-voltage memory term rather than drawing each value independently.","Because $V_{set}$ shows no autocorrelation, set events can be modeled as independent draws in the same statistical framework.","Reported variability figures for reset voltage are not intrinsic constants; the coefficient of variance ranges from 28% to 40% depending on extraction method, so comparisons between devices or studies should quote the extraction technique.","The three-parameter charge-flux model fits experimental $Q$-$\\phi$ curves and provides statistically distributed parameters suitable for circuit-level simulation.","The two-level circuit-breaker simulator reproduces the measured I-V curves and shows that reset corresponds to the rupture of one or several rows of the percolation path, consistent with the abrupt current drop."],"supporting_citations":[{"why":"Supplies the compact charge-flux expression used to fit the experimental Q-phi curves.","marker":"[40]"},{"why":"Establishes the time-series statistical methodology for resistive-switching variability that the ARIMA analysis applies here.","marker":"[15]"},{"why":"Previous study of cycle-to-cycle variability in h-BN memristors that this work extends with full series analysis.","marker":"[23]"},{"why":"Source of the parameter-extraction and variability-estimation techniques used to obtain the three reset voltages.","marker":"[26]"},{"why":"Provides the circuit-breaker simulator with quantum-point-contact conduction used to reproduce the set and reset I-V curves.","marker":"[31]"},{"why":"Defines the ARIMA modeling procedure used to construct the reset-voltage time-series model.","marker":"[42]"}],"fun_headline_variants":["h-BN memristor reset voltage autocorrelates across cycles","Cycle-to-cycle reset variability in h-BN devices: 28-40%","ARIMA model forecasts h-BN memristor reset voltage","Reset voltage shows memory in h-BN memristive devices","Three extraction methods quantify h-BN reset variability"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The argument assumes that the point where the device current collapses to near zero marks the reset transition itself, rather than the end of the reset process; if that point is post-rupture, then the 28% coefficient of variance obtained from the charge-flux method is measuring a different physical quantity than the other two methods, and the comparison that anchors the variability analysis is not apples-to-apples.","fun_headline_variants_meta":{"raw":{"variants":["h-BN memristor reset voltage autocorrelates across cycles","Cycle-to-cycle reset variability in h-BN devices: 28-40%","ARIMA model forecasts h-BN memristor reset voltage","Reset voltage shows memory in h-BN memristive devices","Three extraction methods quantify h-BN reset variability"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000881,"raw_usage":{"total_tokens":3821,"prompt_tokens":974,"completion_tokens":2847,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":590,"completion_tokens_details":{"reasoning_tokens":2758}},"tokens_in":590,"tokens_out":2847,"duration_ms":21489,"temperature":1.0,"reasoning_tokens":2758,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T15:48:10.984263+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Measure, over the same set of switching cycles, the voltage at the onset of the abrupt current decrease during reset and the voltage where $dQ/d\\phi$ first becomes null; if the null point consistently occurs after the current has already dropped by more than half of its low-resistance value, then $V_{reset2}$ is a post-rupture marker and the coefficient-of-variance comparison across the three methods does not describe the same transition.","supporting_citations":[{"cited_title":"Exploring Resistive Switching based Memristors in the Charge -Flux Domain, a modeling Approach","cited_arxiv_id":null,"evidence_quote":"Supplies the compact charge-flux expression used to fit the experimental Q-phi curves."},{"cited_title":"Time series statistical analysis: a powerful tool to evaluate the variability of resistive switching memories","cited_arxiv_id":null,"evidence_quote":"Establishes the time-series statistical methodology for resistive-switching variability that the ARIMA analysis applies here."},{"cited_title":"Time series modeling of the cycle-to-cycle variability in h-BN based memristors,","cited_arxiv_id":null,"evidence_quote":"Previous study of cycle-to-cycle variability in h-BN memristors that this work extends with full series analysis."},{"cited_title":"Variability estimation in resistive switching devices, a numerical and kinetic Monte Carlo perspective","cited_arxiv_id":null,"evidence_quote":"Source of the parameter-extraction and variability-estimation techniques used to obtain the three reset voltages."},{"cited_title":"Comprehensive study on unipolar RRAM charge conduction and stochastic features, a simulation approach","cited_arxiv_id":null,"evidence_quote":"Provides the circuit-breaker simulator with quantum-point-contact conduction used to reproduce the set and reset I-V curves."},{"cited_title":"Time series analysis and forecasting by example","cited_arxiv_id":null,"evidence_quote":"Defines the ARIMA modeling procedure used to construct the reset-voltage time-series model."}],"review_version":1}