{"id":"ab3fd61e-566e-41e2-84e5-5bfffd790881","arxiv_id":"2505.15757","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"A state-dependent diode-resistor model with minimum-variance state estimation improves fitting of SDC memristor current-voltage data compared with two earlier model variants.","lead":"The paper proposes a physics-inspired formula for the current-voltage behaviour of self-directed channel memristors and a noise-aware way to estimate the device state from measurements. A generalist should read it to see how one commercial memristor type can be characterised for memory and neuromorphic applications.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The linear state-scaling assumption in Eq. 5 is untested; the improved in-sample fit in Table II may reflect added flexibility rather than a true physical relation.","rationale":"The reader identified the linear state-scaling assumption as the weakest point, and I concur. The paper is internally coherent: Eq. 5 is a plausible physics-inspired generalization, the corrected diode form (Eq. 6) is reasonable, and the minimum-variance estimator derivation (Eqs. 16-29) is mathematically sound, especially since the relative weights in Eq. 29 do not require knowing the absolute noise magnitude. The fitting errors in Table II support the proposed model on the collected data, but that support is limited by the fact that the evaluation is in-sample. The proposed model's extra state-dependence of the diode term is precisely the kind of added flexibility that would lower training error even if the assumed multiplicative coupling were false. A split-half or two-state-variable test would provide decisive evidence for or against the linear scaling. Given the absence of such a test, and the paper's own acknowledgement of this as an assumption, a conditional verdict is appropriate. The concern does not warrant rejection because the paper's assumptions are clearly stated, the model is a plausible extension of the Generalised MSS, and the reported parameters are physically interpretable. The verdict should remain CONDITIONAL, as originally assessed.","tokens_in":13317,"tokens_out":8001,"duration_ms":75306,"concrete_test":"For each measured state, use the globally fitted parameters from Table I and Eq. 16 to estimate x separately from readings with |v| below half the read amplitude and from readings with |v| above half. If the two estimates differ by more than the propagated measurement noise for a substantial fraction of states, the single linear-scaling model fails. A complementary test: fit the two-state-variable model i = x_1 G_m v + x_2 I_d(v) with per-state (x_1, x_2) and check whether x_2/x_1 is approximately constant across states; a non-constant ratio directly refutes the linear multiplier assumption of Eq. 5.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central physical assumption is that the scalar state x multiplies both the ohmic and diode conduction terms in Eq. 5 (Section III-C), so that higher x simultaneously raises the linear conductance and the Schottky-like exponential terms. The paper motivates this by a qualitative picture of barrier-width modulation, but provides no direct evidence for the linear coupling, or even for a single-parameter state. The only quantitative support is the in-sample fitting comparison of Table II: the proposed model has lower errors than the modified GMSS. However, the proposed model is strictly more flexible than modified GMSS: letting x vary per state also allows the diode branch to be scaled by a state-dependent factor, while modified GMSS fixes the diode branch to be state-independent. A more flexible model is expected to achieve a smaller training error even if the specific linear multiplicative structure is incorrect. The reported improvement therefore cannot, by itself, validate the physical interpretation of x. The manuscript does not report held-out validation, cross-validation, or consistency checks that would show the single-parameter form is not an artifact of the chosen functional form.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper addresses state characterisation of self-directed channel (SDC) memristors. It proposes a physics-inspired conduction model in Eq. (5), i = x(Gm v + alpha1(e^{beta1 v} - 1) + alpha2(1 - e^{-beta2 v})), in which a scalar state x multiplies both the ohmic and diode branches, together with a corrected zero-crossing diode form. Using data from a single Knowm SDC device, the authors fit model parameters by grid search and per-measurement Levenberg-Marquardt state values, then compare fitting errors against the Generalised MSS and a modified version. They invert the fitted model to obtain a state estimate from noisy voltage and current pairs and derive minimum-variance weights (Eq. 29) under a correlated additive noise model. The estimator is demonstrated on 60-minute resistive drift measurements for two initial states.","tokens_in":13495,"tokens_out":6150,"duration_ms":52268,"significance":"If the model and estimator were validated, the work would provide a practical state characterisation protocol for SDC memristors, with a physically motivated nonlinear IV family and an uncertainty-aware state estimate. The paper deserves credit for identifying the zero-crossing defect in the GMSS diode form, for explicitly deriving the variance-proportional weights in Eq. (29), and for demonstrating that a state-dependent diode branch reduces in-sample fitting error. However, the empirical support is currently limited because all model comparisons are in-sample on a single device and the state estimator is the inverse of the same fitted model; the central claims therefore need additional validation before the results can be considered established.","major_comments":[{"comment":"All fitting and evaluation are performed on the same single-device dataset, and the proposed model is strictly more flexible than modified GMSS: it allows the diode branch to scale with x while modified GMSS keeps the diode state-independent. The lower errors in Table II therefore do not, by themselves, establish that the linear multiplicative structure in Eq. (5) is the correct physical form. Please add held-out validation (for example, fitting on a subset of states or cycles and evaluating on the rest) or a model-selection criterion that penalises parameter count, and report parameter uncertainty from the grid-search procedure.","section":"Section V-A, Table II"},{"comment":"The state estimate used in the drift demonstration is the algebraic inverse of the fitted model evaluated with the Table I parameters, so the demonstration in Fig. 7 reuses the same fitted model that produced the data fit. Without an independent ground-truth state measurement or at least a held-out state prediction, the experiment cannot confirm that x corresponds to a physical device state; it only shows that the estimator produces a smooth curve. Please validate the estimator on states not used in fitting, or compare it against an independent conductance or resistance readout.","section":"Section VI-C, Eq. (16)"},{"comment":"The experimental section is incomplete and partly subjective: the number of state measurements is given as \"X separate state measurements\", the read amplitude is \"Aread in [min, max]\" with the actual values left blank, and the filtering criterion is a manual review of Lissajous figures. These gaps make the dataset non-reproducible and make it impossible to assess whether the state range and measurement conditions are representative. Please supply the missing values and an objective, reproducible filtering rule.","section":"Section IV-B, IV-C"},{"comment":"The minimum-variance derivation assumes that the additive noise N is perfectly correlated between the applied-voltage and resistor-voltage measurements, so that the noise cancels exactly in vmemristor. This assumption is stated but not tested, and the subsequent exclusion of low-magnitude measurements (below 30% of the maximum) is an ad hoc patch. If the correlation is imperfect, the variance expression in Eq. (27) and the weights in Eq. (29) are not minimum-variance. Please characterise the measurement noise empirically or justify the assumption from the instrument architecture.","section":"Section VI-A, Eq. (20)"}],"minor_comments":[{"comment":"Equation (2) defines gx(i) = f(x) * i, but the readout is a function from voltage to current; this should be gx(v) = f(x) * v (or equivalent), otherwise the notation is inconsistent with Eq. (1).","section":"Section III, Eq. (2)"},{"comment":"The text says \"choose voltage regions\" but the clustering is applied to currents; please align the terminology with the actual procedure.","section":"Section IV-D1"},{"comment":"Figure 7 has unreadable axis labels made of embedded font tokens such as \"/uni00000013\"; the estimated-state axis and legend need to be rendered properly.","section":"Figure 7"},{"comment":"The sentence stating that multiplying by a constant multiplies noise power by the constant is inconsistent with Eq. (28); the noise power is multiplied by the square of the constant.","section":"Section VI-B"},{"comment":"Table II contains stray spaces in entries such as \"3 .152\" and \"0 .2325\"; please fix the formatting.","section":"Table II"},{"comment":"There is a typo, \"memrsitive\", in the description of modulation of the memristive state.","section":"Section II-A"},{"comment":"The discussion asserts that the results suggest the state-dependence is \"approximately linear\", but no statistical test or confidence interval is given to support this claim.","section":"Section V-B"}],"recommendation":"major_revision","confidential_remarks":"The paper is a reasonable fit for cs.ET, and the underlying idea is publishable if the validation issues are resolved. My main concern is that the empirical core is currently an in-sample curve fit; a revision with held-out validation, a complete dataset description, and an independent state check would substantially strengthen it. The authors should also be asked to provide data or code availability."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should know: this paper does not break new physical ground, but it offers a genuinely useful, physics-inspired state model for SDC memristors. The two modifications to the Generalized MSS framework—state-dependent diode scaling (Eq. 5) and the zero-crossing-corrected diode form (Eq. 6)—are new in this combination, and the weighted minimum-variance estimator (Eq. 29) is a sensible adaptation of a standard result. The authors are also honest about their assumptions: Markovian state, linear state scaling, and additive correlated noise are all flagged explicitly.\n\nThe soft spot is real: the central physical claim—that the scalar state x linearly multiplies both the ohmic and diode conduction terms—is supported only by in-sample fits on a single device. The proposed model is strictly more flexible than the baselines, so a lower training error is expected even if the specific multiplicative structure is wrong. There is no held-out validation, no cross-validation, no error bars, and Figure 7 re-uses the same fitted model to compute the state estimates, so it does not independently confirm the state variable. The manuscript also has mechanical problems: placeholder values like “X”, “min”, and “max” in the experimental section, and a garbled figure caption. These are sloppy but not fatal.\n\nThat said, the math holds up. The estimator derivation is coherent, the fitting comparison, while weak, is not contradicted by the presented data, and the physical motivation is plausible. This is a solid engineering paper, not a breakthrough. It would benefit from a major revision: add held-out validation or at least error bars, report the actual experimental parameters, clean up the placeholders, and either justify the linear scaling with more data or soften the physical interpretation.\n\nThe paper is aimed at researchers working on memristor characterization for memory and neuromorphic applications. It deserves a serious referee and, with revision, could be a useful contribution. I would not recommend rejecting it outright, but acceptance should be conditional on addressing the validation gap.","headline":"A plausible incremental state model for SDC memristors, but the missing held-out validation and single-device data keep the central claim conditional.","tokens_in":14068,"tokens_out":2251,"would_cite":false,"duration_ms":20708,"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":"A diode-aware model recovers SDC memristor state from noisy data","keywords":["memristor","self-directed channel","state characterisation","state estimation","Schottky diode model","Generalised Metastable Switch Model","minimum-variance estimation","resistive switching"],"falsifier":"A direct falsifying experiment would be to program a single SDC device into several distinct states, record full IV curves for each, and check whether the ratio of the diode contribution to the ohmic contribution, $I_d(v)/G_m v$, is exactly constant across states for every voltage; any state-dependent change in that ratio (beyond the common multiplier $x$) would violate Eq. (5). Equivalently, if the fitted diode parameters $\\alpha_1,\\alpha_2,\\beta_1,\\beta_2$ must be re-fit per state to maintain good agreement, the linear multiplier assumption fails.","tokens_in":13096,"feed_emoji":"⚡","tokens_out":6775,"duration_ms":51686,"temperature":0.7,"pith_summary":"This paper aims to establish a physics-inspired way to characterize the resistive state of self-directed channel (SDC) memristors, a type of redox-based resistive memory. The central proposal is Eq. (5), in which a single scalar state $x$ multiplies both an ohmic term and a corrected, zero-crossing Schottky diode term: $i = x(G_m v + \\alpha_1(e^{\\beta_1 v}-1)+\\alpha_2(1-e^{-\\beta_2 v}))$. The authors report that this state-dependent diode form fits measured current–voltage data better than the Generalised Metastable Switch Model and a modified version, and they derive a minimum-variance estimator that recovers $x$ from noisy voltage and current pairs. If correct, this gives a practical readout procedure for tracking the state of SDC memristors in memory and neuromorphic applications.","feed_headline":"A diode-aware model recovers SDC memristor state from noisy data","feed_subtitle":"A scalar state multiplies ohmic and diode terms; minimum-variance weighting recovers it from noisy voltage-current pairs.","key_machinery":"The central object is the state-parameterised current–voltage relation of Eq. (5), $$i = x\\,(G_m v + \\alpha_1($e^{{\\beta_1 v}}$-1) + \\alpha_2(1-$e^{{-\\beta_2 v}}$)),$$ in which the scalar state $x$ multiplies both an ohmic conductance $G_m v$ and a corrected diode component. The diode correction replaces the two independent exponentials of the original Generalised MSS form with two zero-crossing exponential terms, forcing zero current at zero voltage while allowing different forward and reverse parameters. The inverse of this relation, $x = g(v,i)$, is then linearised in the measurement noise to compute per-sample variances, and the minimum-variance state estimate is formed by weighting each sample inversely to its variance (Eq. 29).","core_discovery":"On the paper's own terms, the discovery is that the conduction of SDC memristors is not captured by treating the device as a resistor with memory, nor by a linear ohmic term plus state-independent diode exponentials. Instead, the authors propose and fit a model where the resistive state $x$ acts as a common multiplier on both the ohmic conductance and the two Schottky diode conduction components, with the diode terms written in a zero-crossing form that lets forward and reverse exponential parameters be independent. Fitting this form to measurements yields lower error on all four reported metrics (MSE, MAE, MRE, MRSE) than the Generalised MSS and a modified Generalised MSS baseline, and the inverse of the model, combined with a minimum-variance weighting of multiple noisy measurements, provides a state estimate whose uncertainty is quantified and whose dynamics can be tracked over time.","pith_inferences":["If the linear multiplier relation holds across a wider range of states and devices, the same scalar state could serve as a single readout for both ohmic and barrier-limited conduction, simplifying circuit-level state tracking.","A natural testable extension is to compare this model on the same device at different temperatures, since Schottky barrier width and hopping-site density change with temperature; the model predicts the same functional form with re-scaled $x$ only if the linear multiplier is physical.","The paper's correction to the diode form suggests that previously published memristor models using non-zero-crossing exponentials may have over-parameterised fits; re-examining them with the zero-crossing constraint could change reported state estimates."],"forward_implications":["The state of an SDC memristor can be estimated from ordinary noisy voltage and current readouts, with an explicit variance for the estimate, instead of assuming noiseless measurements.","Multilevel storage and drift monitoring become practical: the estimated state trajectory over time can reveal metastability and resistive drift in the device.","The improved fit suggests that programming protocols should account for the state's effect on the Schottky barrier, not just the ohmic resistance.","The model's state variable $x$ is not an instantaneous resistance (since the VI relation is nonlinear), so it provides a more faithful readout quantity for applications.","The minimum-variance weighting procedure is general and could be applied to other nonlinear memristor models with a similar scalar state."],"supporting_citations":[{"why":"Provides the resistive-state/readout framework that the paper generalises from linear to nonlinear VI relations.","marker":"[7]"},{"why":"Introduces the metastable switch memristor model from which the Generalised MSS baseline derives.","marker":"[18]"},{"why":"Presents the Generalised Metastable Switch Memristor Model whose diode form the paper corrects and extends.","marker":"[19]"},{"why":"Supplies the sinh and Schottky diode conduction picture used to motivate the physical model.","marker":"[9]"},{"why":"Describes the self-directed channel memristor device structure and switching mechanism used to justify the model.","marker":"[17]"},{"why":"Reports the authors' earlier experiments showing the IV characteristic is exponential and asymmetric, motivating the model.","marker":"[21]"},{"why":"The Levenberg–Marquardt algorithm used for fitting the state variable during parameter selection.","marker":"[22]"}],"fun_headline_variants":["Diode-aware model recovers SDC memristor state from noisy data","Scalar multiplier state: new model for SDC memristor characterization","Noise-aware state estimation for self-directed channel memristors","Physics-inspired model with minimum-variance weighting recovers memristor state"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The model assumes that a single scalar state multiplies both the ohmic and the diode conduction terms linearly, motivated by the picture that more hopping sites shrink the Schottky barrier width; if this linear relation does not hold for SDC devices, the improved fit is an artifact of the chosen functional form rather than a physical state characterisation.","fun_headline_variants_meta":{"raw":{"variants":["Diode-aware model recovers SDC memristor state from noisy data","Scalar multiplier state: new model for SDC memristor characterization","Noise-aware state estimation for self-directed channel memristors","Physics-inspired model with minimum-variance weighting recovers memristor state"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000961,"raw_usage":{"total_tokens":4036,"prompt_tokens":833,"completion_tokens":3203,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":449,"completion_tokens_details":{"reasoning_tokens":3125}},"tokens_in":449,"tokens_out":3203,"duration_ms":22839,"temperature":1.0,"reasoning_tokens":3125,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T15:11:26.759604+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A direct falsifying experiment would be to program a single SDC device into several distinct states, record full IV curves for each, and check whether the ratio of the diode contribution to the ohmic contribution, $I_d(v)/G_m v$, is exactly constant across states for every voltage; any state-dependent change in that ratio (beyond the common multiplier $x$) would violate Eq. (5). Equivalently, if the fitted diode parameters $\\alpha_1,\\alpha_2,\\beta_1,\\beta_2$ must be re-fit per state to maintain good agreement, the linear multiplier assumption fails.","supporting_citations":[{"cited_title":"Event-Based Simula- tion of Stochastic Memristive Devices for Neuromorphic Computing,","cited_arxiv_id":null,"evidence_quote":"Provides the resistive-state/readout framework that the paper generalises from linear to nonlinear VI relations."},{"cited_title":"AHaH Computing-From Metastable Switches to Attractors to Machine Learning,","cited_arxiv_id":null,"evidence_quote":"Introduces the metastable switch memristor model from which the Generalised MSS baseline derives."},{"cited_title":"The Generalized Metastable Switch Memristor Model,","cited_arxiv_id":null,"evidence_quote":"Presents the Generalised Metastable Switch Memristor Model whose diode form the paper corrects and extends."},{"cited_title":"Memristive switching mechanism for metal/oxide/metal nanodevices,","cited_arxiv_id":null,"evidence_quote":"Supplies the sinh and Schottky diode conduction picture used to motivate the physical model."},{"cited_title":"Self-directed channel memristor for high temperature operation,","cited_arxiv_id":null,"evidence_quote":"Describes the self-directed channel memristor device structure and switching mechanism used to justify the model."},{"cited_title":"Robust memristor networks for neuromorphic computation applications,","cited_arxiv_id":null,"evidence_quote":"Reports the authors' earlier experiments showing the IV characteristic is exponential and asymmetric, motivating the model."},{"cited_title":"An Algorithm for Least-Squares Estimation of Nonlinear Parameters,","cited_arxiv_id":null,"evidence_quote":"The Levenberg–Marquardt algorithm used for fitting the state variable during parameter selection."}],"review_version":1}