{"id":"b2714549-f95a-4b47-ba2f-89732953d2d5","arxiv_id":"2607.26815","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":3.5,"correctness_risk":"medium","formal_verification":"none","parameter_count":10,"one_line_summary":"A leakage-extended Yakopcic model fitted to volatile TiO2 memristor data is implemented in SPICE and shown to reproduce measured I–V/pulse behavior and support a LIF neuron circuit.","lead":"Researchers built a SPICE model of a volatile TiO2 memristor by fitting a modified Yakopcic model to lab measurements, then showed it can drive a simple leaky integrate-and-fire neuron. The work is a practical device-to-circuit pipeline for people who need to simulate this class of short-term-memory memristors.","discovery_kind":"extension","skeptic_critique":{"model":"grok-4.5","headline":"Deterministic exponential leakage with x_eq=0 is under-constrained by the same noisy, metastable pulse data used to claim neuromorphic suitability.","rationale":"The reader correctly isolates the load-bearing leap: a deterministic single-state ODE plus simple exponential leakage is treated as adequate for neuromorphic circuit use despite the manuscript’s own documentation of stochastic filament behavior, metastable states, and large cycle-to-cycle READ scatter. That is precisely where the argument is least secure. The pipeline (measure → fit modified Yakopcic+Biolek+leak → SPICE → LIF) is real and honestly limited in §IX, so the verdict stays CONDITIONAL rather than REJECT; nothing in the text supplies held-out protocols, quantitative pulse-error metrics, or multi-device statistics that would remove the condition. No stronger internal inconsistency (e.g., algebraic error in the SPICE translation) appears; the negative A_n and hard-to-identify window p are secondary fitting quirks, not the central vulnerability. Concrete test above directly probes whether the omitted stochasticity matters for the exact application used to underwrite the claim.","tokens_in":11292,"tokens_out":681,"duration_ms":34396,"concrete_test":"Re-simulate the §VIII.C LIF circuit while drawing x_eq (or an additive state offset after each spike) from the empirical distribution of end-of-READ resistances in Fig. 2b (roughly 40–100 MΩ mapped through Eq. 3) instead of fixed x_eq=0; if spike rate, amplitude regularity, or the existence of sustained spiking changes materially relative to the deterministic baseline, the suitability claim for neuromorphic simulation weakens.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim (Abstract; §VIII.B–C) is that the leakage-extended Yakopcic SPICE model (Eqs. 3–6, Table 1) reproduces the Pt/TiO2/Ag device well enough for circuit-level neuromorphic use, shown by repeated LIF spiking. That claim rests on l(x)=−(x−x_eq)/τ with x_eq=0 and a single τ≈4.2 s extracted from averaged READ conductances (§VI.C, Eqs. 9–11). The paper itself records cycle-to-cycle READ scatter (40–100 MΩ), metastable intermediate states not captured by the leakage term (§VI.C–D), and stochastic filament fluctuations visible as voltage/resistance jitter inside the compliance region and during decay (§III.D, §VIII.B). Validation remains qualitative visual overlay on the same preprocessed traces used for NMSE coordinate-descent fitting (§VI.A–B, Fig. 8 vs Fig. 12); no held-out pulse protocol, no per-cycle error, and no multi-device statistics are reported. The LIF demo (§VIII.C) is purely simulated and explicitly requires the leakage term for repetitive spikes, yet operates on ~28 ms timescales while τ was fixed from 15 s READ averages. If the unmodeled metastability and cycle variance dominate state trajectories under repeated sub-threshold integration, qualitative average-trace agreement does not establish useful neuromorphic fidelity.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.5","summary":"The manuscript develops an end-to-end workflow for a volatile lateral Pt/TiO2/Ag memristor: experimental I–V and pulse characterization, selection of a Yakopcic-type model with Biolek window and an added exponential leakage term, coordinate-descent NMSE fitting of parameters (plus algebraic extraction of τ from READ decay), translation into a SPICE subcircuit, qualitative comparison of simulated vs. measured traces, and a purely simulated leaky integrate-and-fire neuron in which the memristor replaces the MOSFET threshold element. The central claim is that the fitted leakage-extended model reproduces the device’s pinched hysteresis (below compliance) and volatile relaxation well enough for circuit-level neuromorphic simulation.","tokens_in":11652,"tokens_out":1474,"duration_ms":31514,"significance":"If the fitted SPICE model is reliable under the excitation regimes of interest, the work supplies a practical, reusable subcircuit and a clear measurement-to-model pipeline for volatile TiO2 devices, which is useful for neuromorphic circuit designers who need device-level models rather than idealized non-volatile memristors. Strengths include the explicit leakage extension needed for repetitive spiking, the open SPICE listing, and the side-by-side measurement/simulation figures. Novelty is incremental: Yakopcic, Biolek, and forgetting/leakage terms are established; the contribution is primarily engineering integration and device-specific parameterization rather than a new physical model or multi-device statistical validation.","major_comments":[{"comment":"§VI.B–D and §VIII.B: Parameter identification (NMSE coordinate descent on compliance-stripped, filtered I–V conductance; τ from averaged READ decay via Eqs. 9–11 with x_eq=0) and ‘validation’ use the same device traces and protocols. No held-out pulse protocol, no quantitative error (e.g., per-cycle NMSE or resistance RMSE on unseen sequences), and no multi-device statistics are reported. The Abstract and §VIII claim of suitability for neuromorphic circuit simulation therefore rests on qualitative visual agreement with the fitting data. A minimum fix is a quantitative held-out comparison and explicit error bars or cycle statistics.","section":"§VI.B–D, §VIII.B"},{"comment":"§VI.C–D, §III.D, §VIII.B, §IX: The model assumes deterministic single-exponential leakage l(x)=−(x−x_eq)/τ with fixed x_eq=0 (Eq. 6, Table 1). The paper itself documents cycle-to-cycle READ scatter (40–100 MΩ), metastable intermediate states ‘not captured by the leakage term,’ and stochastic filament fluctuations. These are load-bearing for retention and for any circuit that integrates subthreshold activity over many cycles. Either demonstrate that average-trace fidelity still yields acceptable LIF/network statistics under the observed variance, or treat x_eq (or τ) as stochastic as the authors themselves propose only in future work.","section":"§VI.C, Eq. (6), Table 1"},{"comment":"§VIII.C vs §VI.C: The LIF demonstration uses ~28 ms pulse periods and produces repetitive spikes that explicitly require the leakage term, while τ≈4.2 s was extracted from 15 s READ averages. On the LIF timescale the leakage is almost frozen (Δx/x ~ few percent per cycle unless the state is driven near threshold every cycle). The manuscript should quantify how sensitive spike rate and threshold crossing are to τ and to the unmodeled metastability, and justify that a τ fitted on 15 s decays remains appropriate for tens-of-ms neuromorphic operation.","section":"§VIII.C, §VI.C"},{"comment":"Table 1 and §VI.D: An is negative (−0.0012), which inverts the usual polarity of the negative-threshold branch of g(u) (Eq. 4). The text notes this reflects decreasing memristance under negative bias, but does not show that the resulting dx/dt remains consistent with the window and with 0≤x≤1 under bipolar sweeps, nor that the fit is unique under coordinate descent. A short sensitivity or sign-constraint check is needed so that the SPICE subcircuit does not rely on an accidental local minimum.","section":"Table 1, Eq. (4), §VI.D"}],"minor_comments":[{"comment":"Fig. 9 / Leakage func: SPICE code implements Leakage(x)=−(x+x_eq)/tau while Eq. (6) is −(x−x_eq)/τ. With the fitted x_eq=0 the two coincide, but the listing is inconsistent with the equation and will be wrong if x_eq≠0 is used later.","section":"Fig. 9, Eq. (6)"},{"comment":"§VI.C: τ is reported as τ̄≈−4.2 s then used as τ=4.2 s in Table 1. Clarify that the minus sign is an artifact of the log formula and that the physical time constant is positive.","section":"§VI.C"},{"comment":"Notation: g(t) is used both for conductance and for the threshold function g(u); rename one to avoid confusion in §IV–VI.","section":"§IV–VI"},{"comment":"Fig. 1b and related plots: ‘normalized’ u(t), i(t), g(t) lack stated normalization constants; add them for reproducibility.","section":"Fig. 1b"},{"comment":"References: Biolek window and Yakopcic are appropriately cited; a brief pointer to other volatile/diffusive memristor SPICE models would help position the leakage extension.","section":"§V, References"},{"comment":"Minor typos: ‘Y akopcic’ spacing, ‘Indigital’ → ‘In digital’, ‘memristance’ vs resistance consistency, and arXiv date stamp ‘29 Jul 2026’ looks like a metadata error.","section":null}],"recommendation":"major_revision","confidential_remarks":"Fit for a methods/engineering-oriented venue in materials or neuromorphic circuits is reasonable; novelty is modest. The main risk is over-claiming ‘suitable for neuromorphic circuit simulation’ from single-device, same-data qualitative overlays. I would accept after major revision if the authors add quantitative held-out metrics and a clear statement of the deterministic model’s limits under the reported stochasticity. No integrity concerns."},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"This is a clean end-to-end device-to-SPICE pipeline on a lateral Pt/TiO2/Ag volatile memristor. What is actually new is the artifact: parameters (Table 1), the leakage-extended Yakopcic/Biolek equations in SPICE (Fig. 9), and a working LIF demo that needs the leak term for repeated spikes. The pieces themselves are known—Yakopcic I–V and thresholds, Biolek window, Chen-style forgetting—and the paper is honest about that.\n\nWhat it does well is the workflow. They check the pinched loop, strip compliance artifacts, fit with NMSE coordinate descent, extract τ≈4.2 s from READ decay under x_eq=0, translate to a reusable subcircuit, and overlay sim vs measure for both triangle and pulse. The limitations section is unusually straight: low CC, ON/OFF ~35, cycle scatter 40–100 MΩ, metastable states the leak term misses, and stochastic filament jitter the deterministic ODE cannot capture. That candor helps.\n\nThe soft spots are real but proportionate. Validation is same-distribution visual agreement on the traces used for fitting; no held-out protocol, no multi-device stats, no quantitative error beyond NMSE on preprocessed I–V. An is negative to match the observed polarity, which is fine empirically but undersells mechanism. The LIF runs on ~28 ms pulses while τ came from 15 s averages—illustrative, not a fidelity proof. The stress-test concern lands: if metastability and cycle variance dominate under repeated subthreshold integration, average-trace match does not guarantee neuromorphic usefulness. The paper already flags this for future stochastic work, so it is not hidden.\n\nMath and citations look fine; no invented entities, no load-bearing contradiction. Reproducibility is mid: full SPICE text in-paper, no data/code release.\n\nWho it is for: groups already building circuits around similar volatile oxide devices who need a drop-in subcircuit. Not for people hunting new switching physics or general memristor theory. I would send it to peer review—method is coherent, claims are checkable, limitations are owned. Engage if the device class matches your stack; otherwise skim the SPICE and move on.","headline":"Solid local engineering: a fitted, leakage-extended Yakopcic SPICE subcircuit for one volatile TiO2 stack, useful if you co-design with similar devices, not a modeling breakthrough.","tokens_in":12367,"tokens_out":577,"would_cite":false,"duration_ms":16897,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.5","headline":"A leakage-extended Yakopcic SPICE model fitted to a volatile TiO2 memristor reproduces its measured switching and runs a leaky integrate-and-fire neuron.","keywords":["memristor","TiO2","volatile switching","Yakopcic model","SPICE","neuromorphic computing","leaky integrate-and-fire","parameter fitting"],"falsifier":"Build the same leaky-integrate-and-fire circuit with the physical TiO2 devices and check whether the measured spike timing and recovery match the SPICE prediction under identical current-pulse trains; systematic mismatch in inter-spike intervals or failure to recover would falsify the claim that the fitted model is adequate for neuromorphic simulation.","tokens_in":12062,"feed_emoji":"⚡","tokens_out":919,"duration_ms":22481,"temperature":0.7,"pith_summary":"Circuit designers need device models that match real memristor measurements under both slow voltage sweeps and short pulses if they want trustworthy neuromorphic simulations. This paper takes experimental data from a fabricated volatile Pt/TiO2/Ag memristor, confirms its pinched hysteresis and second-scale resistance relaxation, and fits an extended Yakopcic model whose parameters are optimized to those traces. The fitted equations are packaged as a reusable SPICE subcircuit and shown to reproduce the main features of the measured I–V loops and pulse decay. The same subcircuit is dropped into a leaky integrate-and-fire neuron, where the leakage term lets the device recover and fire repeatedly. The result is a practical end-to-end path from lab data to circuit-level neuromorphic simulation for this class of volatile devices.","feed_headline":"Volatile TiO2 memristor gets a SPICE model that spikes","feed_subtitle":"Fitted leakage term lets the device recover and fire repeatedly in a leaky integrate-and-fire neuron","key_machinery":"The leakage-extended Yakopcic model: a hyperbolic-sine I–V relation coupled to a thresholded state ODE that includes a Biolek window and the relaxation term l(x)=−(x−x_eq)/τ, realized as two behavioral sources plus a 1 F integrator capacitor in a SPICE subcircuit.","core_discovery":"The authors show that a Yakopcic memristor model extended by a Biolek window and a simple exponential leakage term, with parameters optimized to their TiO2 device data, reproduces the measured pinched hysteresis below compliance and the volatile resistance relaxation under pulsed excitation well enough to serve as a SPICE subcircuit for neuromorphic circuit simulation, as demonstrated by repeated spiking in a memristor-based leaky integrate-and-fire neuron.","pith_inferences":["Because the paper already flags metastable intermediate states and cycle-to-cycle scatter, the next natural model upgrade is a stochastic x_eq or a multi-state filament description rather than a finer deterministic fit.","The very low compliance current that produces volatility also caps the ON/OFF ratio; the same modeling pipeline could quantify how raising compliance trades volatility for retention and thereby maps the usable design space for synaptic versus memory use.","Once the SPICE subcircuit is public, standard oscillator and network-stability tools can be applied to memristor LIF arrays without re-deriving device physics each time."],"forward_implications":["Volatile TiO2 memristors characterized at low compliance current can be dropped directly into SPICE netlists for neuromorphic design.","The leakage term is what enables repetitive spiking; without it the neuron would lock into a low-resistance state after the first fire.","The same five-stage workflow (data assessment, model selection, parameter fit, SPICE packaging, validation) can be reused for other volatile memristor technologies.","Circuit-level studies of larger memristive networks become possible once each device is represented by this parameterized subcircuit."],"fun_headline_variants":["TiO2 memristor model with leakage term drives LIF neuron spiking","Yakopcic-Biolek model plus exponential leak fits volatile TiO2 device","SPICE subcircuit matches TiO2 memristor hysteresis and pulse recovery","Optimized volatile TiO2 model enables memristor LIF circuit simulation","Leak-extended memristor model reproduces TiO2 relaxation and neuron fire"],"cache_read_input_tokens":128,"weakest_assumption_plain":"A single deterministic state variable that always relaxes exponentially toward a fixed equilibrium is enough to stand in for the device’s real, noisy, cycle-to-cycle filament dynamics.","fun_headline_variants_meta":{"raw":{"variants":["TiO2 memristor model with leakage term drives LIF neuron spiking","Yakopcic-Biolek model plus exponential leak fits volatile TiO2 device","SPICE subcircuit matches TiO2 memristor hysteresis and pulse recovery","Optimized volatile TiO2 model enables memristor LIF circuit simulation","Leak-extended memristor model reproduces TiO2 relaxation and neuron fire"]},"model":"grok-4.5","effort":"low","cost_usd":0.001894,"raw_usage":{"total_tokens":836,"prompt_tokens":726,"num_sources_used":0,"completion_tokens":90,"cost_in_usd_ticks":18944000,"prompt_tokens_details":{"text_tokens":726,"audio_tokens":0,"image_tokens":0,"cached_tokens":128},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":20,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":726,"tokens_out":90,"duration_ms":4422,"temperature":1.0,"reasoning_tokens":20,"cache_read_input_tokens":128,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-30T20:15:42.971239+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"Build the same leaky-integrate-and-fire circuit with the physical TiO2 devices and check whether the measured spike timing and recovery match the SPICE prediction under identical current-pulse trains; systematic mismatch in inter-spike intervals or failure to recover would falsify the claim that the fitted model is adequate for neuromorphic simulation.","supporting_citations":[],"review_version":1}