{"id":"a60887bb-41c0-4eb1-83e5-5fb02f7f894a","arxiv_id":"2502.03912","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"Low-frequency noise measurements on MIS and MIOS SiN RRAMs are used to argue that multilevel switching comes from nitrogen-vacancy density changes, although the paper's own model predicts a different noise scaling.","lead":"This paper measures low-frequency current noise in silicon nitride resistive memory cells, with and without a thin tunneling oxide, to understand how they switch between four resistance levels. The authors conclude that the switching comes from changes in the number of nitrogen vacancies, but that conclusion does not follow from their own noise model.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Eq. (1) predicts Si/I2 ~ R^2 for the claimed vacancy-density mechanism, yet Fig. 7 shows Si/I2 nearly constant across resistance states; the mechanistic conclusion is internally inconsistent.","rationale":"The reader's weakest_assumption focused on whether Eq. (1), borrowed from oxide RRAMs, applies to SiN and whether the scaling uniquely identifies the switching mechanism. My read finds a sharper problem: even if Eq. (1) is taken at face value and only one mechanism is assumed to change, the reported constancy of Si/I2 across resistance states is incompatible with the paper's own enumerated mechanisms, including the claimed nitrogen-vacancy-density change. The reader's rationale does mention the ~1/n^2 prediction versus roughly constant data, so there is partial overlap, but the formal weakest_assumption stops short of stating the internal inconsistency. My concrete test would settle the point by quantifying how much normalized noise should move under the claimed mechanism and comparing it with the measured scatter; if the test confirms the inconsistency, the paper's mechanistic conclusion fails even before considering model-transfer questions. The data may still be useful as a preliminary experimental report on multi-level SiN RRAMs and on the tunneling oxide's role, but the central mechanistic claim as written is not supported. Therefore I do not change the reader's REJECT verdict; the internal contradiction strengthens it.","tokens_in":5013,"tokens_out":5296,"duration_ms":51915,"concrete_test":"For each device plotted in Fig. 7, take the programmed resistance range from the DC current at 0.1 V and the reported scatter in normalized PSD. Under the paper's mechanism, log10(Si/I2) should shift by roughly 2 * log10(R_max/R_min) between the lowest and highest resistance states. Recompute this predicted shift from the figure's resistance values and compare it with the measured normalized-PSD spread. If the predicted shift exceeds the measured spread by more than a factor of 10 and is not absorbed by error bars, the attribution to nitrogen-vacancy density is refuted. Also report the fitted slope of log10(Si/I2) versus log10(R) with confidence intervals across all MIS and MIOS devices; a slope indistinguishable from 0 while R changes means Eq. (1) cannot explain the resistance modulation.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central attribution in Section III-D does not follow from the paper's own noise model. Eq. (1) gives Si/I2 = 2kT N_it / (n^2 pi r^3 t_SiN f). For fixed filament geometry and interface quality, changing the nitrogen-vacancy density changes the free carrier density n, so the programmed resistance obeys R proportional to 1/n, and the model predicts Si/I2 proportional to 1/n^2, i.e., proportional to R^2. The other two mechanisms enumerated in the same paragraph also predict state-dependent normalized noise: adding CFs gives Si/I2 proportional to R, and changing CF radius gives Si/I2 proportional to R^(3/2). None of these mechanisms yields a constant normalized PSD while R varies. However, Section III-C and Fig. 7 report that normalized noise levels are 'almost the same' across the programmed resistance states, with less than one decade of scatter after normalization. Thus, under Eq. (1), the only way to keep Si/I2 constant is for n, r, and the CF number to remain unchanged, which contradicts the observed resistance change. Invoking nitrogen vacancies as the source of resistance modulation is precisely a change in n, so the conclusion is not just unsupported by the model; it is opposite to the model's prediction. The paper would need an additional, unstated mechanism (for example, correlated changes in N_it or geometry that exactly compensate) to reconcile the data with the conclusion. The dismissal of Dev.#2 as an outlier without a statistical criterion and the absence of error bars further weaken the inference, but the internal contradiction is the decisive problem.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript reports low-frequency noise (LFN) measurements on silicon-nitride-based MIS and MIOS resistive-switching devices programmed to multiple resistance levels. Using a custom measurement setup, the authors record current noise at a fixed bias and find that the normalized power spectral density Si/I2 remains approximately constant across resistance states. They interpret this behavior through a carrier-number-fluctuation model for a cylindrical conductive filament (Eq. (1)), concluding that the multilevel high-resistance switching is mainly caused by a change in the number of nitrogen vacancies that modulate the filament conductivity, without degrading the surrounding interface quality. The presence of a tunneling oxide is argued to improve switching control while keeping noise levels comparable.","tokens_in":5318,"tokens_out":3323,"duration_ms":33234,"significance":"If the central conclusion were sound, it would usefully connect LFN signatures to the microscopic switching mechanism in SiN-based RRAMs and would strengthen the case for using such devices in multilevel and neuromorphic applications. The experimental methodology is careful in several respects: measurements are made at constant bias, spectra are extrapolated at a fixed frequency, and normalized PSDs are compared across device structures. The paper also avoids parameter fitting and does not over-claim circularity; the noise model is adopted from prior work. However, the central inference is internally inconsistent with the paper's own model: Eq. (1) predicts Si/I2 proportional to 1/n^2 for a change in carrier density, while the data show near-constant normalized noise. The small device count and the unsubstantiated exclusion of an outlier further weaken the empirical claims. On balance, the manuscript does not establish its main attribution.","major_comments":[{"comment":"The central mechanistic conclusion is contradicted by the paper's own noise model. Equation (1) gives Si/I2 = 2kT N_it / (n^2 pi r^3 t_SiN f). For fixed filament geometry and interface trap density, changing the nitrogen-vacancy density changes the free carrier density n, and since the programmed resistance R scales as 1/n, the model predicts Si/I2 proportional to R^2. However, Section III-C and Fig. 7 report that the normalized noise is 'almost the same' across resistance states, with less than one decade of scatter after normalization. Thus, under Eq. (1), the data imply that n, r, and filament number remain essentially unchanged, which is incompatible with the claimed vacancy-density-driven resistance change. The conclusion would require an additional, unstated mechanism (e.g., correlated changes in N_it or geometry that exactly compensate) to be consistent with the model; the manuscript does not provide such a mechanism.","section":"III-D, Eq. (1) and Fig. 7"},{"comment":"The dismissal of Dev.#2 as an outlier is not justified by any statistical criterion. The text states that 'the number of measured devices is small in order to reach solid conclusions,' yet it excludes this device when concluding that MIOS devices exhibit a 'superior degree of control over variability.' With only five MIOS devices measured, one excluded datum materially changes the average and the apparent alignment of the remaining devices. A quantitative outlier test (e.g., residual threshold or a robust fit) or a discussion of the sensitivity of the conclusions to this exclusion is needed.","section":"III-B"}],"minor_comments":[{"comment":"The manuscript contains several standalone 'apply.' fragments (e.g., after Fig. 2 and in Section III), which appear to be formatting artifacts and should be removed.","section":"II and III"},{"comment":"The phrase 'declines from this trend' should be 'deviates from this trend'.","section":"III-B"},{"comment":"The sentence about the average noise level derived from a set of five devices does not specify whether the set is MIS, MIOS, or mixed; the composition of the averaged set should be stated explicitly, along with the number of devices in each structure type.","section":"III-C"},{"comment":"The symbol t_SiN is used in Eq. (1) but is not defined in the surrounding text before the equation; it should be introduced as the nitride layer thickness.","section":"III-D, Eq. (1)"},{"comment":"The sentence on fabrication states that details are given in [3] and then refers to [12] for the experimental setup; please ensure the citation correctly identifies the source for each component, as the current phrasing is ambiguous.","section":"II"}],"recommendation":"reject","confidential_remarks":"The paper's main attribution is internally inconsistent with its own Eq. (1), and the outlier handling is not statistically defensible. The experimental data (constant normalized noise) could be reported as an interesting empirical observation, but the conclusion that nitrogen-vacancy density changes the filament conductivity is not supported by the analysis presented. A substantial revision would be needed to either supply a compensating mechanism, revise the model, or substantially weaken the claim; within the current scope, the load-bearing error cannot be fixed without new analysis or new data. The manuscript is also very short for a full journal article; it may be better suited to a conference proceedings if the claims are appropriately restrained."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear colleague,\n\nShort version: if you want an example of a paper where the experiments outrun the interpretation, read this. It compares low-frequency noise in SiN RRAMs with and without a tunneling oxide across four programmed resistance states and reports that normalized noise stays roughly flat. The raw data are probably useful. The conclusion—that multilevel switching comes from a change in nitrogen-vacancy density while the filament interface stays pristine—is not supported; it is actually the opposite of what their Eq. (1) predicts.\n\nWhat is new: an LFN-based comparison of MIS vs MIOS SiN RRAMs at multiple resistance levels, with the observation that the tunneling oxide makes lower-resistance programming easier without raising noise. That is a concrete, useful result for the RRAM subfield, especially the noise-vs-current scaling in Figs. 5 and 6.\n\nThe soft spot is load-bearing. Eq. (1) gives Si/I^2 proportional to 1/n^2, where n is the free carrier density in the filament. If you change the resistance by changing the number of nitrogen vacancies, you change n, so normalized PSD should vary as R^2. The other mechanisms they enumerate also predict state-dependent normalized noise. Their own Fig. 7 shows it is nearly constant across states. That means, under Eq. (1), n, r, and the number of filaments are all effectively fixed, which is incompatible with the claimed vacancy-density mechanism. Invoking nitrogen vacancies is precisely a change in n. To keep Si/I^2 constant you would need an unstated compensatory change in geometry or trap density, and the paper neither states nor measures one.\n\nThere are also smaller issues: Dev.#2 is dismissed as an outlier without a statistical criterion; the device count is thin; no error bars on the normalized PSD averages. These are minor compared to the internal contradiction.\n\nFor a reader: the experimental part is a decent data point on tunneling-oxide effects in nitride RRAMs. The mechanistic section should be rewritten or dropped. I would not cite the vacancy-density claim, but I could see citing the MIS/MIOS noise comparison if I worked in that area.\n\nI would send it to peer review rather than desk-reject: a good referee can force the authors to confront the inconsistency, and the experimental comparison deserves an outlet. But they need to redo the interpretation before publication.","headline":"Nice MIS/MIOS noise data, but the vacancy-density conclusion contradicts the paper's own Eq. (1), which predicts normalized noise should scale as R^2 while the data show it flat.","tokens_in":5871,"tokens_out":2164,"would_cite":false,"duration_ms":20211,"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":"Low-frequency noise measurements on silicon-nitride resistive memories with and without a tunneling oxide show that multilevel high-resistance switching is set by the nitrogen-vacancy density in the conductive filament, with the…","keywords":["Silicon Nitride","RRAM","Multi-Level","Low-Frequency Noise","Random Telegraph Noise","Conductive Filament","Nitrogen Vacancies","Tunneling Oxide"],"falsifier":"Fabricate SiN RRAMs with deliberately varied nitrogen-vacancy content (for example, through deposition stoichiometry or controlled SET currents), measure $S_i/I^2$ at fixed resistance, and check whether the normalized noise tracks the model's inverse-square dependence on carrier density; a flat $S_i/I^2$ across samples with changed vacancy density would falsify the attribution. Alternatively, direct cross-sectional imaging of filaments at different resistance states would show whether the filament radius or interface changes, which the current claim rules out.","tokens_in":4836,"feed_emoji":"⚡","tokens_out":9321,"duration_ms":80149,"temperature":0.7,"pith_summary":"This paper uses low-frequency noise (LFN) measurements to identify what changes when a silicon-nitride resistive memory is programmed into one of four resistance states. Devices with and without a 2.1 nm tunneling oxide both show that the normalized noise power $S_i/I^2$ tracks resistance across states, staying within roughly an order of magnitude. Interpreting this through a carrier-number-fluctuation model for a cylindrical conductive filament, the paper argues that multilevel high-resistance switching is governed by the number of nitrogen vacancies that set the filament's conductivity, not by creating new filaments or degrading the surrounding interface. The tunneling oxide makes lower resistance states easier to set while leaving the noise level unchanged, so it can be added to the stack without a noise penalty.","feed_headline":"Noise data trace multilevel SiN memory switching to nitrogen vacancies","feed_subtitle":"Flat noise across resistance levels points to vacancy-driven switching with an intact filament interface.","key_machinery":"The load-bearing object is Eq. (1), a low-frequency-noise relation for a cylindrical metal-like conductive filament: $S_i/I^2 = S_N/N^2 = kT N_{it} 2\\pi r t_{SiN}/(n^2(\\pi r^2 t_{SiN})^2 f) = 2kT N_{it}/(n^2 \\pi r^3 t_{SiN} f)$. Here $r$ is the filament radius, $t_{SiN}$ its height, $n$ the equivalent free-carrier density, and $N_{it}$ the surface-trap density on the filament's surrounding interface. The relation discriminates among switching mechanisms because adding filaments lowers the normalized noise inversely with filament number, changing the radius alters it as $1/r^3$, and changing the carrier density alters it as $1/n^2$; the paper uses the measured stability of $S_i/I^2$ to select the vacancy-density interpretation with $N_{it}$ unchanged.","core_discovery":"The central claim is that the multilevel high-resistance states of the examined SiN RRAM cells arise from a change in the number of nitrogen vacancies that modulate the conductive filament's conductivity, while the quality of the filament's surrounding interface remains unaffected. This follows from the observation that the 1/f noise component, normalized by $I^2$, stays stable as the programmed resistance changes, and from applying a model in which the normalized power spectral density depends on filament number, radius, and carrier density in distinct ways. The paper also reports that inserting a SiO2 tunneling oxide between the nitride and the silicon bottom electrode lowers the resistance required for setting states and reduces device-to-device variability in the noise, without raising the average normalized noise.","pith_inferences":["A direct test would be to engineer SiN films with different nitrogen-vacancy concentrations and check whether $S_i/I^2$ follows the model's $1/n^2$ dependence; the paper does not report such a control experiment.","The evidence for interface stability is indirect: it is inferred from the stability of normalized noise, not from a direct microscopic measurement of the filament edge. A microscopy or trap-spectroscopy measurement across states would settle this.","The tunneling-oxide result suggests a broader design rule for nitride memories—interface engineering can tame variability without costing noise—but the paper's device count is small, so how far this generalizes is open."],"forward_implications":["If multilevel resistance is set by nitrogen-vacancy density, then programming algorithms can target a desired resistance by controlling the vacancy count in the filament instead of rebuilding the filament each cycle.","The finding that the filament/interface quality stays intact across high-resistance states implies that read noise and retention reliability should be comparable across levels, making multilevel storage more predictable.","Adding a tunneling oxide gives easier access to low-resistance states without increasing normalized noise, so it is a viable stack modification for lowering switching voltages in SiN RRAMs.","Because the insulating interface is not degraded during switching, endurance limits may be set by vacancy replenishment rather than by interface damage."],"supporting_citations":[{"why":"supplies Eq. (1), the low-frequency-noise model for a cylindrical metal-like conductive filament that the interpretation is built on.","marker":"[15]"},{"why":"provides the observed constant R/R fluctuation in high-resistance semiconductor-like filaments used to support the vacancy-density modulation picture.","marker":"[16]"},{"why":"offers prior low-frequency-noise evidence on how a SiO2 sublayer affects Si3N4 memristive devices, used to interpret the MIOS result.","marker":"[14]"},{"why":"establishes the multi-level programming and random-telegraph-noise measurement method employed here.","marker":"[12]"},{"why":"documents the effect of a SiO2 sublayer on silicon-nitride resistance switching and supports the tunneling-oxide role in preventing carrier leakage.","marker":"[10]"}],"fun_headline_variants":["Tunneling oxide steadies SiN ReRAM noise across resistance levels","Noise profile links SiN memory switching to nitrogen vacancy counts","Oxide layer improves SiN ReRAM variability without boosting noise","Flat 1/f noise in SiN RRAM ties high resistance to vacancy changes"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The argument depends on the low-frequency-noise formula for a cylindrical metal-like filament applying to silicon-nitride memories, and on the scaling of normalized noise with resistance uniquely identifying the switching mechanism.","fun_headline_variants_meta":{"raw":{"variants":["Tunneling oxide steadies SiN ReRAM noise across resistance levels","Noise profile links SiN memory switching to nitrogen vacancy counts","Oxide layer improves SiN ReRAM variability without boosting noise","Flat 1/f noise in SiN RRAM ties high resistance to vacancy changes"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00026,"raw_usage":{"total_tokens":1525,"prompt_tokens":818,"completion_tokens":707,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":434,"completion_tokens_details":{"reasoning_tokens":630}},"tokens_in":434,"tokens_out":707,"duration_ms":6725,"temperature":1.0,"reasoning_tokens":630,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-09T00:15:21.256874+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Fabricate SiN RRAMs with deliberately varied nitrogen-vacancy content (for example, through deposition stoichiometry or controlled SET currents), measure $S_i/I^2$ at fixed resistance, and check whether the normalized noise tracks the model's inverse-square dependence on carrier density; a flat $S_i/I^2$ across samples with changed vacancy density would falsify the attribution. Alternatively, direct cross-sectional imaging of filaments at different resistance states would show whether the filament radius or interface changes, which the current claim rules out.","supporting_citations":[{"cited_title":"Low-frequency noise in oxide-based (TiN/HfO x/Pt) resistive random access memory cells,","cited_arxiv_id":null,"evidence_quote":"supplies Eq. (1), the low-frequency-noise model for a cylindrical metal-like conductive filament that the interpretation is built on."},{"cited_title":"Statistical fluctuations in HfOx resistive-switching memory: Part I-Set/Reset variability,","cited_arxiv_id":null,"evidence_quote":"provides the observed constant R/R fluctuation in high-resistance semiconductor-like filaments used to support the vacancy-density modulation picture."},{"cited_title":"Effect of SiO2 sublayer on the retention characteristics of nanometer-sized Si3N4 memristive devices investigated by low-frequency noise spectroscopy,","cited_arxiv_id":null,"evidence_quote":"offers prior low-frequency-noise evidence on how a SiO2 sublayer affects Si3N4 memristive devices, used to interpret the MIOS result."},{"cited_title":"Multi-level resistance switching and random telegraph noise analysis of nitride based memristors,","cited_arxiv_id":null,"evidence_quote":"establishes the multi-level programming and random-telegraph-noise measurement method employed here."},{"cited_title":"Effect of SOI substrate on silicon nitride resistance switching using MIS structure,","cited_arxiv_id":null,"evidence_quote":"documents the effect of a SiO2 sublayer on silicon-nitride resistance switching and supports the tunneling-oxide role in preventing carrier leakage."}],"review_version":1}