{"id":"f1709458-b96e-400e-849a-39d45f433a85","arxiv_id":"2505.06809","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"Electroforming in HfOx/Ti RRAM is governed by a switch from vertical to lateral vacancy motion as the oxide becomes oxygen-poor, and by whether heating is uniform or local.","lead":"This paper combines atomistic simulations with device measurements to explain how oxygen-vacancy stoichiometry and temperature control the voltage needed to form conductive filaments in HfOx/Ti memory cells. The findings could help engineers lower forming voltages and tailor filament shape for more reliable resistive memory.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The vertical-to-lateral transition is shown only in a 2.5-nm cell at 20% vacancies, where edge effects dominate, so the central mechanism may be a finite-size artifact.","rationale":"The reader's weakest assumption (scale transfer) is the right general concern, and this is the precise point where that assumption carries the most weight: the central mechanism is inferred in the smallest domain and at a vacancy concentration beyond any measured device. The paper is otherwise internally consistent: the experimentally matched trends (area-dependent pristine resistance, area-independent LRS, VFORM vs area, sweep rate, and T_env) are reproduced without post-hoc fitting, and the d-KMC framework with NEB-derived barriers is a reasonable tool for the stated purpose. But none of those matched trends discriminates between vertical and lateral vacancy dynamics; only the 2.5-nm, 20% vacancy snapshots do. Since the authors' own finite-size caveat appears in the same section, applying it to their central mechanistic claim is a natural and testable condition. The condition should be added to the verdict; I would keep the overall CONDITIONAL judgement rather than reject, because the proposed check is feasible (the code supports larger domains) and could reasonably resolve the ambiguity.","tokens_in":18939,"tokens_out":6180,"duration_ms":65543,"concrete_test":"Repeat the stoichiometry sweep at [Vo] = 5%, 10%, and 20% in 5×5 and 10×10 nm domains with the same d-KMC solver and random-seed protocol, and quantify the directionality of vacancy motion, e.g., the ratio of summed vertical to lateral vacancy displacement components during forming and the fraction of final filament volume within one lattice constant of a lateral edge. If the vertical-to-lateral crossover disappears at 5% or shifts with domain size, the claim is an edge/scale artifact; if it persists at 5% and in larger cells, the mechanism is robust.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The manuscript's headline mechanism—stoichiometry-driven switch from vertical to lateral vacancy migration—rests on Fig. 3(d)-(e), which uses the smallest 2.5×2.5 nm cell and 20% oxygen vacancies (HfO1.6). In the very same figure discussion the authors note that at this size \"the surface area of lateral boundaries becomes significant compared to the volume of the switching material, leading to filaments preferably forming at edges/corners of the stack.\" Thus the lateral fields and percolating paths invoked to explain the transition are exactly what a few-lattice-spacing-wide domain would artificially produce. The experimental \"leaky\" devices are 170 nm–1.57 µm in size with roughly 5% vacancies (HfO1.9), a composition where the manuscript does not demonstrate lateral dominance; the 20% case is never measured. No quantitative anisotropy metric is reported—only snapshots colored by depth and temperature. Because the abstract generalizes this result to \"sub-stoichiometric oxides\" and to design rules for real stacks, the unvalidated transfer from a strong-edge-effect simulation domain at an unmeasured composition is the most load-bearing weakness.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper combines atomistic driven kinetic Monte Carlo (d-KMC) simulations with electrical measurements of TiN/a-HfOx/Ti/TiN RRAM devices to explain the mechanisms behind reduced electroforming voltage due to oxide stoichiometry and thermal engineering. The authors reproduce qualitative experimental trends in forming voltage as functions of device area, initial oxygen-vacancy concentration, RVS sweep rate, and ambient temperature, and they visualize simulated filament morphologies under these conditions. The central mechanistic claims are (i) a stoichiometry-driven transition from predominantly vertical to predominantly lateral oxygen-vacancy migration during filament formation in sub-stoichiometric oxides, and (ii) a distinction between globally uniform heating (higher T_env) and locally concentrated Joule heating, which are argued to produce different filament geometries and to affect the subsequent HRS/LRS dynamic range. The paper proposes design guidelines for lower-voltage forming protocols and for tuning the resulting switching behavior.","tokens_in":19141,"tokens_out":3496,"duration_ms":37668,"significance":"If the central mechanistic claims are correct, this work would provide an atomistically grounded explanation of how compositional and thermal knobs influence electroforming in a technologically relevant HfOx/Ti stack, connecting nanoscale ion kinetics to device-level forming voltages. The strength of the paper is its combination of a device-scale atomistic solver with a consistent experimental dataset, reporting trends across four independent variables and statistical spreads for both experiment and simulation. The use of NEB-computed activation energies and independent quantum-transport calibration for the current solver is a positive feature, as is the explicit acknowledgement of finite-size effects and the discussion of edge-dominated behavior in the smallest cells. However, the signature mechanistic claim of a vertical-to-lateral transition rests on limited simulation evidence, and the scale gap between simulated and fabricated devices is substantial; these issues currently prevent the broader conclusions from being fully supported.","major_comments":[{"comment":"The claim of a vertical-to-lateral transition in vacancy migration is evidenced only by visual snapshots from a 2.5×2.5 nm cell at 20% oxygen vacancies (HfO1.6), a composition that was not measured experimentally. In the same section the authors note that at this size 'the surface area of lateral boundaries becomes significant compared to the volume of the switching material, leading to filaments preferably forming at edges/corners of the stack.' Because the lateral fields invoked to explain the transition are exactly what such a small domain with strong edge effects would produce, the simulation evidence does not yet establish the transition as a general property of sub-stoichiometric oxides. A quantitative metric of the directional anisotropy of vacancy motion (e.g., a flux ratio or migration-direction histogram) across several stoichiometries and in a larger domain (7.5 or 10 nm) is needed to rule out a finite-size artifact. Until such evidence is provided, the abstract's generalization to 'sub-stoichiometric oxides' overreaches the data.","section":"§Oxide stoichiometry, Fig. 3(d)-(e)"},{"comment":"The scale-transfer assumption is load-bearing for the paper's mechanistic conclusions. The fabricated devices are 170 nm to 1.57 µm in lateral extent, while the simulations are limited to 2.5–10 nm. The authors state that trends measured at different sizes 'can be applied to the ultrascaled dimensions that are achievable via d-KMC simulations,' but the smallest simulated cell already deviates from the monotonic area trend of V_FORM due to edge effects (Fig. 3(c)). This non-monotonicity shows that the simulated regime is not simply the continuation of the experimental trend, and mechanisms inferred from the smallest cells may not transfer to the measured devices. The paper should either demonstrate convergence of the vertical/lateral behavior with increasing lateral size or clearly delimit the size range over which the mechanism is claimed to apply.","section":"§Correlating fabricated and simulated device domains"},{"comment":"The distinction between global and local heating is a central conclusion, but the magnitude of the Joule-heating effect depends on the dissipation fraction α, which is fixed at 0.10 without a reported sensitivity analysis. Since the simulated filament morphologies and the inferred distinction between uniform thermal activation and local Joule-heating-driven lateral diffusion are directly controlled by this parameter, the robustness of the thermal conclusions to the choice of α should be demonstrated (e.g., by varying α over a reasonable range and comparing filament widths or dispersion metrics). Without this, the 'local vs. global heating' narrative is not yet shown to be a parameter-independent finding.","section":"§Model for Joule heating, Eq. (4)"}],"minor_comments":[{"comment":"There is a typo in the text: 'coherant Quantum Transport simulations' should be 'coherent Quantum Transport simulations'.","section":"Methods, Current flow"},{"comment":"The caption reads 'I-V characteristic of the the same devices as in (b)'; the duplicated 'the' should be removed.","section":"Fig. 2(c) caption"},{"comment":"The caption states that V_FORM is extracted 'at the point at which a compliance current I_CC = 20 µA is reached,' but the main text also mentions a 10 kΩ external series resistor; it would be clearer to state whether V_FORM is read at the device or across the series resistor, since the latter would include an IR drop.","section":"Fig. 3(c) caption"},{"comment":"The top-view vacancy profiles are presented without a quantitative description of how the depth coordinate is projected. A brief explanation of the coloring scheme and projection would improve interpretability.","section":"Fig. 5(d)-(e)"}],"recommendation":"major_revision","confidential_remarks":"The manuscript presents an ambitious and largely useful combination of experiments and simulations, and the qualitative reproduction of several experimental trends is commendable. My main concern is that the headline mechanistic claim—the vertical-to-lateral transition—rests on a single simulation at an extreme composition and a very small domain, and the abstract generalizes it beyond the evidence. The authors should be encouraged to strengthen this point with additional simulations or to qualify the claim. I would also note that the paper does not contain an explicit data/code availability statement; given the reliance on a specialized solver, such a statement (or a link to the code) would increase reproducibility and impact."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a solid, useful paper, but the central mechanistic claim is shakier than the paper's tone suggests.\n\nWhat's genuinely new: the paper separates global ambient heating from local Joule heating effects on filament morphology, and it offers a concrete atomistic explanation for why O-poor oxides form at lower voltage. The d-KMC model reproduces four independent experimental trends—area, stoichiometry, sweep rate, temperature—without fitting the forming voltages themselves; the current-solver parameters are calibrated to quantum transport, and the activation energies come from NEB. That is real work and it shows. The experimental side is also well done: 36 devices, several areas, honest error bars.\n\nThe soft spot is the vertical-to-lateral transition. It is demonstrated in Fig. 3(e) for a 2.5×2.5 nm cell with 20% vacancies (HfO1.6), a composition the authors never fabricated. The experimental 'leaky' oxide is around 5% vacancies, and the abstract generalizes the claim to 'sub-stoichiometric oxides' generally. In that same figure discussion, the authors note that at 2.5 nm lateral boundaries are significant and filaments form preferentially at edges and corners. The lateral fields that drive the lateral vacancy migration are exactly what a domain a few lattice spacings wide would artificially produce. There is no quantitative anisotropy metric—only depth- and temperature-colored snapshots. So the mechanism may be a finite-size artifact. This is a load-bearing weakness for the paper's headline claim. It is not fatal to the paper's overall value, because the trends in forming voltage are reproduced without post-hoc fitting, and the thermal global-vs-local separation is supported by consistent behavior across sweep-rate and temperature simulations.\n\nWho should read this: anyone doing atomistic simulations of valence-change memory, and experimentalists thinking about forming protocols. It deserves peer review, but the referee should ask for a sensitivity test at larger domains with the same vacancy fraction, or a quantitative vertical/lateral metric, and a more explicit defense of the scale-transfer assumption from 10 nm to 170 nm+ devices. If that comes, this could be a reference for the field.","headline":"Useful and honest combined experiment/d-KMC study, but the headline vertical-to-lateral mechanism leans on a 2.5 nm cell at 20% vacancies—exactly the regime where edge effects dominate.","tokens_in":19718,"tokens_out":2922,"would_cite":true,"duration_ms":29011,"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":"This paper claims that electroforming in HfOx/Ti RRAM is governed by a stoichiometry-dependent switch from vertical to lateral oxygen-vacancy migration and by whether heating is global or local, providing design rules for lower forming…","keywords":["RRAM","electroforming","hafnia","oxygen vacancies","kinetic Monte Carlo","resistive switching","Joule heating","filament morphology"],"falsifier":"Form HfOx/Ti devices with oxygen-vacancy concentrations from about 1% to 20% and image the filaments after forming (for example by cross-sectional transmission electron microscopy): if the filaments remain predominantly vertical at 20% vacancies, the claimed vertical-to-lateral transition is wrong. Alternatively, carry out electroforming at elevated ambient temperature and measure the HRS/LRS ratio: if the dynamic range does not increase, the thermal-morphology mechanism is not supported.","tokens_in":18735,"feed_emoji":"⚡","tokens_out":7385,"duration_ms":72654,"temperature":0.7,"pith_summary":"This paper attempts to explain, at the level of individual oxygen vacancies, why electroforming voltages in HfOx/Ti resistive memory are lower when the oxide is oxygen-poor or when heat is added, and why those same knobs change the shape of the conductive filament. The proposed mechanism is a transition in the dominant direction of ion motion: in near-stoichiometric HfO2, vacancies drift vertically along the applied field, while in sub-stoichiometric oxides they move laterally along percolating pathways, which lowers the forming voltage but also degrades the resistance contrast. The paper also separates two kinds of heating—uniform ambient heating, which produces a narrow directional filament and increases dynamic range, and localized Joule heating, which disperses the filament into multiple weak pathways. A sympathetic reader would care because these mechanisms turn two empirical tricks (stoichiometry tuning and thermal engineering) into design rules: form at elevated temperature to cut voltage without sacrificing memory contrast.","feed_headline":"Forming voltage in HfOx memory hinges on ion drift direction","feed_subtitle":"Sub-stoichiometric oxide shifts vacancy motion sideways; uniform heat narrows filaments, Joule heat disperses them.","key_machinery":"The argument is carried by a driven kinetic Monte Carlo (d-KMC) solver that represents the entire device as an atomistic graph of lattice and interstitial sites, with each oxygen-vacancy generation, diffusion, or recombination event assigned an activation energy from nudged-elastic-band calculations, and with a trap-assisted tunneling current solver and a Joule-heating term (Fourier heat equation with power dissipation from computed current flow) coupled back into the event rates. After every structural change the electrostatic potential, heat distribution, and current are recomputed, so filament growth emerges from the stochastic evolution of point defects in a structurally inhomogeneous amorphous oxide. This machinery lets the paper simulate domains from 2.5 to 10 nm across, capture finite-size and edge effects, and compare simulated forming sweeps directly against measured trends on devices 170 nm to 1.57 µm across.","core_discovery":"The central claim is that the electroforming kinetics of TiN/a-HfOx/Ti/TiN RRAM stacks are controlled by two independently tunable factors: the initial oxygen-vacancy concentration and the spatial character of heating. In more stoichiometric oxides, forming proceeds by vertical drift of oxygen vacancies toward the counter electrode, with the Ti layer acting as an oxygen reservoir; in sub-stoichiometric oxides, pre-existing vacancies form percolating pathways over short distances, generating lateral fields that redirect migration laterally, so forming becomes limited by migration rather than generation of new vacancies and at roughly 20% vacancy concentration the device is forming-free. Regarding temperature, uniform heating from a raised ambient temperature accelerates generation and migration everywhere, lowering VFORM by about 1 V per 125 °C in the measured devices and leaving a narrow, directional filament with higher HRS/LRS ratio; in contrast, Joule heating created internally at high voltages is localized, drives thermally-induced lateral diffusion of vacancies, and forms multiple dispersed filaments that shrink the dynamic range. Because the compliance current fixes the low-resistance-state current in all cases, the morphology differences manifest in the high-resistance state and hence in the achievable resistance contrast.","pith_inferences":["For analog or neuromorphic applications where gradual conductance tuning is wanted, deliberately promoting dispersed filament growth (fast sweeps, high Joule heating) may be an asset rather than a drawback; the paper notes the analog-operating relevance of dispersed morphologies but does not make this design recommendation explicit.","The preference of filaments for stack edges at 2.5 nm device sizes suggests that at single-nanometer dimensions, electrode geometry and corner sharpness could be engineered independently of stoichiometry as a forming-voltage lever—a testable extension.","The scale-transfer assumption could be checked by fabricating devices with intermediate areas (tens of nanometers) and measuring VFORM vs area; a non-monotonic trend with a small-area upturn would validate the simulation predictions, while a continued rise would invalidate them.","The Joule-heating dissipation fraction (α = 0.10) is a model input; sweeping it would show how sensitive the sweep-rate-induced morphology change is, a sensitivity analysis the paper does not provide."],"forward_implications":["Oxygen-poor HfOx lowers the electroforming voltage because pre-existing vacancies support lateral migration and percolation, and at roughly 20% vacancy concentration devices become forming-free, with SET/RESET happening within the forming voltage range.","Raising the ambient temperature during forming cuts VFORM (about 1 V per 125 °C in measurements) and produces a narrower, more directional filament, which increases the HRS/LRS dynamic range.","Fast voltage ramps raise VFORM and, through localized Joule heating, create multiple dispersed filaments, while slow ramps form a single wider filament; the LRS current stays fixed by the compliance in both cases.","Global (ambient) heating and local (Joule) heating have opposite effects on filament morphology; devices needing high resistance contrast should be formed at high temperature with slow ramps, not fast high-voltage sweeps.","The d-KMC approach with trap-assisted current and Joule heating can be transferred to other valence-change binary-oxide systems to derive the same type of forming guidelines."],"supporting_citations":[{"why":"Supplies the driven kinetic Monte Carlo framework that models field-induced resistive switching in valence-change memory cells, forming the methodological base of this work.","marker":"[18]"},{"why":"Provides atomistic molecular dynamics of the full HfO2 RRAM operation cycle, used as a comparison for switching mechanisms and as an example of lateral filament dynamics.","marker":"[17]"},{"why":"Reports record-small 10x10 nm Hf/HfO2 crossbar RRAM, supplying the ultrascaled dimension baseline and supporting the finite-size scaling discussion.","marker":"[30]"},{"why":"Documents how forming voltage scales with device area, the trend this paper reproduces and extends across measured and simulated sizes.","marker":"[6]"},{"why":"Offers temporally resolved imaging of dielectric breakdown in valence-change memory, backing the oxygen-extraction and primarily vertical filament growth picture.","marker":"[36]"},{"why":"Calculates the reduced energy for oxygen-vacancy generation at the Hf/HfO2 interface, an input used to set the KMC event rates.","marker":"[33]"},{"why":"Observed radial oxygen migration in HfOx memristors, providing experimental support for the lateral-motion component in sub-stoichiometric oxides.","marker":"[52]"},{"why":"Describes a method to reduce forming voltage without degrading performance in HfO2 1T1R devices, the experimental counterpart to the high-temperature forming study.","marker":"[57]"},{"why":"Three-dimensional imaging of hourglass-shaped conductive filaments in nanoscale RRAM, the morphological reference for simulated filament shapes.","marker":"[22]"},{"why":"Microscopy study showing multiple conductive filaments in HfO2 devices, supporting the multi-filament growth seen in larger simulation domains.","marker":"[35]"}],"fun_headline_variants":["Lateral ion drift lowers forming voltage in HfOx memory","Uniform heat trims filaments; Joule heat spreads them in RRAM","Vacancy concentration dictates electroforming path in HfOx/Ti","Thermal engineering tailors filament morphology in oxide RRAM"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The paper's mechanistic conclusions assume that trends measured on fabricated devices 170 nm to 1.57 µm across transfer directly to the 2.5–10 nm simulation domains; if smaller fabricated devices do not follow the same size trends (for example because corner effects or defect landscapes differ), the inferred vertical-to-lateral transition and filament morphologies would not describe real devices.","fun_headline_variants_meta":{"raw":{"variants":["Lateral ion drift lowers forming voltage in HfOx memory","Uniform heat trims filaments; Joule heat spreads them in RRAM","Vacancy concentration dictates electroforming path in HfOx/Ti","Thermal engineering tailors filament morphology in oxide RRAM"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000242,"raw_usage":{"total_tokens":1595,"prompt_tokens":1084,"completion_tokens":511,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":700,"completion_tokens_details":{"reasoning_tokens":439}},"tokens_in":700,"tokens_out":511,"duration_ms":5485,"temperature":1.0,"reasoning_tokens":439,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T22:32:09.573925+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Form HfOx/Ti devices with oxygen-vacancy concentrations from about 1% to 20% and image the filaments after forming (for example by cross-sectional transmission electron microscopy): if the filaments remain predominantly vertical at 20% vacancies, the claimed vertical-to-lateral transition is wrong. Alternatively, carry out electroforming at elevated ambient temperature and measure the HRS/LRS ratio: if the dynamic range does not increase, the thermal-morphology mechanism is not supported.","supporting_citations":[{"cited_title":"An Atomistic Model of Field-Induced Resistive Switching in Valence Change Memory","cited_arxiv_id":null,"evidence_quote":"Supplies the driven kinetic Monte Carlo framework that models field-induced resistive switching in valence-change memory cells, forming the methodological base of this work."},{"cited_title":"L.; Islam, M","cited_arxiv_id":null,"evidence_quote":"Provides atomistic molecular dynamics of the full HfO2 RRAM operation cycle, used as a comparison for switching mechanisms and as an example of lateral filament dynamics."},{"cited_title":"10x10nm Hf/HfO _2 crossbar resistive RAM with excellent performance, reliability and low-energy operation","cited_arxiv_id":null,"evidence_quote":"Reports record-small 10x10 nm Hf/HfO2 crossbar RRAM, supplying the ultrascaled dimension baseline and supporting the finite-size scaling discussion."},{"cited_title":"Forming voltage scaling of resistive switching memories","cited_arxiv_id":null,"evidence_quote":"Documents how forming voltage scales with device area, the trend this paper reproduces and extends across measured and simulated sizes."},{"cited_title":"A.; Lodico, J","cited_arxiv_id":null,"evidence_quote":"Offers temporally resolved imaging of dielectric breakdown in valence-change memory, backing the oxygen-extraction and primarily vertical filament growth picture."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Calculates the reduced energy for oxygen-vacancy generation at the Hf/HfO2 interface, an input used to set the KMC event rates."},{"cited_title":"P.; Vine, D.; Kilcoyne, A","cited_arxiv_id":null,"evidence_quote":"Observed radial oxygen migration in HfOx memristors, providing experimental support for the lateral-motion component in sub-stoichiometric oxides."},{"cited_title":"M.; Liu, H.-W.; Chen, P.-H.; Chang, T.-C.; Tsai, T.-M.; Chu, T.-J.; Pan, C.-H.; Wu, C.-H.; Yang, C.-C","cited_arxiv_id":null,"evidence_quote":"Describes a method to reduce forming voltage without degrading performance in HfO2 1T1R devices, the experimental counterpart to the high-temperature forming study."},{"cited_title":"Three-Dimensional Observation of the Conductive Filament in Nanoscaled Resistive Memory Devices","cited_arxiv_id":null,"evidence_quote":"Three-dimensional imaging of hourglass-shaped conductive filaments in nanoscale RRAM, the morphological reference for simulated filament shapes."},{"cited_title":"C.; Kirsch, P","cited_arxiv_id":null,"evidence_quote":"Microscopy study showing multiple conductive filaments in HfO2 devices, supporting the multi-filament growth seen in larger simulation domains."}],"review_version":1}