{"id":"48b02494-c341-4a58-a058-74ea0865c33e","arxiv_id":"2412.00168","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"LCAO-based Ehrenfest dynamics within GPAW reproduces grid-based results at modest ion velocities with more than an order of magnitude speedup in large vacuum cells.","lead":"A new implementation of Ehrenfest molecular dynamics using localized atomic-orbital basis sets inside the projector augmented-wave method is presented and tested in the GPAW code. It offers a large speedup over real-space-grid dynamics for ion irradiation simulations, with accuracy that degrades for projectile velocities above about 4 angstroms per femtosecond.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Neglect of velocity-dependent forces is not validated at the claimed 4 Å/fs limit; LCAO energy drift reaches ~1 eV at 1.0 keV, so the 'satisfactory accuracy' claim is not yet established.","rationale":"The reader's weakest assumption is exactly the neglect of velocity-dependent terms in the nuclear forces, and I agree that this is the most load-bearing concern for the central claim. The paper's own energy-conservation tables show that LCAO-PAW-ED exhibits energy drifts up to ~1 eV at 1.0 keV (3.98 Å/fs), which is the upper end of the claimed validity range of 4 Å/fs. These drifts are two orders of magnitude larger than FD-PAW-ED and are directly attributable to the missing force terms. The paper does not separately verify the effect of this neglect on the reported kinetic energy loss, so the agreement with FD up to 1 keV is not yet demonstrated to be robust. The reader's CONDITIONAL verdict is appropriate: the manuscript should be corrected (including the index error in Eq. 7 and the inconsistent energy-conservation prose) and ideally should include a quantitative test of the force approximation. My concern does not move the verdict; it reinforces the need for the conditional changes. Therefore, the verdict should remain CONDITIONAL (UNCHANGED relative to the reader's assessment).","tokens_in":11923,"tokens_out":11115,"duration_ms":86103,"concrete_test":"Re-run the 1.0 keV H and H+ on graphene impacts with the nuclear forces augmented by the velocity-dependent terms (the P and G contributions, or equivalently by computing forces as -dE_tot/dR with E_tot including the connection terms), and compare the kinetic energy loss and Emax to the reported values. If the energy loss changes by more than ~0.5 eV or the energy drift remains above ~0.1 eV at 1.0 keV, the 'up to 4 Å/fs' accuracy claim is not supported. A complementary check is to recompute the NaCl vibration with the full forces and compare the oscillation period to BO-MD.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central accuracy claim (LCAO-PAW-ED reproduces FD-PAW-ED electronic stopping up to ~4 Å/fs) is not yet established because the nuclear forces omit the velocity-dependent terms arising from the moving LCAO basis and PAW projectors. Equations (3)–(7) show that the electronic propagation includes the P(t) and G terms, but Section I.A states forces are computed from the static energy derivative (Ref. 35) without these terms, adding only the qualification that this 'may only be valid within specific velocity ranges.' The energy-conservation data in Tables I–II quantify the consequence: for 1.0 keV H (3.98 Å/fs), LCAO Emax = 1.07 eV versus 0.0092 eV for FD; for 0.5 keV H+ (2.81 Å/fs), LCAO Emax = 1.03 eV. These are orders of magnitude larger than FD and grow rapidly with velocity. The prose claims LCAO violations are '0.011–0.372 eV,' which is inconsistent with these tabulated values. Because the energy loss (the reported observable) is extracted from trajectories generated with these incomplete forces, agreement with FD up to 1 keV could be fortuitous if the missing terms bias the trajectory. No test against a method that includes the velocity-dependent forces (e.g., the gauge-potential integrator of Ref. 28 or a full Lagrangian force derivation) is provided.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper reports an implementation of Ehrenfest molecular dynamics within the projector augmented-wave formalism of GPAW using a localized atomic-orbital (LCAO) basis, following the approximate Tomfohr–Sankey scheme. The authors benchmark LCAO-PAW-ED against the existing real-space grid FD-PAW-ED implementation for NaCl vibrations, CH2NH2+ isomerization, and H/H+ irradiation of a periodic graphene monolayer, and they document substantial savings in runtime and memory for large supercells. The central claim is that, for modest projectile velocities (up to about 4 Å/fs), LCAO-PAW-ED reproduces the electronic stopping and charge-transfer behavior of FD-PAW-ED at much lower cost, while becoming unreliable at higher velocities.","tokens_in":12271,"tokens_out":6377,"duration_ms":56274,"significance":"If the claim survives scrutiny, this is a useful and timely implementation: LCAO-PAW-ED could enable Ehrenfest simulations of ion irradiation with large vacuum cells, which are currently impractical with grid-based FD-PAW-ED. The work has notable strengths: it is benchmarked against an independent existing implementation and against Born-Oppenheimer dynamics, no parameters are fitted to the target results, the comparison is performed on well-defined physical observables (kinetic-energy loss, charge transfer, energy conservation), and the underlying data are openly deposited. The performance advantage is quantified concretely, including wall-clock time and memory. The main weaknesses are that the printed derivation of the velocity-dependent term contains an index inconsistency, the neglected velocity-dependent nuclear forces are not validated against a method that includes them, and a prose statement about energy-conservation errors is inconsistent with the paper's own tables.","major_comments":[{"comment":"The printed equation and the surrounding text have an index inconsistency in the velocity-dependent term. Starting from Eq. (6), the term involving ∂|Φμ⟩/∂t should lead to a coupling of Gνμ = ⟨Φν|Ŝ|∇Φμ⟩ with the velocity of the atom hosting basis function μ, not ν. The text currently says the matrix Gνμ is 'proportional to the velocity of the atom vν hosting basis function ν'. Please correct the index in Eq. (7) and in the sentence defining G, and clarify whether the implementation solves Eq. (7) explicitly or replaces the G·v term with the Tomfohr–Sankey Löwdin step of Eq. (8).","section":"§I.B, Eq. (7)"},{"comment":"The nuclear forces are computed without the velocity-dependent terms that arise from the moving LCAO basis and PAW projectors, and Section I.A states this 'may only be valid within specific velocity ranges'. The paper does not provide a test against any formulation that includes those terms, such as the gauge-potential integrator of Ref. 28 or a full Lagrangian force derivation. This is load-bearing because the reported energy-conservation violations grow rapidly with velocity: at 1.0 keV (3.98 Å/fs) the LCAO maximum violation is 1.07 eV for H and 2.18 eV for H+, whereas the FD values are 0.0092 and 0.0139 eV. The claim that LCAO-PAW-ED is 'satisfactory' up to 4 Å/fs is therefore not yet established; either add a validation against a method that includes the velocity-dependent forces, or narrow the claimed velocity range and the accuracy statement accordingly.","section":"§I.A and §II.C, Tables I–II"},{"comment":"The sentence 'even for lower velocities, energy conservation depends on velocity and is also worse (0.011–0.372 eV)' is inconsistent with the table entries. Tables I and II list LCAO Emax values of 0.0038 eV at 0.01 keV H, 0.38 eV at 0.5 keV H, 1.03 eV at 0.5 keV H+, and 2.18 eV at 1.0 keV H+. The prose range of 0.011–0.372 eV does not match any subset of the tabulated data. The authors should correct either the prose or the tables, and the corrected numbers should be used when assessing the accuracy of the method in the modest-velocity regime.","section":"§II.C, prose after Tables I–II"},{"comment":"The criterion for 'reproducing' FD-PAW-ED kinetic-energy loss should be stated more quantitatively. For example, at 0.5 keV H+ the LCAO and FD losses differ by about 2.4 eV (11.3 vs 8.9 eV), which is a substantial relative difference, and at 0.5 keV H they differ by 1.6 eV (11.3 vs 9.7 eV). If the intended claim is qualitative agreement, this should be said explicitly; if quantitative agreement is claimed, a tolerance or uncertainty estimate is needed.","section":"§II.C, Figure 7 and Tables I–II"}],"minor_comments":[{"comment":"The phrase 'To approximate the transformation of the basis in a square bracket' is unclear; please rewrite to refer explicitly to the action of the G·v term and its replacement by the Löwdin orthogonalization step.","section":"§I.B, text after Eq. (8)"},{"comment":"The note about the ×10−2 factor applies only to the final column, but the formatting is easy to misread; please restate the units in each column header and, if possible, use consistent scientific notation for the LCAO and FD columns.","section":"Tables I–II"},{"comment":"The labels 'EDLCAO' and 'EDFD' in Figures 5 and 6 are visually cramped; consider using distinct line styles or a legend with clearer spacing.","section":"Section II.C"},{"comment":"Reference 32 is missing the journal title and volume/page information; please complete the citation.","section":"References"},{"comment":"The blue shaded region is described as the van der Waals radius of carbon, but it is not explained how that radius is defined; a brief definition would help the reader interpret the plots.","section":"Section II.C, Figure 8 caption"}],"recommendation":"major_revision","confidential_remarks":"This is a promising implementation paper with honest benchmarking and useful performance data. The main issue is that the central accuracy claim is not fully supported because the velocity-dependent force terms are neglected and not independently validated, and because the printed derivation and the energy-conservation prose contain internal inconsistencies. These are fixable with additional tests or a more carefully bounded claim. I do not see a novelty or scope concern; the paper fits the journal. A revised version that resolves the Eq. (7) index error, reconciles the text with the tables, and either validates or explicitly restricts the 4 Å/fs claim would be suitable for publication."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper delivers what it says: a working LCAO-based Ehrenfest implementation in GPAW using the Tomfohr-Sankey integrator, benchmarked against the existing real-space-grid Ehrenfest code and BO-MD. The speedups are real (3x in small cells, 16x in a 29.5 Å box, less than half the memory), and the open data (Phaidra) makes it checkable. The authors are also honest about the main limitation: the TS integrator and the neglect of velocity-dependent force terms restrict the method to modest projectile velocities, and they say so in the abstract and conclusion. The comparison to FD-PAW-ED for H and H+ on graphene shows similar electronic stopping up to about 4 Å/fs, which is a practically useful result for ion-irradiation simulations of 2D materials.\n\nThe soft spots are real but mostly manuscript-level. Equation (7) has an index mismatch: the G matrix is defined with a derivative on the µ basis function but the velocity is labeled by ν. That needs fixing. More substantively, the prose claims energy-conservation violations of 0.011–0.372 eV for LCAO at low velocities, but Tables I and II list Emax values up to 1.07 eV (H) and 2.18 eV (H+) at 1.0 keV. Those are not fatal if the authors mean the final drift ∆E rather than the maximum violation, but as written the numbers don't match the claim.\n\nThe bigger methodological concern, which the paper itself flags, is that nuclear forces are computed from the static energy derivative without the velocity-dependent terms from the moving basis and PAW projectors. The paper validates against grid-based ED, which doesn't have those terms, and the agreement in stopping power up to 1 keV is encouraging. But because the missing terms directly affect energy conservation, and the violations grow quickly with velocity, the 'satisfactory accuracy' claim would be stronger with a test against a method that includes those terms (e.g., the gauge-potential integrator or a full Lagrangian force derivation), even for a single trajectory. As it stands, the claim is plausible but not airtight.\n\nWho is this for? Groups doing Ehrenfest simulations of ion irradiation in large cells with vacuum, especially with GPAW. It's a solid implementation paper, not a conceptual breakthrough. With the typos fixed and the energy-conservation statement clarified, it deserves publication. The missing-force-term caveat should either be verified or clearly stated as a quantitative limitation.\n\nRecommendation: send it to peer review. The referee should ask for the Eq. (7) fix, a corrected energy-conservation statement, and ideally a short test or explicit discussion of the velocity-dependent force neglect. None of this is load-bearing enough for rejection.","headline":"A useful, honest implementation paper with real speedups, but the central equation has a typo, the energy-conservation numbers don't match the prose, and the neglect of velocity-dependent forces deserves a sharper caveat.","tokens_in":12739,"tokens_out":3188,"would_cite":true,"duration_ms":27313,"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 paper implements Ehrenfest dynamics with localized atomic-orbital basis sets in the PAW method, showing that for projectile velocities up to about 4 Å/fs it reproduces grid-based results while running far faster.","keywords":["Ehrenfest dynamics","linear combination of atomic orbitals","projector augmented-wave method","Tomfohr–Sankey integrator","electronic stopping","ion irradiation of graphene","time-dependent density functional theory","Löwdin orthogonalization"],"falsifier":"Run the same H-on-graphene impact at 0.5 keV with an Ehrenfest implementation that explicitly includes the velocity-dependent connection terms from the moving basis and projectors, and compare the kinetic-energy loss, trajectory, and total-energy conservation with the Tomfohr–Sankey LCAO-PAW-ED results; if the stopping energies differ by more than about one eV or total-energy violations grow beyond the reported few-tenths-of-an-eV level, the neglect is the limiting assumption. A simpler proxy is to check whether LCAO-PAW-ED's energy loss converges to FD-PAW-ED as the grid spacing is tightened at 1 keV, which would indicate the discrepancy is basis-set rather than velocity-term driven.","tokens_in":11755,"feed_emoji":"⚛️","tokens_out":8457,"duration_ms":69594,"temperature":0.7,"pith_summary":"This paper implements Ehrenfest molecular dynamics (ED), in which electrons evolve by time-dependent Kohn–Sham equations while nuclei follow classical Newtonian trajectories, using a linear combination of atomic orbitals (LCAO) basis inside the projector augmented-wave (PAW) method of the GPAW code. The electron propagation follows Tomfohr and Sankey's approximate scheme, which applies a Löwdin orthogonalization to the expansion coefficients at each nuclear step to keep the wavefunction continuous as the basis moves. The central claim is that for modest atomic velocities—up to roughly 4 Å/fs for hydrogen projectiles on graphene—this LCAO-PAW-ED method reproduces the kinetic-energy loss and charge-transfer behavior of the established finite-difference grid implementation, at a fraction of the cost. In a 29.52 Å cubic cell the authors measure a 16-fold speedup and a memory footprint below 50% of the grid method, making large-vacuum ion-irradiation simulations practical. The paper also delineates the method's limit: at projectile energies of 5 keV and above the approximate integrator degrades, so the usefulness is confined to slower impacts.","feed_headline":"Localized-orbital Ehrenfest runs 16x faster on big cells","feed_subtitle":"Reproduces grid-based stopping power up to ~4 Å/fs while using less than half the memory.","key_machinery":"The load-bearing mechanism is the Tomfohr–Sankey time-propagation scheme with a Löwdin orthogonalization step. When atoms move, the LCAO expansion coefficients are mapped by $c_n(t_0+\\Delta t)=S^{-1/2}(t_0+\\Delta t)S^{1/2}(t_0)\\tilde{c}_n(t_0+\\Delta t)$, where $S$ is the PAW overlap matrix; this approximately preserves wavefunction continuity without computing the velocity-dependent connection terms $\\tilde{P}(t)$ and $G_{\\nu\\mu}$ that would appear in a full gauge-potential treatment. The electronic coefficients are otherwise propagated by the semi-implicit Crank–Nicolson scheme, and nuclear forces are taken as the negative derivative of the electronic energy with respect to nuclear positions, deliberately omitting velocity-dependent corrections. This combination is what makes the method tractable within PAW—where the projectors move with the nuclei—and sets the velocity ceiling where the approximation breaks down.","core_discovery":"The paper's discovery is that the Tomfohr–Sankey integrator, already known in pseudopotential LCAO codes, can be adapted to the projector augmented-wave formalism and yields an Ehrenfest dynamics method that is quantitatively reliable in the low-velocity regime. Expanding the PAW time-dependent Kohn–Sham equation in localized orbitals introduces two velocity-dependent connection terms—one from the moving PAW projectors, $\\tilde{P}(t)$, and one from the moving basis functions, $G_{\\nu\\mu}$—and the adaptation sidesteps them by applying the Löwdin step $c_n(t_0+\\Delta t)=S^{-1/2}(t_0+\\Delta t)S^{1/2}(t_0)\\tilde{c}_n(t_0+\\Delta t)$ after each nuclear displacement. The authors validate the implementation against the pre-existing finite-difference PAW-Ehrenfest method on NaCl bond oscillations, CH$_2$NH$_2^+$ twisting through a conical intersection, and H/H$^+$ impacts on graphene. For projectile velocities up to about 4 Å/fs the two methods agree on electronic-stopping energy loss and on the projectile's charge state, while for 5–10 keV projectiles the LCAO results deviate strongly and total-energy conservation worsens.","pith_inferences":["The measured speedup should grow with the fraction of vacuum in the cell, so the method is likely to make fully ab initio studies of slow highly charged ion impacts on suspended monolayers feasible for the first time—this is our inference, as the paper only hints at it.","A natural extension, not explored in the paper, would be to add back the velocity-dependent force terms (the connection terms) and test how far the reliable velocity range extends; the current results bracket where those terms start to matter.","Because the forces omit velocity-dependent terms by construction, energy conservation is not guaranteed; users of LCAO-PAW-ED should monitor total-energy drift as a standard diagnostic, and the 0.1–1 keV window would benefit from a dedicated benchmark against a method that includes the missing terms."],"forward_implications":["LCAO-PAW-ED can replace the grid-based method for ion-irradiation simulations with projectile speeds up to roughly 4 Å/fs, cutting wall-clock time by more than an order of magnitude in large supercells.","Cells with much more vacuum become tractable, which matters for simulating collisions of ions with suspended monolayer materials where the grid method's cost scales badly with empty space.","The weak dependence of LCAO-PAW-ED on the integration timestep in molecular tests means additional savings can come from using longer timesteps without losing accuracy.","Simulations that track projectile charge state, such as H$^+$ neutralization before impact on graphene, can be run in the same framework as neutral projectiles, since the method captures charge oscillations below the velocity limit.","For impacts at 5 keV and above, LCAO-PAW-ED should not be trusted; the finite-difference implementation or a method with full connection terms remains necessary."],"supporting_citations":[{"why":"Tomfohr and Sankey's time-dependent conduction scheme; supplies the Löwdin-orthonormalization step used for electron propagation.","marker":"29"},{"why":"Ojanperä et al.'s implementation of Ehrenfest dynamics with finite-difference PAW; the reference method that this paper extends and benchmarks against.","marker":"32"},{"why":"Larsen et al.'s localized atomic-orbital basis in PAW; provides the LCAO-PAW framework and the force convention adopted here.","marker":"35"},{"why":"Qian et al.'s derivation of the $\\tilde{P}$ operator for time-dependent PAW; defines the projector-motion term that the TS approximation avoids.","marker":"36"},{"why":"Ojanperä's thesis; supplies the general PAW time-dependent Kohn–Sham derivation used for the equations of motion.","marker":"37"},{"why":"Halliday and Artacho's comparison of gauge-potential and Tomfohr–Sankey integrators; motivates the TS choice and documents its accuracy limits.","marker":"31"},{"why":"GPAW code description; the implementation platform for both FD-PAW-ED and the new LCAO-PAW-ED.","marker":"25"}],"fun_headline_variants":["LCAO Ehrenfest: big-cell speed, low-velocity accuracy","Atomic-orbital Ehrenfest: fast, validated to 4Å/fs","Localized-orbital Ehrenfest: cheap for large cells, limited to slow ions","PAW-LCAO Ehrenfest: grid reliability at a fraction of the cost","Ehrenfest with atomic orbitals: faster, but only for modest speeds"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The central load-bearing premise is that the velocity-dependent terms in the forces—arising from the moving LCAO basis and moving PAW projectors—can be neglected for the modest velocities studied; the paper states this may only hold within specific velocity ranges and does not verify it against a method that includes those terms.","fun_headline_variants_meta":{"raw":{"variants":["LCAO Ehrenfest: big-cell speed, low-velocity accuracy","Atomic-orbital Ehrenfest: fast, validated to 4Å/fs","Localized-orbital Ehrenfest: cheap for large cells, limited to slow ions","PAW-LCAO Ehrenfest: grid reliability at a fraction of the cost","Ehrenfest with atomic orbitals: faster, but only for modest speeds"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001222,"raw_usage":{"total_tokens":5025,"prompt_tokens":944,"completion_tokens":4081,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":560,"completion_tokens_details":{"reasoning_tokens":3988}},"tokens_in":560,"tokens_out":4081,"duration_ms":24895,"temperature":1.0,"reasoning_tokens":3988,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T05:48:35.230623+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same H-on-graphene impact at 0.5 keV with an Ehrenfest implementation that explicitly includes the velocity-dependent connection terms from the moving basis and projectors, and compare the kinetic-energy loss, trajectory, and total-energy conservation with the Tomfohr–Sankey LCAO-PAW-ED results; if the stopping energies differ by more than about one eV or total-energy violations grow beyond the reported few-tenths-of-an-eV level, the neglect is the limiting assumption. A simpler proxy is to check whether LCAO-PAW-ED's energy loss converges to FD-PAW-ED as the grid spacing is tightened at 1 keV, which would indicate the discrepancy is basis-set rather than velocity-term driven.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Ojanperä et al.'s implementation of Ehrenfest dynamics with finite-difference PAW; the reference method that this paper extends and benchmarks against."},{"cited_title":"Ojanperä , author V","cited_arxiv_id":null,"evidence_quote":"Larsen et al.'s localized atomic-orbital basis in PAW; provides the LCAO-PAW framework and the force convention adopted here."},{"cited_title":"Ojanperä , author A","cited_arxiv_id":null,"evidence_quote":"Ojanperä's thesis; supplies the general PAW time-dependent Kohn–Sham derivation used for the equations of motion."},{"cited_title":"Artacho \\ and\\ author D","cited_arxiv_id":null,"evidence_quote":"Halliday and Artacho's comparison of gauge-potential and Tomfohr–Sankey integrators; motivates the TS choice and documents its accuracy limits."},{"cited_title":"Castro , author H","cited_arxiv_id":null,"evidence_quote":"GPAW code description; the implementation platform for both FD-PAW-ED and the new LCAO-PAW-ED."}],"review_version":1}