{"id":"e9b896ca-1a9f-4513-b674-d80a26fb309d","arxiv_id":"2510.03479","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"Automated active learning produces transferable MTPs for EC/EMC/LiPF6 electrolytes, while explicit electrostatics (QRd) matches accuracy with fewer parameters only where its MD remains stable.","lead":"This paper shows that an active-learning loop can automatically build training sets for machine-learned potentials of battery electrolytes, giving densities within about 6% and ionic conductivities within 11% of experiment. It also tests a charge-aware potential that saves parameters in pure solvents but turns unstable for the LiPF6 solution.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 6% MTP-QRd electrolyte-conductivity claim rests on 2.8 ns of unstable MD; Green-Kubo convergence is asserted but not demonstrated.","rationale":"The reader's weakest assumption exactly matches the concern I identified: the reliability of MTP20-QRd conductivity depends on Green-Kubo convergence despite acknowledged MD instability. The paper itself contains explicit limitation statements (Sec. 2.3.3, Sec. 4.6, Table 1) that the trajectory statistics are insufficient, which must be weighed as in-scope evidence. The MTP-QRd claim is the central differentiator of the paper (the abstract's 6% figure), and it is fragile. The AL pipeline for MTP is well-supported: stable densities, 25 ns MD, RDF agreement; that part of the paper is strong. Therefore the verdict remains CONDITIONAL: the conductivity comparison needs verification or re-analysis before the 6% claim can be accepted. I would not move to REJECT because the underlying AL/electrostatics claim for EC/EMC mixtures (fewer parameters, comparable accuracy) is supported by RMSE analysis in Sec. 2.2, and the electrolyte claim might survive if a re-analysis of the short trajectories still gives agreement. The concrete test would settle whether the 6% is an artifact or a genuine prediction.","tokens_in":20217,"tokens_out":1545,"duration_ms":10925,"concrete_test":"Compute the ionic conductivity of the 1M LiPF6 in 3EC:7EMC system at 300 K using the stable MTP20, but truncate the Green-Kubo integral at 100 ps, average over only the first 200 ps of 17 independent trajectories, and use only 100 ps equilibration. If the resulting conductivity differs from the converged MTP20 value by more than the reported ~5–11% deviation band, the MTP-QRd 6% value cannot be distinguished from a sampling artifact. Additionally, run MTP-QRd for at least 10 ns (or until stable) if possible and recompute the conductivity; if the value shifts outside the 6% band, the claim fails.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim that explicit electrostatics (MTP-QRd) reproduces ionic conductivity within 6% of experiment relies on Green-Kubo integrals computed from MD that the paper itself says is unreliable. In Sec. 2.3.3 and Sec. 4.6, MTP-QRd MD trajectories failed beyond a few hundred ps, requiring 100 ps equilibration (vs 5 ns for MTP) and only 17/14/6 completed runs out of 50 at 280/300/320 K, giving total sampling of 3.4/2.8/1.2 ns. The correlation time was reduced to 100 ps and averaging was over the last 20 ps, whereas for MTP the correlation time was 500 ps and averaging over 300 ps. The paper explicitly states this 'undermines the reliability of the ionic conductivity values obtained with MTP20-QRd.' The Green-Kubo integral's convergence requires the current autocorrelation function to decay to zero; with such short, unstable trajectories and acknowledged unphysical charges (Sec. 2.3.2, Fig. S6), the 6% agreement with experiment may be an artifact of insufficient sampling, incomplete equilibration, or biased dynamics. If this is unsupported, the headline advantage of QRd over MTP for the electrolyte is unsubstantiated.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript presents a D-optimality/MaxVol active-learning workflow for training system-specific moment tensor potentials (MTPs) for EC/EMC binary solvents and 1M LiPF6 in EC/EMC, and compares the resulting short-range MTPs with MTP-QRd, an MTP augmented by a fixed-charge redistribution (QRd) term. The MTPs are reported to give stable densities (within ~6%) for pure solvents and mixtures, accurate Li–O RDFs, and ionic conductivities within 11% of experiment across temperatures and solvent compositions. The MTP-QRd model is reported to reach comparable or better accuracy than MTP at lower MTP level for salt-free mixtures and to reproduce the 3:7 EC:EMC LiPF6 conductivity within 6% of experiment, but its MD is acknowledged to be unstable, with only 2.8 ns of usable trajectory at 300 K and a much shorter equilibration/correlation protocol.","tokens_in":20560,"tokens_out":6124,"duration_ms":52709,"significance":"If upheld, the paper would establish a practical, hands-off AL pipeline for building system-specific MLIPs for liquid electrolytes, with strong validation of the short-range MTP: 25 ns of stable MD, ~6% density error, 11% conductivity error, and Li–O RDFs that improve on ReaxFF relative to AIMD. The D-optimality/MaxVol procedure is well-defined, the MLIP-2 and MLIP-4 codes are public, and the training sets are promised for release; these are concrete strengths. The MTP-QRd parameter-efficiency claim for EC/EMC mixtures is plausible but incompletely demonstrated because some QRd models could not cover all compositions. The headline 6% explicit-electrostatics conductivity result is the main liability: the underlying trajectories are short, unstable, and the paper itself states that this undermines the result's reliability.","major_comments":[{"comment":"The central abstract claim of 6% ionic conductivity for MTP20-QRd rests on 17/14/6 completed runs out of 50 at 280/300/320 K, total sampling of 3.4/2.8/1.2 ns, 100 ps equilibration (versus 5 ns for MTP), a 100 ps correlation time, and averaging over the last 20 ps of the Green–Kubo integral. The manuscript itself states that this 'undermines the reliability of the ionic conductivity values obtained with MTP20-QRd' (Sec. 2.3.3). The subsequent assertion that the value is 'unlikely to be severely affected' is not supported by any demonstrated decay of the current autocorrelation function to zero within 100 ps, nor by convergence checks with respect to equilibration length or starting configuration. The 6% figure and the related statement in Sec. 3 that MTP-QRd 'outperformed MTP by 5%' should be removed from the abstract/discussion or replaced by a clearly labeled preliminary estimate accom","section":"2.3.3, 4.6, Table 1"},{"comment":"The parameter-efficiency advantage claimed for MTP-QRd is broader than the evidence. For EC/EMC mixtures the paper notes that 'some MTP-QRd models could not model all mixture compositions' (Sec. 2.2), and for LiPF6 solution MTP20-QRd was stable only in the single 3EC:7EMC composition corresponding to its training set (Sec. 2.3.3). The abstract's wording that the extended MTP 'achieves accuracy comparable to MTP ... with fewer parameters' therefore needs an explicit scope restriction to the compositions where QRd MD is actually stable. As written, the claim suggests a general advantage that the paper's own stability data contradict.","section":"2.2, 2.3.3, 3"},{"comment":"The front-matter abstract states that charge redistribution was assessed 'using either fixed or environment-dependent charges' and that 'environment-dependent charges further improve accuracy and the stability of simulations.' The body implements only the fixed-charge QRd scheme (Sec. 4.2, Eq. 8) and reports that QRd MD is unstable for LiPF6 in pure EC (Sec. 2.3.2); no environment-dependent charge model is trained or tested. This is an internal inconsistency that attributes to the paper a result it does not contain. The abstract must be corrected to describe only the fixed-charge QRd model that was actually used.","section":"Abstract (front matter) vs 4.2, 2.3.2"}],"minor_comments":[{"comment":"The caption contains two panels labelled (d): 'd) Coordination number (CN) distributions in ionic pairs present in training set' and then 'd) Force error magnitudes...' The second should be relettered (e) or (f), and subsequent panel labels adjusted.","section":"Figure 5 caption"},{"comment":"The text 'where xx% of the EC and xx% of the EMC configurations are inherited from the initial training set' contains unprocessed placeholders. Please replace with actual percentages.","section":"Supplementary Information, S1"},{"comment":"Several typographical errors need correction: 'unpysical' (Sec. 2.3.2), 'Maxvell' (Sec. 4.6), 'primitivity' (Sec. 2.3.3), 'respressed' (Sec. 4.2), 'neigboring' (Sec. 4.1), 'compering' (Sec. 9), and 'on the over hand' (Sec. 2.3.3).","section":"Throughout"},{"comment":"The caption says RMSEs are 'compared with PBE-D3 calculations,' which is clear, but the text could state explicitly that reference energies/forces are DFT values computed on MTP20-generated validation configurations, so that readers do not mistake the validation set for AIMD-sampled configurations.","section":"Figure 3"}],"recommendation":"major_revision","confidential_remarks":"The paper's MTP/AL half is solid and likely publishable after revision. The decision hinges on whether the authors are willing to demote the MTP-QRd conductivity result to a preliminary observation or provide the missing convergence evidence. The current abstract overstates a result that the authors themselves describe as unreliable, and the front-matter abstract additionally claims an environment-dependent charge model that was never implemented. The reliance on unpublished reference [38] for the QRd model is a secondary concern; the referee could not independently verify the QRd charge parametrization."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"This is two papers in one. The first is a genuinely useful demonstration that D-optimality active learning, without manual density inflation/deflation, builds stable, composition-transferable MTPs for EC, EMC, their mixtures, and LiPF6. That half holds up: densities within ~6%, MTP20 conductivities within 11% of experiment, 25 ns stable MD, and Li–O RDFs that beat ReaxFF against AIMD. The head-to-head RMSE analysis against DFT is also credible and shows QRd reaching a given accuracy at a lower MTP level, which is a real parameter-efficiency argument for the salt-free solvent mixtures.\n\nWhere it gets shaky is the QRd leg for the LiPF6 solution. The headline 6% conductivity deviation rests on 2.8 ns of usable trajectory from 17/14/6 completed runs out of 50, with 100 ps equilibration instead of 5 ns and a 100 ps correlation time instead of 500 ps. The paper itself says this 'undermines the reliability of the ionic conductivity values obtained with MTP20-QRd.' That is not a minor caveat; it is the basis for the abstract's central quantitative claim. The SI figure showing a plateau in sigma(t) for QRd is not convincing because the dynamics are unstable and the charges are acknowledged to be unphysical, so the autocorrelation function may not have decayed for the right reason. The structural differences in ion-pair populations between MTP and QRd—SSIPs only versus CIPs and AGGs present—also suggest the two models are not sampling the same physics, yet they produce nearly identical conductivities.\n\nI want to give credit where it is due. The paper is honest about the instability and even includes a limitation statement. The AL pipeline is the real contribution, and it should survive peer review. The QRd parameter-efficiency result for EC/EMC mixtures is well-supported. But the 6% conductivity claim for the electrolyte is overclaimed, and the abstract needs rephrasing or the authors need to generate longer stable QRd trajectories (perhaps with a different charge-assignment scheme or AL tailored to QRd).\n\nWho is this for? Anyone building system-specific MLIPs for battery electrolytes, especially those wanting to automate training-set collection. It deserves a serious referee—not a desk reject—but the referee should push hard on the QRd conductivity numbers and require either more statistics or a softer claim. I would accept a revised version with that fixed.","headline":"Solid active-learning pipeline for electrolyte MTPs; the 6% explicit-electrostatics conductivity claim is softer than the abstract suggests.","tokens_in":21075,"tokens_out":2404,"would_cite":true,"duration_ms":20917,"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":"Active learning can automatically build machine-learned potentials that simulate carbonate electrolytes, with ionic conductivities within 11% of experiment (6% when explicit electrostatics is added).","keywords":["active learning","moment tensor potentials","machine-learned interatomic potentials","ionic conductivity","electrolytes","charge redistribution","Green-Kubo","EC/EMC solvent"],"falsifier":"Run MTP20-QRd (or an environment-dependent-charge variant) for a 25 ns trajectory at 300 K in 3EC:7EMC with a 500 ps correlation time, matching the MTP20 protocol used here, and check whether the mean absolute deviation from experimental conductivity stays below 6%; if the Green-Kubo plateau shifts or the simulation blows up, that number is an artifact of short trajectories.","tokens_in":20097,"feed_emoji":"⚡","tokens_out":6680,"duration_ms":78727,"temperature":0.7,"pith_summary":"The paper makes two connected claims. First, D-optimality-based active learning—adding configurations only when a model's extrapolation grade crosses a threshold—can replace the manual, physically motivated dataset augmentation that earlier EC/EMC machine-learned potential work required, and can yield moment tensor potentials with stable liquid densities for ethylene carbonate, ethyl methyl carbonate, mixtures, and LiPF6 solutions. Second, explicitly adding electrostatics through a charge-redistribution (QRd) term delivers equal or better accuracy with fewer parameters in solvent mixtures, and predicts ionic conductivity with a 6% mean deviation from experiment. A reader should care because composition screening for battery electrolytes is bottlenecked by expensive training-set construction and by local potentials' neglect of long-range electrostatics. The salt-solution test is more equivocal: the fixed-charge QRd model produced only 2.8 ns of total trajectory, and the paper itself warns this undermines the reliability of its 6% conductivity figure.","feed_headline":"Active learning hits electrolyte conductivity within 11%","feed_subtitle":"D-optimality sampling builds carbonate-electrolyte potentials without manual tuning; explicit electrostatics cuts it to 6%.","key_machinery":"Two components carry the argument. (1) D-optimality active learning: from the matrix of energy derivatives with respect to MTP parameters, the MaxVol algorithm selects the most linearly independent rows; the extrapolation grade γ of any new configuration is the largest component of its derivative vector expressed in that basis, with thresholds γ_save≈2 and γ_break≈10 controlling when configurations are added and when AL-MD is stopped. This automates dataset growth. (2) The QRd charge-redistribution term: point charges qi = b_zi + s_zi (Q_total − Σb)/(Σs), fixed per atomic type, are added to the short-range MTP energy and fitted alongside the MTP parameters. Its analytic simplicity is what le","core_discovery":"On its own terms, the paper establishes that the D-optimality active-learning criterion, using the MaxVol algorithm and an extrapolation-grade threshold, can automatically assemble compact training sets (hundreds to about 6300 configurations) that keep MTP-based molecular dynamics in the liquid phase without manual density inflation or isolated-molecule augmentation. The resulting potentials transfer across EC/EMC ratios and 280–320 K, matching measured densities to about 6% and ionic conductivities to within 11% (MTP20). The paper further argues that augmenting a level-16 MTP with the QRd fixed-charge Coulomb term reaches the accuracy of a level-20 plain MTP with 389 rather than 651 machine","pith_inferences":["If environment-dependent charges (charge equilibration or equivariant charge prediction) stabilize the long-range term, the parameter-count and training-set advantages seen here could extend to salt solutions; the paper's own discussion points the same way.","The 6% conductivity figure should be read as provisional until a fixed-charge or environment-dependent QRd model sustains multi-nanosecond trajectories with a correlation time of at least 500 ps; a stable rerun is a direct test.","The near-identical conductivities despite clearly different ion-pair populations (MTP20: 100% solvent-separated pairs; MTP20-QRd: about 95% SSIPs plus some contact pairs and aggregates) suggest ionic conductivity at this salt concentration is insensitive to pairing in short trajectories—useful for prediction but weak as a structural discriminator.","The submitted metadata abstract says environment-dependent charges 'further improve accuracy and stability,' but the body reports MD instability for the implemented fixed-charge QRd scheme and lists environment-dependent charges as future work; readers should weigh the body's evidence over the abstract's wording."],"forward_implications":["The AL pipeline removes the need for manual training-set augmentation (density inflation/deflation, isolated molecules) that earlier EC/EMC MTP work required; training sets of roughly 500–1400 configurations sufficed for pure solvents and mixtures.","A short-range MTP without explicit electrostatics is enough for roughly 11% ionic-conductivity accuracy at multiple EC:EMC ratios and three temperatures, supporting the view that most electrostatics is captured within a 5 Å cutoff.","Explicit electrostatics buys parameter efficiency in the stable regime: a level-16 QRd model with 389 parameters matches a level-20 plain MTP with 651 parameters on energies and forces of solvent mixtures.","The QRd model's transport accuracy is coupled to a stability penalty in salt solutions: it reached 6% conductivity only at the trained composition and on trajectory statistics the paper itself calls insufficient.","Force errors concentrate on carbonyl carbons (and, in the salt solution, phosphorus), yet the Li–O radial distribution function still matched AIMD, so substantial relative force errors on Li+ (~40%) were tolerable for structure prediction by MTP20."],"fun_headline_variants":[],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The 6% conductivity result for MTP20-QRd rests on assuming that 2.8 ns of accumulated trajectory (14 completed runs at 300 K, 100 ps equilibration, 100 ps correlation time) is enough for a converged Green-Kubo integral—an assumption the paper itself says undermines the reliability of that number.","fun_headline_variants_meta":{"error":"'choices'"},"cache_creation_input_tokens":0},"created_at":"2026-08-04T11:39:30.536673+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run MTP20-QRd (or an environment-dependent-charge variant) for a 25 ns trajectory at 300 K in 3EC:7EMC with a 500 ps correlation time, matching the MTP20 protocol used here, and check whether the mean absolute deviation from experimental conductivity stays below 6%; if the Green-Kubo plateau shifts or the simulation blows up, that number is an artifact of short trajectories.","supporting_citations":[],"review_version":1}