{"id":"d66ba27b-17b6-4e2f-b340-ba486a2d55a0","arxiv_id":"2607.20591","paper_version":1,"verdict":"ACCEPT","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"Li-projected phonon spectra encoded with Wasserstein distance predict experimental solid-electrolyte conductivity with held-out R²=0.44, far above total-DOS (0.01) and static descriptors (0.18).","lead":"The paper tests whether the lithium-specific vibration spectrum (Li-PDOS) predicts which solid electrolytes conduct lithium ions, and reports it substantially outperforms total vibration spectra and static chemistry descriptors. The proposed descriptor is a cheap way to rank electrolyte candidates before expensive molecular-dynamics or first-principles screening.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"MatterSim Li-projection bias, validated on only 10 phase-matched references, could explain the Li-PDOS conductivity signal; a broader DFT check is needed.","rationale":"The reader correctly identified the MatterSim Li-projection proxy as the weakest assumption. I agree that this is the single most load-bearing point: the scientific claim is that lithium-resolved spectral distributions carry reproducible information about ionic conductivity, but the only direct evidence for the Li-projection fidelity comes from ten phase-matched references showing systematic softening. The paper is transparent about this limitation and carefully avoids causal language, but the descriptor claim still rests on the assumption that the MLIP Li projection is not biased in a way that is correlated with the target variable. The family-demeaning attenuation and the exact-cohort HSIC reversal are internal warning signs that the signal may be partly confounded by composition or family. Because the paper is otherwise methodologically strong and the concern is testable, the appropriate verdict is conditional acceptance rather than rejection: the main result should be confirmed with a broader DFT-projection validation and a leave-one-family-out predictive test. If those checks pass, the current ACCEPT verdict would stand; if they fail, the descriptor claim would need to be substantially weakened.","tokens_in":14130,"tokens_out":4826,"duration_ms":48299,"concrete_test":"Expand the external validation to at least 30 phase-matched reference crystals spanning the conductivity range and electrolyte families of OBELiX, computing both MatterSim and DFT Li-PDOS and total DOS. Then (i) regress the MatterSim-vs-DFT Li Wasserstein distance and centroid error on experimental log conductivity and on the Li fraction below 5 THz; (ii) for the overlapping materials, retrain the Wasserstein model on DFT-based Li-PDOS and evaluate on the OBELiX test set. If the MatterSim Li-projection error correlates with conductivity, or if DFT-based Li-PDOS loses its R² advantage over total DOS, the claimed robust Li-projection descriptor is not supported. As a cheaper first check, report leave-one-family-out held-out R² for Li-Wasserstein versus static and total-DOS models.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim depends on MatterSim Li-projected harmonic phonon DOS being a faithful proxy for true lithium vibrational spectra. The only external validation compares 10 phase-matched materials (reported as 20 comparisons, double-counting total and Li projections from the same calculation) and finds the Li projection is systematically softer than the total DOS: mean centroid error -0.711 THz vs -0.410 THz, and mean Wasserstein distance 0.731 THz vs 0.542 THz. If this projection error is correlated with composition or conductivity — e.g., because fast-ion conductors are more anharmonic or disordered, or because MLIP force errors vary by chemistry — then the low-frequency Li spectral redistribution and the held-out R²=0.444 advantage over total DOS (R²=0.012) could be a computational artifact rather than a physical descriptor signal. The paper's own family-demeaning analysis is consistent with this worry: the Li fraction below 5 THz correlation drops from 0.538 to ~0.150 after family adjustment, and the Li-vs-total HSIC ordering reverses in the exact-cohort. Ten reference materials are too few to rule out a conductivity-correlated bias in the Li projection.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper tests whether the complete lithium-projected phonon density of states (Li-PDOS), treated as a distribution and compared via Wasserstein geometry, is a robust descriptor of experimental room-temperature ionic conductivity. Using MatterSim MLIP forces and Phonopy, the authors compute harmonic total and Li-projected spectra for OBELiX entries after an audit of composition/structural fidelity, preserving the official train/test split. They construct nested cohorts (primary 260, strict 241, exact 168) and report that higher conductivity is associated with low-frequency redistribution of normalized Li spectral weight. In the untouched test set, a Li-PDOS Wasserstein kernel achieves R²=0.444, versus 0.012 for total DOS and 0.181 for a static composition–structure kernel, with similar R² values in the strict and exact cohorts. The paper includes extensive sensitivity analyses, multiplicity control, external DFT reference comparisons, and acknowledges limitations related to MLIP softening and family confounding.","tokens_in":14353,"tokens_out":4545,"duration_ms":50566,"significance":"If the claims hold, the paper provides a practical screening descriptor for solid electrolytes and a methodologically careful template for evaluating distribution-valued descriptors. The validation-first design is a clear strength: the official split is respected, frequency-window selection is training-only, permutation/FDR corrections are applied, and the code/data are released. External comparison with NIMS/PhononDB, though limited in size, is a valuable addition. The authors are appropriately cautious about causal interpretation, and the paper's explicit discussion of MLIP softening and family attenuation is honest. However, the central claim depends on the Li projection being a faithful proxy for true lattice dynamics, and the current validation leaves a nontrivial risk of conductivity-correlated bias that could explain part of the Li-PDOS advantage.","major_comments":[{"comment":"The external validation is too limited to rule out a conductivity-correlated bias in the Li projection. The mean W1 distance is 0.731 THz and the mean centroid error is -0.711 THz for Li-PDOS, based on only 10 phase-matched materials (20 comparisons counting total and Li separately). This is a systematic, projection-dependent softening. If the MatterSim Li-force error correlates with composition or conductivity—e.g., because fast-ion conductors are more anharmonic or contain heavier elements—the held-out R²=0.444 advantage over total DOS could reflect a computational artifact rather than physical information. The paper itself notes that \"total-DOS agreement cannot validate a species projection\" but does not provide a test of whether the Li projection error is conditionally independent of the target. Please expand the reference set to a broader, conductivity-stratified sample, or demonstr","section":"§2.4, §3.2, Fig. 1"},{"comment":"The family-confounding analysis is not extended to the prediction comparison. The marginal training correlation for Li fraction below 5 THz is ρ=0.538, but after family demeaning it drops to approximately 0.150, and the test interval crosses zero. Moreover, the Li-vs-total HSIC ordering reverses in the exact cohort (Li 0.125 vs total 0.184). These observations show that the pooled Li-PDOS advantage may be driven by family composition. The paper reports residual HSIC in some cohorts, but not a family-demeaned or within-family prediction comparison between Li and total spectra. To support the claim that Li-PDOS contains information \"not captured by total DOS,\" please report leave-one-family-out or family-demeaned R² for the Li-Wasserstein vs total-Wasserstein and static models on the official test set, or otherwise quantify how much of the Li advantage survives control for family identity.","section":"§3.5, Fig. 4"},{"comment":"The p-values in Figure 4c are presented without clearly distinguishing which permutation null (global vs within-family) each p-value refers to. The text states HSIC is significant \"under both global and within-family permutations,\" but the figure symbol legend and caption imply all listed p-values come from the same permutation procedure. Please clarify by explicitly labeling the permutation type for each p-value, or moving the within-family p-values to a separate panel/table.","section":"§3.5, Fig. 4c"}],"minor_comments":[{"comment":"Typo: \"rooot-mean squared\" should be \"root-mean squared\" in the first paragraph.","section":"§3.6"},{"comment":"Heading contains a stray space: \"F requency-Resolved\" should be \"Frequency-Resolved.\"","section":"§3.4 heading"},{"comment":"The software is referred to as both \"PhonoPy\" and \"Phonopy\" in the text; please standardize to the official spelling \"Phonopy.\"","section":"§2.2"},{"comment":"The description \"20 independently generated phonon-calculation database comparisons\" is somewhat misleading because the 20 comparisons come from 10 materials with two representations each. The text later clarifies this, but readers may initially interpret it as 20 independent crystal structures.","section":"§2.4"},{"comment":"The exact-cohort HSIC reversal is discussed in the text but not shown quantitatively in Figure 4; please point more explicitly to the supporting figure or table showing the exact-cohort HSIC values and the permutation intervals.","section":"§3.5"}],"recommendation":"major_revision","confidential_remarks":"The paper is methodologically strong and transparent, but the central claim hinges on the Li-projected MLIP spectra being unbiased, and the current validation base is too thin to establish that. The family-confounding issue also tempers the 'not captured by total DOS' phrasing. I would not reject, but I would require additional validation or recalibration before accepting."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The short version: this is a careful, validation-first study showing that the full Li-projected phonon spectrum, encoded with a Wasserstein kernel, predicts experimental log conductivity much better than total DOS or a static composition-structure baseline (R²=0.444 vs 0.012 and 0.181 on the official OBELiX test set). The design is genuinely out-of-sample: frequency windows are chosen on training only, the official split stays untouched until the end, and the strict and exact-composition cohorts reuse fixed hyperparameters. That is more discipline than most papers in this space.\n\nWhat is new is the demonstration that the complete mobile-ion-resolved spectrum, not a scalar softness descriptor, carries transferable information about conductivity. Prior work mostly compressed phonons into band centers or used total DOS. The Wasserstein kernel on Li-PDOS is a sensible way to treat spectra as distributions.\n\nThe soft spot is real, and the reader's stress-test note points at it accurately. The external validation of the MatterSim Li projection uses only 10 phase-matched materials, and the paper calls that 20 comparisons by counting total DOS and Li-PDOS separately. That wording overstates independent reference calculations. More importantly, the comparison shows systematic softening that is larger for Li (mean centroid error -0.711 THz vs -0.410 THz), and ten references is too few to rule out that the projection error correlates with composition or with conductivity. The paper's own family-demeaning analysis feeds the same worry: the Li fraction below 5 THz correlation drops from 0.538 to about 0.150 after family adjustment, and the Li-vs-total HSIC ordering reverses in the exact cohort. So the mechanistic interpretation is fragile. But the paper itself bends over backward to avoid a causal claim, explicitly calling the result a comparative screening descriptor and not a universal softness law. The comparative claim, on this dataset with this MLIP, is supported.\n\nMinor issues: the '20 independent comparisons' sentence should be rewritten, and I could not verify the GitHub contents from the text alone. Neither undermines the central evidence.\n\nWho is this for: people doing solid-electrolyte screening with MLIPs, and anyone else evaluating whether species-projected phonons from universal potentials can be used as descriptors. It deserves a serious referee. The main revision should be a broader DFT validation set — even non-phase-matched references would help — and a fix to the comparison-count wording.","headline":"A carefully designed comparative study showing full Li-PDOS beats total-DOS and static baselines on OBELiX, but the Li-projection validation rests on only 10 references and the mechanistic gloss is fragile — worth a serious referee, not a desk reject.","tokens_in":14856,"tokens_out":2492,"would_cite":true,"duration_ms":24256,"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 complete lithium-projected phonon spectrum, viewed through Wasserstein geometry, carries reproducible information about experimental room-temperature ionic conductivity that scalar softness measures and total phonon spectra miss.","keywords":["solid electrolyte","ionic conductivity","lithium-projected phonon density of states","Wasserstein kernel","phonon softness","machine-learning interatomic potential","materials informatics","lattice dynamics"],"falsifier":"A direct test would be to recompute lithium-projected phonon spectra for a few dozen representative materials using first-principles density-functional-theory forces and check whether the low-frequency lithium weight redistribution and the Wasserstein-model R² survive; if the correlation with conductivity collapses under the more accurate spectra, the descriptor is an artifact of the approximate force model.","tokens_in":13986,"feed_emoji":"🔋","tokens_out":3582,"duration_ms":29405,"temperature":0.7,"pith_summary":"The paper tests whether the full frequency-resolved vibrational spectrum of lithium ions—not just a single 'softness' number—can predict measured room-temperature ionic conductivity across chemically diverse solid electrolytes. Using machine-learned forces to generate harmonic spectra for hundreds of crystallographically resolved materials, it finds that higher-conductivity materials systematically shift normalized lithium spectral weight toward low frequencies, a pattern that replicates in an untouched test set. A Wasserstein kernel on the lithium-projected density of states yields held-out R²=0.444, versus 0.012 for the total phonon spectrum and 0.181 for a static composition–structure baseline, and remains stable under stricter composition auditing. The paper argues that mobile-ion-resolved spectral distributions are robust comparative screening descriptors, while cautioning that they are not a universal causal law.","feed_headline":"Lithium phonon spectrum, not total, predicts electrolyte conductivity","feed_subtitle":"A Wasserstein kernel on Li-projected modes triples held-out R² versus the total phonon spectrum in a 260-material test.","key_machinery":"The central object is the unit-area-normalized lithium-projected phonon density of states, treated as a probability distribution over frequency. The Wasserstein-1 distance between such distributions, implemented through quantile embeddings, provides a kernel that respects the ordering of the frequency axis, unlike bin-wise vector representations. The paper also uses cumulative-frequency descriptors such as the fraction of spectral weight below 2 and 5 THz and the 5% quantile frequency. These carry the argument by converting raw spectra into distribution-aware features that separate conductive from non-conductive materials.","core_discovery":"The central claim is that the complete lithium-projected phonon density of states (Li-PDOS)—the distribution of vibrational modes weighted by lithium motion—contains transferable information about experimental ionic conductivity that is lost when the spectrum is compressed to a band center or other scalar, and also lost when only the total density of states is used. In the paper's primary cohort of 260 materials, the Li-PDOS Wasserstein model achieves held-out R²=0.444 versus 0.012 for total DOS and 0.181 for a static kernel. The low-frequency redistribution direction and cumulative-onset shift replicate in the untouched test set, and the R² remains near 0.45 under strict and exact-compositi","pith_inferences":["If the Li-PDOS Wasserstein signal is reproducible across datasets beyond the one used here, it could be combined with anharmonic and defect-aware simulations to separate genuine lattice-dynamical softness from compositional confounding.","The family-attenuation of scalar descriptors suggests the full-distribution model may work by capturing cross-family differences; a natural extension is to test whether the Wasserstein kernel still adds value for substitutions within a single electrolyte family.","Because the harmonic spectrum is computed at zero temperature, the descriptor may be a proxy for something else—such as coordination or polarity—so a decisive test would be to correlate the same low-frequency redistribution with independently measured migration barriers or ab initio molecular-dynamics conductances.","The systematic lithium-projection softening reported for the approximate force model suggests that calibrating against first-principles spectra for a small representative set could improve the descriptor's quantitative accuracy without losing its ranking power."],"forward_implications":["Screening workflows should retain mobile-ion-projected spectra instead of only total DOS or scalar softness descriptors.","A Wasserstein-kernel model on Li-PDOS can rank candidate electrolytes before expensive molecular dynamics or first-principles diffusion calculations.","The observed low-frequency redistribution provides a specific, testable spectral signature (increased Li weight below ~5 THz, earlier cumulative onset) for prioritizing materials.","Audit and validation practices—checking composition, integerization, and comparing against first-principles references—become standard parts of MLIP-based phonon screening.","Total-DOS agreement with reference calculations cannot validate a mobile-ion projection; species-resolved validation is necessary."],"fun_headline_variants":["Li-projected phonons triple held-out R² over total spectrum","Full Li-phonon spectrum beats total DOS for conductivity","Li-weighted phonons outperform scalars in electrolyte screening","Phonon distribution on Li modes predicts ion transport better","Li-phonon kernel doubles static descriptor accuracy"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The calculated lithium phonon spectra, generated from a machine-learned interatomic potential applied to simplified ordered versions of the real crystal structures, accurately represent the true lattice vibrations of these diverse electrolytes.","fun_headline_variants_meta":{"raw":{"variants":["Li-projected phonons triple held-out R² over total spectrum","Full Li-phonon spectrum beats total DOS for conductivity","Li-weighted phonons outperform scalars in electrolyte screening","Phonon distribution on Li modes predicts ion transport better","Li-phonon kernel doubles static descriptor accuracy"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000197,"raw_usage":{"total_tokens":1272,"prompt_tokens":886,"completion_tokens":386,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":630,"completion_tokens_details":{"reasoning_tokens":320}},"tokens_in":630,"tokens_out":386,"duration_ms":4310,"temperature":1.0,"reasoning_tokens":320,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-01T10:21:11.431988+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A direct test would be to recompute lithium-projected phonon spectra for a few dozen representative materials using first-principles density-functional-theory forces and check whether the low-frequency lithium weight redistribution and the Wasserstein-model R² survive; if the correlation with conductivity collapses under the more accurate spectra, the descriptor is an artifact of the approximate force model.","supporting_citations":[],"review_version":1}