{"id":"3778868a-09f8-441c-9172-a8bc29d6fb48","arxiv_id":"2505.18571","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A dual-stage ML workflow combining fine-tuned CHGNet and an interpretable structure-to-diffusion model screens 4575 high-entropy LZSP compositions and identifies Li2.625Zr0.25Hf0.1875Sn0.1875Ti0.1875Nb0.1875Si2PO12 with predicted high ionic conductivity.","lead":"This paper builds a two-stage machine learning pipeline that first relaxes high-entropy solid electrolyte structures with a tuned neural network potential, then predicts lithium mobility from simple structural features. The authors use it to screen thousands of Li3Zr2Si2PO12-based compositions and claim a specific candidate with a projected room-temperature conductivity of 4.53 mS/cm, three orders of magnitude above the starting material.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 4.53 mS/cm room-temperature value depends entirely on an Arrhenius extrapolation whose temperature range, fit quality, and uncertainties are not reported; the reported 0.24 eV barrier is not established at 300 K.","rationale":"The reader's weakest assumption is exactly the Arrhenius extrapolation; I agree. The paper's contribution is a two-stage screening workflow, and that part is supported by internal validation: fine-tuned CHGNet is tested against DFT/AIMD on a few structures, and SF-MSD is tested against held-out structures. However, the headline application example is only as strong as the extrapolated room-temperature conductivity. The text gives no details of the Arrhenius fit, and the exponential sensitivity of the 300 K value makes this the single most load-bearing step. Other concerns, such as the SF-MSD model's larger absolute errors in the high-MSD region, affect screening confidence but do not by themselves invalidate the framework. The recommended verdict remains CONDITIONAL: the discovery claim should not be accepted as a quantitative result until the Arrhenius fit details are provided and the linearity assumption is tested. Because this does not change the reader's verdict, I mark the adjustment as UNCHANGED.","tokens_in":15955,"tokens_out":4782,"duration_ms":41921,"concrete_test":"Recompute the Arrhenius plot in Fig. 6(c) using fine-tuned CHGNet NVT-MD at 400, 500, 600, 700, 800, 900, and 1000 K for both LZHSTNSP and pristine LZSP, with 100+ ps trajectories and bootstrap error bars on each conductivity point. Then fit ln(sigma*T) vs 1/T and report the 300 K intercept with its uncertainty. If the low-temperature points deviate from the line fitted above 800 K, or the extrapolated 300 K conductivity changes by more than a factor of 2 from 4.53 mS/cm, the discovery claim is unsupported. Also compute the Li RDF and cell symmetry at each temperature to rule out a phase transition in the extrapolation window.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central discovery claim is that LZHSTNSP reaches 4.53 mS/cm at 300 K, with the migration barrier reduced from 0.46 to 0.24 eV. That quantitative claim is produced by an Arrhenius fit: Section 4.2 states that all dataset MD runs used 1000 K except for 'the final MD simulation that was used to fit the Arrhenius curve,' but it does not state how many temperatures were used, which temperatures, the fit range, the fitted prefactor, or any uncertainty on the 300 K intercept. Extrapolating a slope fitted at high temperature across a roughly 700 K gap assumes a single diffusion mechanism, no phase transition or order-disorder change, and no temperature dependence of attempt frequency or correlation factor; none of these is checked. The only AIMD consistency check in the Supplementary (Fig. S10) is for pristine LZSP at 1000/1100 K, not for the candidate at lower temperature. Because the 300 K conductivity depends exponentially on the barrier, a plausible ±0.05 eV uncertainty in the fitted slope changes the extrapolated value by a factor of roughly exp(0.05/k_B*300 K) ≈ 7, so the claim of 'three orders of magnitude' is not robust without reporting the fit. If the true low-temperature mechanism is blocked or changes, the number could be off by far more.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes a dual-stage machine-learning framework for discovering high-entropy solid electrolytes in the Li3Zr2Si2PO12 (LZSP) family. In the first stage, a pretrained CHGNet potential is fine-tuned on DFT relaxation and MD data to provide fast structural relaxation and molecular dynamics. In the second stage, a structure-feature-to-mean-squared-displacement (SF-MSD) surrogate model, comparing SISSO, gradient boosting, and random forest, is trained on CHGNet MD results and used to screen 4575 quinary HE-LZSP compositions. The authors identify Li2.625Zr0.25Hf0.1875Sn0.1875Ti0.1875Nb0.1875Si2PO12 (LZHSTNSP) and report an extrapolated room-temperature ionic conductivity of 4.53 mS/cm, three orders of magnitude higher than pristine LZSP, with a migration barrier reduced from 0.46 to 0.24 eV. The paper also presents SHAP analyses linking transport to features such as lithium vacancy concentration and octahedral distortion.","tokens_in":16241,"tokens_out":3459,"duration_ms":30487,"significance":"If the screening results are reliable, this work would provide a practical template for accelerating discovery in large composition spaces of high-entropy solid electrolytes, where exhaustive DFT MD is intractable. The framework is appealing in combining a fine-tuned universal potential with a lightweight interpretable surrogate, and the authors report several useful validation checks: CHGNet relaxation errors are compared against DFT for selected compositions, an AIMD comparison is provided for a quaternary structure, and model selection among SISSO, GB, and RF is documented with test-set RMSE values. The manuscript also makes a substantial computational dataset available (about 500 MD-simulated HE-LZSP compositions and over 5000 relaxed structures with predicted MSDs), which is a clear strength. However, the central quantitative discovery claim rests on an Arrhenius extrapolation whose details and uncertainties are not reported, and the validation chain is largely circular because the surrogate is trained and evaluated on the same MLIP-MD methodology used for the final candidate. These issues must be resolved before the quantitative conductivity claim can be accepted.","major_comments":[{"comment":"The central claim that LZHSTNSP has a room-temperature ionic conductivity of 4.53 mS/cm and a migration barrier of 0.24 eV is obtained by an Arrhenius extrapolation, but the manuscript does not report the number of MD temperatures used for the fit, the temperature range, the fitted prefactor, the goodness of fit, or any uncertainty on the extracted barrier and intercept. Section 4.2 only states that 'all simulations were set at 1000 K with the exception of the final MD simulation that was used to fit the Arrhenius curve.' This omission is load-bearing: the conductivity depends exponentially on the barrier, so a plausible ±0.05 eV uncertainty in the fitted slope changes the 300 K extrapolation by a factor of roughly exp(0.05/k_B·300 K) ≈ 7. Without reporting the fit details and statistical uncertainties, the 'three orders of magnitude' improvement over pristine LZSP cannot be critically assessed.","section":"§2.3, Fig. 6(c); §4.2"},{"comment":"The only AIMD consistency check for the LZSP framework is performed on pristine LZSP at 1000 K and 1100 K (Fig. S10), not on the candidate LZHSTNSP or at lower temperatures. All CHGNet MD data in the dataset are generated at 1000 K (except the final Arrhenius simulation, whose temperature is not specified). Extrapolating over roughly 700 K assumes that the diffusion mechanism, attempt frequency, and correlation factor are temperature-independent and that no phase transition or order–disorder change occurs. None of these assumptions is tested. To support the quantitative room-temperature claim, the authors should provide at least one independent check for the candidate, such as AIMD or DFT NEB barrier calculations, or explicitly reframe the reported value as a high-temperature screening metric rather than an established 300 K conductivity.","section":"§2.3; Supplementary Fig. S10"},{"comment":"The validation of the SF-MSD model is circular with respect to the transport property being predicted. The model is trained on MSD values obtained from fine-tuned CHGNet MD, tested on further fine-tuned CHGNet MD, and the final candidate is selected by the surrogate and then 'verified' by the same fine-tuned CHGNet MD potential. Consequently, systematic errors in CHGNet's prediction of ionic transport are not probed. Moreover, Table 2 reports multinary test-set RMSEs of 20.3 Ų (GB) and 20.1 Ų (RF), with quaternary test-set RMSEs of 23.9–24.8 Ų, and Fig. 6(a) states that the largest absolute errors occur in the high-MSD region, which is exactly where the candidate lies. The authors should report the per-region errors, the actual MD MSD for LZHSTNSP, and an independent validation (e.g., AIMD or DFT NEB) for the candidate to break the circularity.","section":"§2.2, §2.3; Table 2; Fig. 6(a,b)"},{"comment":"The screening procedure for the 4575 quinary compositions is not fully specified. The manuscript categorizes predicted MSD values into high, medium, and low regions (MSD>120, 40–120, <40 Ų) and selects 10 random structures per region for MD validation, but it does not state how the final candidate LZHSTNSP was chosen from the high-MSD region, nor does it quantify the screening's false-discovery rate. Given that only 30 of 4575 candidates are MD-verified, the paper should report the selection rule and show how many of the 30 validation samples were correctly classified, so the reader can assess the reliability of the high-throughput screening claim.","section":"§2.3"}],"minor_comments":[{"comment":"The caption contains a typo ('Comparation') and should define the notation 'Lix', 'csm_Xavg', 'csm_Xsd', and 'X-O' explicitly.","section":"Fig. 4 caption"},{"comment":"The sentence 'This dataset was partitioned, allocating a small subset as the training set for the SF-MSD model and the majority as the predicted targets' is unclear; it should specify the number of training, validation, and test compositions and clarify the role of the 'predicted targets.'","section":"§2.2"},{"comment":"The MSD analysis protocol is described for the 100 ps dataset runs (averaging the 20–80 ps window), but it is not stated whether the same protocol was used for the final Arrhenius MD simulations. This should be clarified for reproducibility.","section":"§4.2"},{"comment":"The figure numbering in the supplementary material is disordered (Fig. S10 appears after Fig. S11 in the supplied text), and the caption for Fig. S10 should be placed with the figure. Please renumber all supplementary figures consistently.","section":"Supplementary Information"},{"comment":"The sentence 'The results confirm that the current model maintains strong predictive performance, as effective dataset generation strikes an optimal balance between computational cost and model accuracy' is vague; the cross-validation RMSEs (28.7 Ų for GB, 26.1 Ų for RF) are not compared statistically with the main model's RMSEs, so the claim of 'optimal balance' is not supported by a quantitative comparison.","section":"§2.3"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is within the scope of the journal and addresses a timely problem. The main concern is that the headline result (4.53 mS/cm at room temperature) is a high-temperature Arrhenius extrapolation with no reported fit range or uncertainty, and the validation chain is largely self-referential because the surrogate and the final MD verification both use the fine-tuned CHGNet potential. I am not questioning the authors' integrity; the issue is that the manuscript as written does not provide enough independent evidence to support the quantitative conductivity claim. The recommendations in the report are intended to be fixable within the scope of a revision: report the Arrhenius fit details and uncertainties, add an independent check for the candidate, and clarify the screening selection rule."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe paper is worth a look, but the famous number—4.53 mS/cm at room temperature for LZHSTNSP—rests on an Arrhenius extrapolation described in one sentence. That is the main thing to know.\n\nWhat is genuinely new: the two-stage workflow (fine-tuned CHGNet for relaxation/MD, then an interpretable SF-MSD model) applied to the HE-LZSP space, a dataset of 500+ MD runs and 5000+ relaxed structures, and a specific predicted composition. The fine-tuning is careful: they sample transition-state-rich images from DFT relaxation paths for the MD training set, and they validate the potential against AIMD on both pristine LZSP and a quaternary HE structure. The SF-MSD model choice is reasonable, and the SHAP analysis gives useful mechanistic hints. The authors are also honest that the surrogate has larger absolute errors in the high-MSD region, though they argue relative errors stay stable.\n\nThe soft spots are real and concentrated in the central discovery claim. Section 4.2 says all training MD ran at 1000 K “with the exception of the final MD simulation that was used to fit the Arrhenius curve,” but it never states which temperatures, how many, the fit range, the fitted prefactor, or any uncertainty on the 300 K intercept. The activation-energy drop from 0.46 to 0.24 eV is the same extrapolation. Given an exponential sensitivity—a plausible ±0.05 eV changes the 300 K value by roughly 7×—the “three orders of magnitude” statement is not robust as reported. The auxiliary AIMD checks are for the host and one quaternary composition at 1000–1100 K, not for the candidate at lower temperature. So the room-temperature value is a simulation-based projection, not a well-characterized result.\n\nThere is also a circularity concern: the SF-MSD model is trained on CHGNet MD and validated largely on further CHGNet MD, so the screening tightly tracks the potential's biases. The AIMD cross-checks mitigate this but do not eliminate it. No data or code are released, which limits reproducibility.\n\nBottom line: the framework is a plausible, useful engineering contribution for high-entropy solid electrolyte screening, and the dataset could be valuable if shared. The quantitative discovery claim needs substantial support before it can be taken as is. I'd send this to peer review with a request for full Arrhenius details, uncertainty propagation, and ideally direct AIMD validation on the candidate at intermediate temperatures.","headline":"A useful screening workflow and a new dataset, but the banner 4.53 mS/cm conductivity claim rests on an undocumented Arrhenius extrapolation.","tokens_in":16813,"tokens_out":2151,"would_cite":false,"duration_ms":18450,"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":"A two-stage machine-learning pipeline can screen thousands of high-entropy solid electrolytes and points to a candidate with 4.53 mS/cm room-temperature ionic conductivity.","keywords":["high-entropy solid electrolytes","machine learning interatomic potentials","CHGNet","ionic conductivity","NASICON","Li3Zr2Si2PO12","structure-property prediction","high-throughput screening"],"falsifier":"Synthesize LZHSTNSP and measure its ionic conductivity at 300 K, or run the same CHGNet molecular dynamics at 400, 600, and 800 K and check whether the Arrhenius plot remains linear with the same 0.24 eV slope; a bend in the plot or a measured conductivity far from 4.53 mS/cm would falsify the extrapolation.","tokens_in":15737,"feed_emoji":"🔋","tokens_out":5706,"duration_ms":45689,"temperature":0.7,"pith_summary":"This paper tries to show that a two-stage machine-learning pipeline can replace trial-and-error in the search for high-entropy solid electrolytes. The first stage fine-tunes a pretrained neural-network interatomic potential on DFT data for the NASICON electrolyte Li$_3$Zr$_2$Si$_2$PO$_{12}$ (LZSP), so that atomic relaxations and molecular dynamics run at near-DFT accuracy but at a fraction of the cost. The second stage maps simple relaxed-structure features, lithium content, cell parameters, and oxygen-coordination distortions, onto lithium mean squared displacement, producing a fast predictor for ionic transport. Applied to 4,575 quinary high-entropy LZSP compositions, the pipeline singles out Li$_{2.625}$Zr$_{0.25}$Hf$_{0.1875}$Sn$_{0.1875}$Ti$_{0.1875}$Nb$_{0.1875}$Si$_2$PO$_{12}$, whose extrapolated room-temperature conductivity is 4.53 mS/cm, about three orders of magnitude above pristine LZSP, with a migration barrier reduced from 0.46 to 0.24 eV. If true, this would cut the evaluation of one candidate composition from days to minutes and give a transferable recipe for other solid-electrolyte families.","feed_headline":"ML screens 4,575 electrolytes, finds one 1,000x more conductive","feed_subtitle":"Fine-tuned neural potential plus structural descriptors cut per-candidate evaluation from days to minutes.","key_machinery":"The load-bearing machinery is a dual-stage pipeline. Stage one is a fine-tuned CHGNet interatomic potential: the pretrained universal potential is refined first for structural relaxation and then for molecular dynamics on DFT-generated images of LZSP-family structures, including images chosen specifically for lithium transition-state geometry. Stage two is the SF-MSD model, a regression on 17 structural descriptors of the relaxed cell that outputs the lithium mean squared displacement, replacing direct molecular-dynamics screening. The bridge between stages is the assumption that relaxed-structure descriptors encode the bottleneck geometry and percolation network that control diffusion. The final conductivity figure comes from an Arrhenius fit of CHGNet molecular dynamics at elevated temperatures down to 300 K.","core_discovery":"The central claim is that ionic transport in high-entropy LZSP can be predicted from the relaxed structure alone, without running expensive molecular dynamics for every candidate. By fine-tuning CHGNet on a purpose-built dataset of relaxation paths and lithium transition-state images, the authors obtain a potential that reproduces DFT energies, forces, cell volumes within about 1%, and molecular-dynamics mean-squared-displacement curves for doped and high-entropy variants. On top of that, the SF-MSD model uses structural descriptors such as lithium content, unit-cell parameters, bond lengths, and continuous symmetry measures of the oxygen polyhedra to predict the 100-ps, 1000 K mean squared displacement, with gradient-boosting and random-forest variants outperforming a compressed-sensing descriptor baseline on out-of-sample quaternary and multinary tests. The framework's payoff is the identification of LZHSTNSP in the 4,575-composition quinary space, with a predicted migration barrier of 0.24 eV and extrapolated 300 K conductivity of 4.53 mS/cm. The paper also uses feature-importance attribution to argue that the conductivity gain comes from a favorable lithium vacancy concentration together with uniform Zr-site octahedral distortions, not from any single dopant.","pith_inferences":["The paper reports the 4.53 mS/cm value as an extrapolation; a natural next step the authors do not take is to run CHGNet molecular dynamics at intermediate temperatures such as 400, 600, and 800 K to check whether the Arrhenius line stays straight and whether the same diffusion mechanism persists.","Because the SF-MSD model is trained on 1000 K mean-squared-displacement values, applying it to room-temperature design implicitly assumes that ranking by high-temperature MSD equals ranking by room-temperature conductivity, an assumption that could be tested directly by comparing predicted and measured conductivities on a handful of compositions.","A similar descriptor-based shortcut could be built for other framework families, but the 17 selected descriptors would need re-derivation because bottleneck geometry and coordination chemistry differ.","The framework does not include thermodynamic stability or synthesis feasibility, so composition screening would need to be paired with phase-stability checks before experimental follow-up."],"forward_implications":["Evaluation time per candidate drops from days to minutes, so a 4,575-composition space becomes screenable in one pass.","The fine-tuned potential and SF-MSD predictor are released as a queryable dataset, letting researchers search high-entropy LZSP compositions by element, proportion, and formula.","The identified composition LZHSTNSP is a concrete candidate for experimental synthesis and impedance testing.","The same two-stage recipe is claimed to transfer to other high-entropy solid electrolytes such as garnets and argyrodites, and to other properties of high-entropy cathode materials.","The feature-importance analysis identifies lithium vacancy concentration and octahedral distortion uniformity as design levers, rather than composition alone."],"supporting_citations":[{"why":"Supplies the pretrained universal neural-network potential that the paper fine-tunes for the LZSP family.","marker":"[20]"},{"why":"Provides the initial pristine LZSP structure and the observation that ionic transport is sensitive to bottleneck sizes.","marker":"[31]"},{"why":"Establishes the high-entropy mechanism for boosting ionic conductivity, the target phenomenon the framework aims to exploit.","marker":"[16]"},{"why":"Documents the trial-and-error constraints and high-dimensional composition space that motivate a computational screening strategy.","marker":"[14]"},{"why":"Supplies the 20-to-80 ps averaging protocol used to extract statistically reliable mean squared displacements from molecular dynamics.","marker":"[33]"},{"why":"Supports the corner-sharing framework screening rationale and the use of structural features to identify superionic conductors.","marker":"[13]"},{"why":"Supplies the continuous symmetry measures used to describe oxygen-polyhedra distortions in the structural-feature set.","marker":"[36]"},{"why":"Provides the softening analysis that motivates stepwise fine-tuning of the universal potential for relaxation and dynamics.","marker":"[32]"}],"fun_headline_variants":["Dual-stage ML finds fast ion conductor among 4,575 high-entropy solids","Structure predicts ionic transport: ML speeds up high-entropy SE search","Fine-tuned neural potential trims MD, pinpoints 4.53 mS/cm electrolyte","Machine learning screens 4,575 electrolytes without full molecular dynamics","AI maps ionic transport from structure alone in high-entropy solid electrolytes"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The headline room-temperature conductivity assumes that lithium diffusion follows the same unchanged mechanism from the 1000 K simulation temperature down to 300 K, without any phase transition or ordering effect in between.","fun_headline_variants_meta":{"raw":{"variants":["Dual-stage ML finds fast ion conductor among 4,575 high-entropy solids","Structure predicts ionic transport: ML speeds up high-entropy SE search","Fine-tuned neural potential trims MD, pinpoints 4.53 mS/cm electrolyte","Machine learning screens 4,575 electrolytes without full molecular dynamics","AI maps ionic transport from structure alone in high-entropy solid electrolytes"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000331,"raw_usage":{"total_tokens":1900,"prompt_tokens":1058,"completion_tokens":842,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":674,"completion_tokens_details":{"reasoning_tokens":742}},"tokens_in":674,"tokens_out":842,"duration_ms":7366,"temperature":1.0,"reasoning_tokens":742,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T14:29:01.453776+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Synthesize LZHSTNSP and measure its ionic conductivity at 300 K, or run the same CHGNet molecular dynamics at 400, 600, and 800 K and check whether the Arrhenius plot remains linear with the same 0.24 eV slope; a bend in the plot or a measured conductivity far from 4.53 mS/cm would falsify the extrapolation.","supporting_citations":[{"cited_title":"Ouyang, Y","cited_arxiv_id":null,"evidence_quote":"Documents the trial-and-error constraints and high-dimensional composition space that motivate a computational screening strategy."},{"cited_title":"He, Y .Z","cited_arxiv_id":null,"evidence_quote":"Supplies the 20-to-80 ps averaging protocol used to extract statistically reliable mean squared displacements from molecular dynamics."},{"cited_title":"Jun, Y .Z","cited_arxiv_id":null,"evidence_quote":"Supports the corner-sharing framework screening rationale and the use of structural features to identify superionic conductors."}],"review_version":1}