{"id":"fa6b947a-d012-43f0-a0c2-d8bda4a7ec2c","arxiv_id":"2506.11199","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"Machine learning molecular dynamics on amorphous Ti-doped Li-P-S shows 10-20% Ti doping optimizes Li-ion transport through free-volume diffusion in disordered Li-S polyhedra.","lead":"This simulation study uses a machine-learned force field to run large-scale molecular dynamics on amorphous Ti-doped lithium phosphorus sulfide, a mixed ionic-electronic conductor for solid-state batteries. It reports that 10% and 20% Ti doping give faster lithium transport and more stable transport channels than 0% or 30%, and links this to disordered Li-S coordination environments.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The stability claim rests on an entropy-only free-energy argument in Sec. 3.5.2 that never compares enthalpy; Eq. 9 is a coordination-number Shannon entropy, not a demonstrated configurational entropy, so the 10/20% Ti stabilization conclusion is currently unsupported.","rationale":"The reader's conditional verdict and weakest-assumption analysis align with my reading. The main numerical results - MLFF accuracy, MSD-derived conductivities, activation energies, and the qualitative free-volume/coordination picture - are plausible and would hold up if the entropy section were corrected or reframed. The load-bearing flaw is in Section 3.5.2, where a Shannon entropy of coordination numbers is equated with configurational entropy and used to draw a thermodynamic stability conclusion without computing enthalpy. This is not an internal contradiction; it is a missing comparison that the central claim explicitly requires. Specifically, the paper's own text says 'the Gibbs free energy mostly depends on Sconfig' and 'When Sconfig increases, Delta Gconfig decreases' without ever evaluating H. Since the claim is about stability, not just disorder, the enthalpy term cannot be omitted. The proposed test - computing Delta H and comparing it with T Delta S across compositions - directly settles whether the 10/20% Ti preference survives. If it does, the claim becomes credible; if not, the manuscript should be revised to present the entropy analysis as a structural descriptor rather than a thermodynamic stabilization mechanism. Because the reader already identified exactly this issue and assigned conditional acceptance, my stress-test does not move the verdict; it reinforces the condition that the entropy/free-energy treatment must be corrected.","tokens_in":12187,"tokens_out":2653,"duration_ms":37143,"concrete_test":"Use the same MLFF (or DFT on representative snapshots) to compute the average enthalpy H per atom for LPS:Ti00%, Ti10%, Ti20%, and Ti30% at 300-500 K, along with S_vib from the VDOS and S_config from Eq. 9. Then evaluate Delta G = Delta H - T(S_vib + S_config) relative to the undoped or 30% case. If the ordering of Delta G does not place 10% and 20% Ti as most stable, or if |Delta H| is comparable to or larger than T|Delta S| with opposite sign, the entropy-only stability argument fails. As an additional check, estimate a many-body configurational entropy from a supercell enumeration or thermodynamic integration to test whether Eq. 9 is a valid proxy for S_config at all.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central stability claim depends on Section 3.5.2, but the argument there has two missing links. First, Eq. 9 computes a Shannon entropy of the one-body Li-S coordination-number distribution P(nc). Labeling this Sconfig and treating it as the configurational entropy appearing in the Gibbs free energy is not justified: the true configurational entropy of an amorphous material counts accessible microscopic configurations, including many-body correlations among coordination environments, not just the spread of a local coordination histogram. Second, the paper never computes enthalpy. The text states 'When Sconfig increases, Delta Gconfig decreases' and concludes that 10% and 20% Ti doping reduce Gibbs free energy, but Delta G = Delta H - T Delta S. A larger coordination-number entropy does not imply a lower free energy unless the enthalpy difference is smaller in magnitude or has the same sign. The paper also asserts that translational and electronic entropy contributions are negligible without computing them, and the vibrational entropy is estimated but the enthalpy term is entirely absent. Thus, even if the conductivity numbers and the free-volume hopping picture are correct, the specific claim that 10% and 20% Ti doping 'enhance the stability of transport channels' by configurational-entropy stabilization is not supported by the calculations presented.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript develops a DeePMD-based machine-learned force field (MLFF) for amorphous Li-Ti-P-S mixed ionic-electronic conductors, trained on AIMD trajectories at 500-900 K, and uses it to run 12,000-atom DLMD simulations at 300-500 K for compositions with 0%, 10%, 20%, and 30% TiS2. From these simulations the authors extract Li MSDs, diffusion coefficients, Nernst-Einstein ionic conductivities, Arrhenius activation energies, Li-S coordination-number distributions, and a Shannon-entropy-based 'configurational entropy.' They report that the computed conductivities and activation energies agree with their own experimental values, that Li transport occurs by free-volume diffusion through disordered Li-S polyhedra, and that 10% and 20% Ti doping stabilize the transport channels by increasing configurational entropy and lowering the Gibbs free energy. The paper concludes that 10-20% Ti doping is optimal for this MIEC.","tokens_in":12461,"tokens_out":5013,"duration_ms":56976,"significance":"If the results are fully supported, the paper would be a useful demonstration that a DeePMD force field trained on high-temperature AIMD data can be applied to large-scale MD simulations of an amorphous sulfide MIEC, and it would identify a doping window (10-20% Ti) for optimizing Li transport. The transport calculations use a standard and appropriate Arrhenius/Nernst-Einstein framework, and the direct comparison with the authors' prior experimental work [13] is a strength. The main conceptual contribution, however, is the thermodynamic stability argument based on coordination-number entropy, and this is currently not established. The paper would also be strengthened by quantitative error reporting and by validation of the MLFF in the production temperature range. The computational pipeline itself is potentially valuable, but the central stability claim needs substantial revision before the manuscript can be accepted.","major_comments":[{"comment":"The central stability conclusion is not supported by the entropy analysis as presented. Equation (9) defines S_config as the Shannon entropy of the single-particle Li-S coordination-number distribution P(n_c). For an amorphous network, the configurational entropy that enters the Gibbs free energy counts accessible microscopic configurations, including many-body correlations among coordination environments, and one cannot simply identify it with the spread of a local coordination histogram. Moreover, the text states 'When S_config increases, ΔG_config decreases' and concludes that 10% and 20% Ti doping 'reduce the Gibbs free energy,' but ΔG = ΔH - TΔS, and no enthalpy difference ΔH is computed anywhere in the paper. The assertions that translational and electronic entropy contributions are negligible are not computed; only Svib (Eq. 10) is estimated, and Svib is reported to be of order 1-3 m k_B, i.e., negligible compared with k_B, so it cannot compensate for the missing enthalpy term. Unless an enthalpy term is computed or the claim is reframed as a structural descriptor rather than a thermodynamic stability argument, the 10/20% stabilization conclusion in Sections 3.5.2 and 4 is unsupported.","section":"Section 3.5.2 (Eqs. 8-10)"},{"comment":"The MLFF is trained exclusively on AIMD trajectories at 500, 600, 700, 800, and 900 K, but all production DLMD simulations are performed at 300-500 K. The force/energy MAEs in Table 1 measure accuracy on the training distribution and do not establish accuracy in the extrapolation regime. A direct check of the MLFF against AIMD at 300-400 K, for example by comparing forces and energies on short AIMD trajectories or by running short AIMD simulations at low temperature and comparing with DLMD predictions, is needed before the quantitative conductivity and activation-energy values can be taken as reliable.","section":"Sections 2.2-2.3 and 3.1"},{"comment":"The quantitative comparison with experiment is presented without any statistical uncertainty. There are no error bars on σ(T), Ea, σ0, or the Shannon entropy, and the text does not state how many independent MD runs or trajectory blocks were used to estimate D. The reported composition ranking, with Ea = 0.30 eV for 10% Ti and 0.32 eV for 20% Ti versus higher values for 0% and 30% Ti, may lie within the statistical noise of a single 3 ns trajectory per state point. Block averaging or multiple independent simulations are required to support the claim that 10% and 20% Ti are optimum.","section":"Section 3.3, Fig. 4"},{"comment":"The free-volume diffusion mechanism is inferred primarily from visual inspection of the trajectory of a single Li ion per composition, without a quantitative free-volume analysis. The manuscript does not compute void-size distributions, hopping distances, or the correlation between Li displacements and local free-volume regions, and the conclusion in Fig. 5 that 10% and 20% Ti 'cover more area in 2D or volume in 3D' is not backed by a statistical or ensemble-averaged metric. A quantitative analysis of free volume and hopping statistics is needed to support the mechanistic claim.","section":"Section 3.4, Fig. 5"}],"minor_comments":[{"comment":"The text states that σ0 is determined from the y-intercept of a ln(σT) versus 1/T plot, but Eq. (4) is written as σ = σ0 exp(-Ea/kBT); if the plot uses σT, the pre-factor should be defined consistently, and the Arrhenius expression should be given in the same form.","section":"Section 3.3, Eq. (4)"},{"comment":"The caption uses the label '(a)' twice, once for the 2D ball-and-stick model and once for the LPS:Ti00% heatmap; the panel labels (b)-(e) also appear to be misaligned with the described composition order.","section":"Fig. 6 caption"},{"comment":"The word 'psudo-boundaries' should be 'pseudo-boundaries,' and the caption should clarify that the cubic boundaries are only guides to visualize a single Li-ion trajectory.","section":"Fig. 5 caption"},{"comment":"The sentence stating that configurational entropy 'explicitly refers to the degree of disorder in the Li-S polyhedra' is a definition introduced only after Eq. (9); it should appear before the entropy is computed.","section":"Section 3.5.2"},{"comment":"References [29] and [30] appear to be the same DeePMD-kit citation in two forms; one should be removed and the other consistently formatted.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The entropy-based thermodynamic stability argument in Section 3.5.2 is the main obstacle: as written, it does not support the central claim that 10% and 20% Ti doping stabilize transport channels. The issue is fixable by either computing enthalpy changes or explicitly reframing the coordination-number Shannon entropy as a structural descriptor rather than a free-energy driver. I would not recommend rejection, because the MLFF/MD pipeline and the transport calculations are valuable and the manuscript can be revised within its scope."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: the paper gives useful MLFF-MD transport numbers for a new amorphous MIEC, and the 10-20% Ti sweet spot for low activation energy is plausible. But the entropy-based stability claim — the part that turns a simulation study into a design rule — does not hold up as written.\n\nWhat is actually new: composition-dependent conductivities, activation energies, MSD analysis, and coordination statistics for Ti-doped LPS at 0/10/20/30% Ti. The MLFF is standard DeePMD, but the force errors are small and the 12,000-atom MD is a reasonable way to sample amorphous dynamics. The computed sigma and Ea trends track their own experimental data, which is a useful consistency check even if it is not independent validation.\n\nThe weak section is 3.5.2. Equation 9 is a Shannon entropy of the one-body Li-S coordination-number distribution. Labeling that 'configurational entropy' and then arguing Delta G decreases when this quantity increases skips the enthalpy term entirely. Delta G = Delta H - T Delta S; without estimating Delta H, a larger S does not imply a lower free energy. The assertion that vibrational, translational, and electronic entropies are negligible is also not substantiated beyond a single sentence. So the conclusion that 10% and 20% Ti doping 'enhance stability' is unsupported, even if the conductivities are correct. That said, this is a local flaw: the free-volume mechanism and the composition trend from Ea/sigma stand independently of the entropy argument.\n\nOther issues are more minor: no error bars on sigma, Ea, or entropy (at minimum block averaging over the 3 ns trajectory); no AIMD check of MLFF accuracy at 300-500 K when the training set covers 500-900 K; and the only experimental comparison is to their own prior paper. These are fixable or at least should be acknowledged.\n\nBottom line: worth a serious referee if the authors are willing to either fix the thermodynamics or cut the stability claim down to what the data actually show. It is a reasonable MLFF application paper, not a breakthrough.","headline":"Useful MLFF-MD transport data for Ti-doped LPS, but the entropy-based stability argument is not supported by the calculations.","tokens_in":13002,"tokens_out":2966,"would_cite":false,"duration_ms":33237,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Machine-learned molecular dynamics shows that lithium transport in amorphous Ti-doped Li-P-S proceeds by free-volume hopping, with 10–20% Ti doping giving the most stable transport channels.","keywords":["machine learning force field","molecular dynamics","lithium solid electrolyte","mixed ionic-electronic conductor","Li-Ti-P-S","ionic conductivity","configurational entropy","free-volume diffusion"],"falsifier":"A concrete check would be to compute, with the same force field or with DFT, the enthalpy and total entropy of amorphous Li-Ti-P-S at 0%, 10%, 20%, and 30% Ti; if 10–20% Ti does not give the lowest Gibbs free energy, or if the coordination-number entropy does not track the free-energy ordering, the channel-stability conclusion fails.","tokens_in":11958,"feed_emoji":"🔋","tokens_out":6946,"duration_ms":68254,"temperature":0.7,"pith_summary":"This paper argues that in an amorphous lithium titanium phosphorus sulfide mixed ionic-electronic conductor, lithium ions move by a free-volume diffusion mechanism: they hop through voids in the disordered structure rather than along fixed lattice pathways. Using a machine-learning force field trained on ab initio molecular dynamics, the authors simulate 12,000-atom cells at three titanium doping levels and six temperatures, and their computed ionic conductivities and activation energies match recent experiments. They further claim that 10% and 20% Ti doping produce the most stable transport channels, because those compositions show higher configurational entropy of disordered Li-S coordination polyhedra, which they equate with lower Gibbs free energy. If correct, the work identifies an optimal doping window for the material and demonstrates that large-scale machine-learning molecular dynamics can explain transport mechanisms in amorphous solid electrolytes.","feed_headline":"Lithium hops through free volume; 10–20% Ti best for sulfide MIEC","feed_subtitle":"Simulations with a 99%-accurate force field match experiments and explain why mid-level Ti doping stabilizes Li-S channels.","key_machinery":"The load-bearing machinery is a deep learning molecular dynamics workflow: a neural-network potential trained on ab initio MD data, then used in 3 ns NVT simulations of 12,000-atom amorphous cells at three Ti concentrations and six temperatures. Transport is quantified from mean-square displacement via the Nernst-Einstein relation, and channel stability is quantified from the Li-S coordination-number distribution $P(n_c)$ computed by k-nearest-neighbor counting. The identity that carries the stability argument is Eq. (9), $S_{\\mathrm{config}} = -k_B \\sum_{n_c} P(n_c)\\ln P(n_c)$, supplemented by a vibrational-entropy term; the paper argues that higher $S_{\\mathrm{config}}$ lowers the configurational Gibbs free energy and hence makes the 10% and 20% Ti channels more favorable.","core_discovery":"On the paper's own terms, the discovery is that amorphous lithium titanium phosphorus sulfide conducts lithium not through fixed crystalline pathways but through free-volume diffusion: individual Li ions hop among voids in the disordered structure, and the hopping medium is a set of disordered Li-S_n polyhedra with n ranging from 1 to 6. The authors report that a machine-learning force field trained on ab initio molecular dynamics reproduces experimentally measured ionic conductivities and activation energies across six temperatures, and that the lowest activation energy occurs at 10% Ti (0.3 eV), followed closely by 20% Ti (0.32 eV), while 0% and 30% Ti give higher barriers. They interpret the stability of the transport channels through coordination statistics: at 10% and 20% Ti, more than half of the Li atoms sit in four-coordinated S environments, whereas at 0% and 30% Ti the majority are three-coordinated. The accompanying configurational entropy of the coordination-number distribution is highest at 10% and next highest at 20% Ti, which the authors read as a decrease in Gibbs free energy and therefore a thermodynamically stabilized channel.","pith_inferences":["Editorial inference: the entropy-stability ranking could be checked by computing the enthalpy of each composition; if enthalpy differences are large enough to outweigh temperature times the configurational entropy, the claimed 10–20% window may not survive a full free-energy comparison.","Editorial inference: the coordination-entropy descriptor is cheap to compute from any trajectory and could serve as a screening metric for other amorphous solid electrolytes, not just Ti-doped LPS.","Editorial inference: because the paper does not compute electronic conductivity, the MIEC label rests on prior experiments; a simulation-based test of how Ti doping affects electron transport in the same 10–20% window would complete the picture."],"forward_implications":["If the central claim is right, the optimal Ti doping for this amorphous MIEC lies near 10–20%, where activation energy is lowest and the Li-S channel network is most stable; 30% Ti overdoping and 0% Ti both degrade transport.","Free-volume diffusion implies that amorphous disorder is a design lever: increasing the diversity of Li-S coordination environments should raise configurational entropy and improve ionic transport.","The machine-learning force field approach becomes a validated tool for screening other amorphous sulfide electrolytes at 12,000-atom scale, where direct ab initio molecular dynamics would be prohibitively expensive.","Computed conductivities and activation energies matching experiment strengthen confidence that the machine-learned force field captures the relevant physics, not just the training set."],"supporting_citations":[{"why":"Supplies the experimental ionic conductivity, activation energy, and mixed ionic-electronic behavior that the simulation results are compared against.","marker":"[13]"},{"why":"Establishes the prior machine-learning molecular dynamics approach to Li-ion transport that this study extends to Ti-doped LPS.","marker":"[14]"},{"why":"Provides the ab initio molecular dynamics method used to generate training data for the force field.","marker":"[20]"},{"why":"Companion method paper for the ab initio molecular dynamics simulations used in training-data generation.","marker":"[21]"},{"why":"Provides the deep neural network potential training package used to build the machine-learning force field.","marker":"[25]"},{"why":"Introduces the end-to-end symmetry-preserving potential energy model that defines the force field's architecture.","marker":"[27]"},{"why":"Supplies the Nernst-Einstein relation used to convert diffusion coefficients into ionic conductivities.","marker":"[34]"}],"fun_headline_variants":["Li hops via free volume; mid-Ti doping stabilizes sulfide MIEC","Amorphous Li-Ti-P-S: free-volume diffusion, 10-20% Ti best","ML force field reveals why 10-20% Ti stabilizes Li channels","Free-volume Li transport in Ti-doped LPS; 10-20% Ti optimal","Ti doping tunes Li-S coordination for stable MIEC conduction"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The stability claim rests on treating the Shannon entropy of the Li-S coordination-number distribution as the dominant term in the Gibbs free energy, with enthalpy and other entropy contributions neglected.","fun_headline_variants_meta":{"raw":{"variants":["Li hops via free volume; mid-Ti doping stabilizes sulfide MIEC","Amorphous Li-Ti-P-S: free-volume diffusion, 10-20% Ti best","ML force field reveals why 10-20% Ti stabilizes Li channels","Free-volume Li transport in Ti-doped LPS; 10-20% Ti optimal","Ti doping tunes Li-S coordination for stable MIEC conduction"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00019,"raw_usage":{"total_tokens":1396,"prompt_tokens":1055,"completion_tokens":341,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":671,"completion_tokens_details":{"reasoning_tokens":239}},"tokens_in":671,"tokens_out":341,"duration_ms":4070,"temperature":1.0,"reasoning_tokens":239,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T04:12:48.991479+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A concrete check would be to compute, with the same force field or with DFT, the enthalpy and total entropy of amorphous Li-Ti-P-S at 0%, 10%, 20%, and 30% Ti; if 10–20% Ti does not give the lowest Gibbs free energy, or if the coordination-number entropy does not track the free-energy ordering, the channel-stability conclusion fails.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the experimental ionic conductivity, activation energy, and mixed ionic-electronic behavior that the simulation results are compared against."},{"cited_title":"Selvaraj, V","cited_arxiv_id":null,"evidence_quote":"Establishes the prior machine-learning molecular dynamics approach to Li-ion transport that this study extends to Ti-doped LPS."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the ab initio molecular dynamics method used to generate training data for the force field."},{"cited_title":"Burke, J.P","cited_arxiv_id":null,"evidence_quote":"Provides the deep neural network potential training package used to build the machine-learning force field."},{"cited_title":"Abadi, A","cited_arxiv_id":null,"evidence_quote":"Introduces the end-to-end symmetry-preserving potential energy model that defines the force field's architecture."}],"review_version":1}