{"id":"ba95deef-9389-495a-8d67-ff2d3f59d32b","arxiv_id":"2506.21248","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"A machine-learning potential for bulk In2Se3 reveals previously unknown beta' polymorph variants, a lower-energy beta'' structure, and a strain-induced reversible beta' to beta'' transition.","lead":"The authors built a machine-learning potential for In2Se3 and used it to find new crystal structures and phase transitions, including a strain-driven switch between two polymorphs. A generalist reader might care because In2Se3 is a candidate material for ferroelectric memories and phase-change devices, and the paper suggests a new way to control its phases.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"MLP extrapolation error is unquantified in the exact regions used for the phase diagram and strain transition; the reported transition temperatures rest on finite-rate MD and an unrepresentative validation set.","rationale":"The static β″ structure ordering is explicitly checked against DFT (Supplementary Fig. 7 reports DFT-calculated total energies), so the 11 meV/f.u. Type-I/Type-II energy difference is not itself an MLP artifact; the reader's weakest_assumption is therefore slightly mislocated when it ties that specific number to MLP extrapolation. The genuinely load-bearing issue is the MLP's accuracy for the free-energy differences and transition-path barriers used to construct the phase diagram and the strain-induced transition. The manuscript itself documents the two key weaknesses: the validation set covers only a fraction of the training phase space, and the β–β′ boundary is obtained from direct MD heating/cooling runs rather than free-energy integration. No independent DFT check is provided for the free-energy differences at 300–700 K or for the 3–4% strained configurations in Fig. 6. The paper has independent support where DFT verification is provided: lattice parameters, phonon dispersions, STM simulations, and the static energy ordering of the new polymorphs. But none of this validates the MLP in the extrapolative regime where the phase diagram and strain predictions live. The reader's CONDITIONAL verdict is appropriate; requiring release of the MLP and training data, plus uncertainty estimates and an out-of-sample DFT check along transition paths, is the right remedy. I do not see grounds to reject, because the concern remains a missing validation rather than a demonstrated error.","tokens_in":24654,"tokens_out":10283,"duration_ms":130230,"concrete_test":"Run short DFT-MD simulations (or static DFT relaxations) on ~300 configurations sampled from the actual MD trajectories along the β–β′ and β′–β″ transition paths, including the 3–4% strained states of Fig. 6, and compare MLP energies, forces, and stresses against DFT, ensuring these configurations are withheld from the MLP training set. Then recompute the β–β′ phase boundary using thermodynamic integration between the two phases at 0 GPa and 300–500 K. If the held-out energy RMSE exceeds ~1 meV/atom, or if the equilibrium β–β′ transition temperature differs from the MD heating/cooling average by more than 50 K, the phase diagram and strain-transition claims are not supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim depends on the MLP being accurate for free-energy differences and transition barriers in configurations not represented by the validation set. Supplementary Fig. 1a states that the validation set covers only a fraction of the phase space of the training data, and the validation RMSE (0.61 meV/atom for energies, 0.039 eV/Å for forces) is lower than the training RMSE (1.17 meV/atom, 0.067 eV/Å), which is the opposite of what one expects if the test set were probing extrapolation. The phase diagram (Fig. 2c) is built from MLP thermodynamic integration/reversible scaling, but the β–β′ boundary is obtained as the mean of heating/cooling MD runs in 20 K steps with only 1 ns production (Methods), so it is a kinetic estimate, not an equilibrium free-energy boundary. Likewise, the strain-induced β′–β″ transition (Fig. 6) is observed at 3–4% strain under a 0.01 ns⁻¹ strain rate; no DFT check is reported for these strained, out-of-sample configurations. Because energy differences among the competing β′ variants are only a few meV/f.u., a modest MLP bias in this region would change the phase boundaries and the strain response without being visible in the reported validation metrics.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper develops a moment tensor machine-learning potential (MTP-MLP) for bulk van der Waals In2Se3, trained by active learning on 5,262 DFT structures, and validates it against lattice parameters, energy differences, energy-volume curves, and phonon dispersions. Using this MLP, the authors identify a Type-II structure for the beta'' polymorph that is 11 meV/f.u. lower in energy than the literature Type-I model, report numerous new beta' superstructure variants with dynamic stability, and interpret them through flat imaginary phonon bands and a Mexican-hat-like potential energy surface. Large-scale MLP molecular dynamics is used to observe alpha-to-beta, beta-to-beta', and beta'-to-beta'' transitions, to construct a temperature-pressure phase diagram by thermodynamic integration and reversible scaling supplemented by direct MD, and to predict a reversible strain-induced beta'-to-beta'' transition with accompanying polarization modulation. The manuscript concludes with proposals for strain-programmable electronic devices based on these transitions.","tokens_in":24945,"tokens_out":4310,"duration_ms":55824,"significance":"If the MLP is accurate in the regions where it is applied, the paper is significant: it unifies many experimentally observed polymorphs through a common Mexican-hat PES, provides a candidate corrected structure for the beta'' phase, reports an ab initio temperature-pressure phase diagram, and predicts a strain-driven reversible transition that could be exploited in devices. The paper has real strengths: multi-pronged validation against DFT (lattice parameters, energies, phonons, STM simulations), reproduction of experimentally known transitions, inclusion of additional InSe structures to improve transferability, and use of thermodynamic integration rather than direct fitting of transition temperatures. However, the load-bearing premise is extrapolative accuracy of the MLP in configurations not well represented by the validation set. The reported validation metrics characterize interpolation on a subset of the training manifold, and energy differences among competing variants are only a few meV/f.u.; without DFT checks of the specific new structures, strained states, and transition trajectories, the new phases and phase boundaries could be MLP artifacts.","major_comments":[{"comment":"The validation set is explicitly described as covering only a fraction of the phase space represented by the training dataset, and its RMSEs are lower than the training RMSEs. These metrics therefore demonstrate interpolation accuracy on the training manifold, not extrapolation accuracy to unsampled configurations. Since the central discoveries (the beta'' Type-II structure 11 meV/f.u. below Type-I, the new beta' variants with meV-scale energy differences, and the 3-4% strained states) live precisely in such unsampled regions, the paper should supply DFT calculations for these specific structures and for representative strained MD configurations, or an uncertainty metric such as the active-learning extrapolation grade along the MD trajectories. As written, the claim that these phases and transitions are not MLP artifacts is not fully supported.","section":"Development of MLP; Supplementary Fig. 1a"},{"comment":"The beta-beta' boundary in Fig. 2c is not obtained by thermodynamic integration or reversible scaling; the Methods state that this transition is determined from five heating and five cooling MD runs in 20 K steps with 100 ps equilibration and 1 ns production per step. This is a finite-rate kinetic estimate, not an equilibrium free-energy boundary, and the near-overlap of heating and cooling curves in Fig. 4a does not by itself remove hysteresis. Please report the spread over independent seeds and, ideally, compute the free-energy difference with an order-parameter-based thermodynamic integration scheme that can handle the interchangeable nature of the two phases.","section":"Methods, 'Thermodynamic phase diagram calculations'"},{"comment":"The quantitative agreement with experiment is weaker than the text suggests: the ambient-pressure alpha-to-beta transition temperature is computed as 683 K versus the cited experimental 473 K, and the later beta'-to-beta'' estimate is 277 K versus the cited 180 K. The phrase 'showing good agreement with experiments' should be qualified, and convergence checks (system size, heating rate, finite-size corrections) should be reported. If the overestimates are attributed to the PBE-D3 functional or to kinetic effects, this should be stated explicitly in the text.","section":"Thermodynamic phase diagram, Fig. 2c"},{"comment":"The strain-induced beta'-to-beta'' transition is simulated at a strain rate of 0.01 ns^-1 with critical strains of 3.81-4.02% in one direction and 0.59% in the reverse direction, and the associated polarization modulation is computed entirely from MLP trajectories. No DFT benchmark is reported for strained cells at these strains. Because the competing variants are separated by only a few meV/f.u., representative strained configurations at the endpoints and at the nucleation strain should be revalidated by DFT before the strain-driven transition and the polarization-switching claims are accepted as quantitative predictions.","section":"Dynamical polymorphic phase transitions, Fig. 6 and Supplementary Figs. 13-14"}],"minor_comments":[{"comment":"The phrase 'pays the way for the design' should read 'paves the way for the design'.","section":"Introduction"},{"comment":"The sentence describing the reverse transition says 'the reverse beta'' to beta'' phase transition'; this should read 'beta'' to beta' phase transition'.","section":"Dynamical polymorphic phase transitions, text accompanying Fig. 6"},{"comment":"There are small typographical errors in the text: 'T echnology' in the affiliation, 'MLlP' in the code availability section, and inconsistent spacing in 'V ASP'; these should be corrected.","section":"Affiliation and Code availability"},{"comment":"The heating rate is 0.039 K/ps and the cooling rate is 0.39 K/ps, differing by an order of magnitude; the main text should state both rates when quoting the averaged beta'-beta'' transition temperature of 277 K, since the asymmetry likely affects the averaged value.","section":"Supplementary Fig. 12 caption"},{"comment":"Fig. 2a uses absolute energies (approximately -19.3 eV/f.u.) while Fig. 2b uses enthalpies relative to alpha phase; for consistency, the captions should explicitly define the reference state for each panel.","section":"Fig. 2 caption and axes"}],"recommendation":"major_revision","confidential_remarks":"This is an ambitious and potentially impactful paper, but the strongest claims rest on MLP extrapolation that is not yet demonstrated. I do not see the concern as circularity: the transition temperatures are not direct fits to experiment, and the thermodynamic integration framework is appropriate. The main risk is that the MLP is silently inaccurate in the very regions used for the new phases and phase boundaries. I would request DFT verification of the key structures and strained states, plus a clearer separation of equilibrium free-energy boundaries from kinetic MD estimates, before publication."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear colleague,\n\nThis is a genuinely useful MLP study of In2Se3. The new stuff: a beta'' structure (Type-II) that is 11 meV/f.u. lower than the accepted Type-I and matches the zig-zag STM contrast better for one set of experiments; a family of beta' superstructure variants, several not reported before; and a strain-driven reversible beta' ↔ beta'' transition with polarization switching. The authors also reproduce the known α-β, β-β', and β'-β'' transformations in MD, and they give a thermodynamic T-P phase diagram from free-energy integration. That is a solid body of work, and the MLP appears to be well constructed: training includes InSe and split-structure motifs, and the validation covers lattice parameters, energy-volume curves, phonons, and STM images for key phases.\n\nThe soft spots are real but not fatal. The validation set covers only a subset of the training phase space, and the reported validation RMSE is actually lower than the training RMSE. That is the opposite of what you want to see if the test set is probing extrapolation. For the energy differences that matter here—a few meV/f.u. between competing beta' variants—a small MLP bias could shift the ordering without showing up in the global RMSE. Second, the β-β' boundary in the phase diagram is obtained from heating/cooling MD runs in 20 K steps with 1 ns production, i.e., a kinetic estimate, not an equilibrium free-energy boundary. The reported α-β transition at 683 K vs 473 K experiment and β'-β'' at 277 K vs 180 K are moderate agreements, not \"good.\" Third, the strain-induced transition is seen at 3-4% strain under a fast 0.01 ns^-1 strain rate, and no DFT check is reported for those strained configurations. Since this is the paper's most novel prediction, a few DFT single-point calculations on the strained intermediate states would go a long way.\n\nThe abstract overstates things: \"first-principles accuracy\" and \"ab initio temperature-pressure phase diagram\" are too strong, because the MLP is fitted to DFT and the phase boundaries carry the kinetic and fitting uncertainties above. That language should be tempered.\n\nBottom line: this paper deserves a serious referee. The new structures and the strain-switching mechanism are plausible and worth reporting, but the phase diagram and transition temperatures should be presented as semi-quantitative predictions until the MLP's extrapolation error in the relevant regions is quantified or artifacts are released. I would recommend acceptance with major revisions: release the MLP and training/validation data, add DFT checks on strained and out-of-sample configurations, and tone down the \"ab initio\" and \"good agreement\" phrasing.\n\nBest,\n[You]","headline":"Solid MLP study of In2Se3 with genuinely new polymorph structures and a strain-driven transition, but the phase diagram and transition temperatures rest on unquantified MLP extrapolation and kinetic MD, so they should be presented as semi-quantitative predictions rather than first-principles results.","tokens_in":25531,"tokens_out":2136,"would_cite":true,"duration_ms":25126,"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":"A machine-learning potential trained on 5,262 DFT structures maps In2Se3's polymorph landscape, resolving the β″ structure and the phase diagram.","keywords":["In2Se3","van der Waals layered materials","machine-learning potential","polymorphic phase transitions","ferroelectric–antiferroelectric","temperature–pressure phase diagram","molecular dynamics","Mexican-hat potential energy surface"],"falsifier":"Directly recompute with DFT the total energies of the proposed Type-II and Type-I β″ structures and the reported β′ superstructure variants using the same exchange-correlation functional; if Type-II is not lower by roughly 11 meV per formula unit, or if the β′ variants are not dynamically stable in the MLP phonon spectra, the claim fails. Experimentally, measure the β′↔β″ transition under controlled uniaxial strain in few-layer In2Se3 and compare the critical strains and polarization changes to the MD predictions.","tokens_in":24429,"feed_emoji":"⚛️","tokens_out":5594,"duration_ms":62996,"temperature":0.7,"pith_summary":"The paper claims that a machine-learning potential trained on 5,262 density-functional-theory structures can map the polymorphic energy landscape of van der Waals In2Se3 accurately enough to resolve an 11 meV per formula unit energy difference between competing β″ structures, uncover previously unknown β′ superstructure variants, and simulate phase transitions over nanosecond timescales. If true, this yields the first ab initio-quality temperature–pressure phase diagram for In2Se3 and a strain-based route to switch between antiferroelectric and nonpolar phases. The broader payoff would be atomic-scale understanding of how nucleation and domain-wall motion drive transitions among layered ferroic phases.","feed_headline":"One MLP maps In2Se3's polymorphs and phase diagram","feed_subtitle":"Simulations identify the lower-energy β″ structure and a reversible strain switch between polar phases.","key_machinery":"The central object is the moment tensor machine-learning potential (MLP), a fitted interatomic potential that reproduces DFT energies, forces, and stresses for In2Se3 and runs molecular dynamics on supercells of tens of thousands of atoms. The mechanism it exposes is the Mexican-hat-like potential energy surface of the β phase: the central Se atoms sit on a circular energy ridge with many shallow minima, so freezing the flat imaginary phonon branches at different wavevectors yields the family of β′ nanostripe superstructures and the β″ phase. Thermodynamic integration and reversible scaling turn MLP-mapped free energies into the temperature–pressure phase diagram.","core_discovery":"The central discovery is that In2Se3's β polymorph is a high-symmetry parent phase whose central Se atoms sit on a Mexican-hat-like potential energy surface, so freezing its flat imaginary phonon modes at different wavevectors generates the family of β′ nanostripe superstructures and the β″ phase. The paper identifies a β″-1T structure (Type-II) that is 11 meV per formula unit lower in energy than the previously proposed Type-I structure, supports this with simulated STM images that better match one set of experiments, and reports several previously unreported β′ superstructure variants across the 1T, 2H, and 3R polytypes. Using MLP-driven molecular dynamics, it observes α→β→β′→β″ transformations, constructs a temperature–pressure phase diagram via thermodynamic integration, and finds reversible β′↔β″ transitions under uniaxial strain with accompanying polarization changes.","pith_inferences":["If the MLP extrapolates as claimed, other layered III2-VI3 compounds with flat imaginary phonon bands and Mexican-hat-like surfaces should show similarly large families of closely spaced metastable polar and antiferroelectric superstructures.","The strain-induced β′↔β″ transition suggests a mechanical-writing scheme for ferroelectric memories without applied electric fields; the paper leaves device implementation unexplored.","Because the validation set covers only a fraction of the phase space represented by the training dataset, an independent check of transition barriers and nucleation rates (for example, comparing MD-derived kinetics against experimental hysteresis) would be needed before relying on the predicted transition temperatures quantitatively.","The coexistence of Type-I and Type-II β″ structures in different STM experiments hints that both may appear depending on growth or quench history, a hypothesis testable by controlled cooling-rate experiments."],"forward_implications":["The MLP places the β″-2H, β″-3R, and β″-1T Type-II structures as dynamically stable phases, with Type-II 11 meV/f.u. more stable than the previously proposed model.","Several β′ nanostripe superstructures (2d-2d, 3d-3d, 4d-4d, 5d-5d, and mixed variants) are dynamically stable and nearly degenerate in energy, explaining observed coexistence of ferroelectric and antiferroelectric contrast in experiments.","The computed temperature–pressure phase diagram gives α stable at low temperature and pressure, β at high temperature and pressure, and β′ in between, with the α→β boundary at about 683 K at ambient pressure and about 0.9 GPa at 300 K, both consistent with experimental data.","Molecular dynamics shows that all these phase transitions proceed by in-plane nucleation followed by layer-by-layer growth along the c axis, including the formation and motion of 0°, 60°, and 120° domain walls.","Uniaxial strain reversibly converts β′ to β″ (a-axis strains of 3.81–4.02%) and back (b-axis strains near 0.59%), with ferroelectric polarization appearing and disappearing during the transitions.","The strain-induced β′↔β″ transition decouples phase switching from electric fields, pointing toward mechanically programmable ferroelectric devices."],"supporting_citations":[{"why":"Defines moment tensor potentials, the MLP class used in this work.","marker":"[76]"},{"why":"Provides the active learning scheme that selects which structures enter the training set.","marker":"[77]"},{"why":"Supplies the MLIP package used for active learning and moment tensor potential fitting.","marker":"[78]"},{"why":"Establishes the Mexican-hat potential energy surface for central Se atoms in In2Se3 monolayers, the basis for the polymorph landscape.","marker":"[67]"},{"why":"Predicts a Star-of-David CDW phase with energy comparable to β″, motivating exploration of competing metastable phases.","marker":"[66]"},{"why":"Provides the experimental β″ structure and reversible β′↔β″ transition that the paper's Type-II structure improves upon.","marker":"[30]"},{"why":"Documents β′ ferroelasticity and domain walls, the experimental baseline for the MD simulations.","marker":"[12]"},{"why":"Supplies experimental transition temperatures and β′ superstructure periodic lengths used to validate the phase diagram.","marker":"[26]"},{"why":"Provides the experimental pressure-induced α→β transformation point used to benchmark the computed phase boundary.","marker":"[42]"},{"why":"Proposes the intralayer splitting mechanism for α→β transitions; the paper includes these structures in training and tests them with MD.","marker":"[40, 41]"}],"fun_headline_variants":["MLP reveals In2Se3 phase diagram","In2Se3's lower-energy β'' phase revealed by MLP","Strain reversibly flips In2Se3 between β′ and β''","MLP-driven MD maps In2Se3 polymorphs and phase transitions"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The machine-learning potential remains accurate outside the configuration space it was trained on, down to energy differences of about 11 meV per formula unit and the free-energy differences that set transition temperatures.","fun_headline_variants_meta":{"raw":{"variants":["MLP reveals In2Se3 phase diagram","In2Se3's lower-energy β'' phase revealed by MLP","Strain reversibly flips In2Se3 between β′ and β''","MLP-driven MD maps In2Se3 polymorphs and phase transitions"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001133,"raw_usage":{"total_tokens":4744,"prompt_tokens":1016,"completion_tokens":3728,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":632,"completion_tokens_details":{"reasoning_tokens":3652}},"tokens_in":632,"tokens_out":3728,"duration_ms":30904,"temperature":1.0,"reasoning_tokens":3652,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T22:29:29.990589+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Directly recompute with DFT the total energies of the proposed Type-II and Type-I β″ structures and the reported β′ superstructure variants using the same exchange-correlation functional; if Type-II is not lower by roughly 11 meV per formula unit, or if the β′ variants are not dynamically stable in the MLP phonon spectra, the claim fails. Experimentally, measure the β′↔β″ transition under controlled uniaxial strain in few-layer In2Se3 and compare the critical strains and polarization changes to the MD predictions.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines moment tensor potentials, the MLP class used in this work."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the active learning scheme that selects which structures enter the training set."},{"cited_title":"S., Gubaev , K., Podryabinkin, E","cited_arxiv_id":null,"evidence_quote":"Supplies the MLIP package used for active learning and moment tensor potential fitting."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Establishes the Mexican-hat potential energy surface for central Se atoms in In2Se3 monolayers, the basis for the polymorph landscape."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Predicts a Star-of-David CDW phase with energy comparable to β″, motivating exploration of competing metastable phases."},{"cited_title":"M., Teklemichael, S","cited_arxiv_id":null,"evidence_quote":"Provides the experimental pressure-induced α→β transformation point used to benchmark the computed phase boundary."}],"review_version":1}