{"id":"f1ad1584-ee62-4719-a381-180404e8256d","arxiv_id":"2412.05332","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"Neural quantum state simulations indicate that two-dimensional helium-4 freezes via a first-order transition, with no hexatic or supersolid phase surviving in larger systems.","lead":"Using neural quantum states, the authors map the zero-temperature phase diagram of two-dimensional helium-4 and find a first-order liquid-solid transition. They also find that the Rényi-2 entanglement entropy drops sharply on freezing, and that an apparent intermediate hexatic/supersolid phase seen in small simulations disappears in larger ones.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The Maxwell construction that establishes first-order melting relies on N=80 liquid/solid branches obtained by transfer learning in a translation-invariant NQS ansatz, while Appendix C shows a 12-14 mK variational bias specific to the N=80 solid and no independent check excludes that the branches…","rationale":"The reader's weakest assumption targets the same load-bearing premise: whether the presence (N=80) and absence (N=30) of metastable branches reflect the true free-energy landscape rather than variational bias in the translation-invariant NQS ansatz. I agree with that assessment. My stress-test sharpens it with a quantitative observation from the paper's own benchmark table: the NQS solid branch at N=80 is 12-14 mK above DMC while the liquid branch is essentially exact, so the Maxwell construction is built on a branch whose variational error is phase-dependent and largest exactly where the common tangent is anchored. A reverse-transfer and pinning-removal protocol would directly test whether the branches are physical or artifacts. I do not see a reason to change the reader's CONDITIONAL verdict: the paper is honest about finite-size ambiguity, reports DMC benchmarks for many but not all coexistence points, and does not provide code or data, so the central claim is plausible but not fully secured.","tokens_in":22607,"tokens_out":7945,"duration_ms":89450,"concrete_test":"Run the fixed-pressure VMC protocol at N=30 initialized from the converged N=80 solid branch at P≈0.50 KÅ^-2 (reverse transfer learning) and independently re-derive the N=80 branches by starting each coexistence pressure from a Gaussian-pinned solid trial with the pinning removed after convergence, then compare density and S(G) hysteresis loops against the main-text branches. If the N=30 solid-seeded run melts and the N=80 branches reproduce without transfer learning, the first-order scenario is supported; if not, the branches are optimization artifacts.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim of a first-order superfluid-to-solid transition rests on the Maxwell construction over metastable liquid and solid branches observed for N=80 (Appendix D, Eq. D1-D2), while N=30 shows no such branches and instead a continuous crossover with hexatic order and a condensate fraction. For the claim to hold, the N=80 branches must faithfully represent the true metastable free-energy branches, and the N=30 absence of branches must be a genuine finite-size effect. Two facts make this insecure. First, the N=80 branches are generated by transfer learning from deeper liquid/solid densities (Appendix D), so the optimizer is explicitly guided to preserve the phase; there is no independent check, such as bidirectional sweeps or unbiased initialization, that both branches are actual variational minima at the same pressure. Second, Appendix C, Table I shows the NQS ansatz is less accurate for the N=80 solid: at n=0.070 and 0.075 Å^-2 the NQS energies are 0.518(3) and 1.085(3) K versus DMC 0.504(2) and 1.073(1) K, a 12-14 mK upward bias, while the liquid branches agree within 1-3 mK. This phase-dependent variational bias is largest on the solid branch that anchors the high-density side of the common tangent, so the Maxwell construction could shift or close. The N=30 absence of metastability could therefore be either a finite-size effect or an optimizer artifact. The first-order conclusion is not yet independently supported at the accuracy claimed.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies the zero-temperature phase diagram of two-dimensional helium-4 using a translation-invariant neural quantum state (NQS) in both canonical and isobaric variational Monte Carlo. The authors report a first-order superfluid-to-solid transition, with freezing and melting densities and a critical pressure, and argue that the apparent continuous crossover with hexatic order and a condensate fraction seen for N=30 is a finite-size artifact, since N=80 simulations exhibit metastable liquid and solid branches that support the Maxwell construction. They also compute the Rényi-2 entanglement entropy and observe a sharp drop at freezing. The main evidence is the constant-slope region in E(n), the abrupt density jump in isobaric runs, and the N=80 metastable branches; NQS energies are benchmarked against DMC in Table I.","tokens_in":22954,"tokens_out":12513,"duration_ms":118742,"significance":"If the conclusions hold, this is a valuable methodological and physical contribution: a single symmetric NQS functional form describes both liquid and solid without imposing symmetry, fixed-pressure NQS optimization allows cell-shape relaxation, and the paper makes concrete, falsifiable predictions for the coexistence densities and melting pressure. The DMC benchmarks and the variance reductions reported in Table I are strengths, as is the improved Rényi-2 estimator with random-origin averaging. The significance is tempered because the quantitative coexistence interval and the exclusion of a hexatic phase rest on N=80 metastable branches that are obtained by transfer learning and are not independently validated against the known phase-dependent variational bias; additional checks are needed before the phase diagram can be considered quantitative.","major_comments":[{"comment":"The Maxwell construction in Appendix D uses spline fits to the N=80 liquid and solid branches, but Table I shows that the NQS energies on the N=80 solid branch are systematically above DMC by 14(3) mK at n=0.070 Å^-2 and 12(3) mK at n=0.075 Å^-2, while the N=80 liquid entries agree with DMC to within 1-3 mK at n=0.065 and 0.068 Å^-2. Because the solid branch anchors the high-density end of the common tangent, a constant upward bias of about 13 mK in the solid branch changes Pc by roughly δ/(n_m^-1 - n_f^-1) ≈ 0.018 KÅ^-2, which is comparable to the quoted 0.01 KÅ^-2 uncertainty and to the gap between the N=30 and N=80 pressure estimates. Please provide a bias-corrected Maxwell construction, for example by shifting the solid branch to the DMC energies, or otherwise demonstrate that the quoted coexistence interval and Pc are stable against this phase-dependent variational bias.","section":"Appendix C, Table I"},{"comment":"The N=80 liquid and solid branches are generated by transfer learning from variational parameters optimized at deeper densities of the same phase, and no bidirectional initialization or unbiased optimization from the opposite phase is reported. In a translation-invariant ansatz, a solid is represented only if the Monte Carlo chain remains in a crystalline basin, so seeding from a solid-optimized state favors the solid branch by construction, and similarly for the liquid branch. This matters because the central finite-size argument for first-order melting and against a hexatic phase is precisely that N=80 exhibits two metastable branches while N=30 does not. Please demonstrate that each branch is a genuine local minimum at the relevant densities, for example by performing bidirectional sweeps (liquid initialization at densities above n_f and solid initialization below n_m) or by reporting order-parameter histograms and convergence from both initialization directions.","section":"Appendix D, Fig. 7"},{"comment":"The paper reports materially different values for the critical pressure: 0.485(5) KÅ^-2 for N=30 in the isobaric ensemble (Fig. 1b), 0.50(1) KÅ^-2 for N=30 from canonical data (Appendix E), 0.53(1) KÅ^-2 for N=80 from the Maxwell construction, and 0.54(1) KÅ^-2 after finite-size extrapolation. The spread of about 0.04-0.05 KÅ^-2 is several times the quoted error bars. Since Pc and the coexistence interval are central quantitative predictions, please reconcile these values, state whether the discrepancy is due to finite-size effects, the ensemble, or the potential truncation, and report the N=30 and N=80 coexistence densities with the same definition. The statement in the Fig. 1 caption that the uncertainty on Pc is the pressure grid size appears to account only for grid resolution, not for statistical or variational uncertainty.","section":"Main text and Appendices D/E"}],"minor_comments":[{"comment":"Please state explicitly which potential form, the full periodized version of Eq. (A4) or the shifted and truncated version of Eq. (A8), is used for each figure and for the isobaric simulations; Table I and Fig. 8 clearly use the shifted and truncated potential, while the main-text equation of state does not specify this. This is necessary to reproduce the Pc values.","section":"Appendix A and main text"},{"comment":"The legends of Figs. 2 and 4 include N=56 data, but the text never discusses the N=56 results. Please state what the N=56 calculations add and whether they support the finite-size trend between N=30 and N=80.","section":"Figs. 2 and 4"},{"comment":"The 1/x fit f(x)=c1+c2(1/x+1/(min(L)-x)) is used to distinguish liquid from solid order, but the fitted values and uncertainties of c1 are not reported; giving them would make the liquid/solid distinction quantitative.","section":"Fig. 2"},{"comment":"Please clarify the notation in the second exponential factor of Eq. (2): it is not clear whether the intended form is exp(-θ2/d_sin(ri,rj)^5) or exp(-θ2 d_sin(ri,rj)^5), and the definition of d_sin should be stated in the main text where Eq. (2) appears rather than only in Appendix B.","section":"Eq. (2)"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is potentially suitable for publication after a revision that addresses the validation of the N=80 metastable branches. The main risk is that the quantitative coexistence interval and the reported critical pressure depend on branches with a known variational bias, and the current transfer-learning protocol does not by itself establish that those branches are true variational minima. I would encourage the editor to require the bias-corrected Maxwell construction and bidirectional initialization checks before acceptance. There is no circularity issue: the phase diagram is obtained by energy minimization of a standard Hamiltonian and benchmarked against DMC."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Bottom line: this is a serious paper and the first neural quantum state treatment of 2D helium-4 that uses one symmetry-unbiased ansatz for both liquid and solid. The fixed-pressure isobaric VMC machinery, with cell geometry as variational parameters, is a real methodological step. But the headline claim — a first-order superfluid-to-solid transition with no hexatic or supersolid intermediate phase in the thermodynamic limit — is not yet as secure as the abstract suggests. I would send it to referees, not desk reject it.\n\nThe paper does several things well. The NQS energies are benchmarked against DMC in Table I and, in the liquid, agree within a few mK; the per-particle energy variance is orders of magnitude below the U2/U3 VMC wavefunctions. Using the same functional form for both phases is exactly the right way to avoid the symmetry bias that plagued earlier DMC studies with explicit crystal or hexatic trial functions. The authors also report the N=30 continuous crossover with hexatic signatures and a retained condensate fraction honestly, and interpret it as a finite-size effect rather than hiding it. The Rényi-2 estimator with a truncated swap and random-origin averaging is a genuine technical contribution.\n\nThe soft spots are concentrated where the phase diagram is built. The N=80 metastable liquid and solid branches come from transfer learning, with no independent check — no unbiased initialization, no bidirectional sweeps — that both are actual variational minima at the same pressure. For a translation-invariant ansatz, that leaves room for the optimizer to create or suppress branches. Appendix C makes this concrete: on the solid side, N=80 NQS sits 12–14 mK above DMC, while the liquid is within 1–3 mK; at N=30 and n=0.070 Å^-2, NQS is actually 11 mK below the quoted DMC. A variational bias that is phase-dependent and largest exactly on the solid branch can shift the common tangent. The N=30 critical pressure is 0.485(5) KÅ^-2, while N=80 and the extrapolated value are 0.53(1) and 0.54(1); that gap is not really reconciled by \"agrees within resolution.\" There are also no error bars on S(G), G6, or condensate fractions, and no code or data, which makes the absence-of-hexatic claim hard to falsify.\n\nI think the first-order scenario is probably right — the constant-slope E(n), the isobaric density jump, and the sharp entanglement drop all point that way. For people working on NQS methods or on 2D quantum melting, this is worth reading. It deserves a serious referee, and the review should ask for error bars on order parameters, code/data, an independent study of the N=80 branches, and a clearer account of the Pc discrepancy.","headline":"A genuinely new isobaric NQS method for 2D helium, with an honest finite-size discussion, but the quantitative phase boundary needs independent branch checks and error bars before the first-order claim is taken at face value.","tokens_in":23509,"tokens_out":5177,"would_cite":true,"duration_ms":53456,"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":"Two-dimensional helium-4 freezes abruptly at zero temperature, with no hexatic phase in the thermodynamic limit.","keywords":["helium-4","two-dimensional quantum solid","neural quantum states","first-order phase transition","hexatic phase","Bose-Einstein condensation","Rényi-2 entanglement entropy","isobaric variational Monte Carlo"],"falsifier":"Perform the same isobaric neural quantum state simulation with an explicitly symmetry-broken ansatz, such as adding Gaussian pinning to triangular lattice sites, for $N=30$ in the coexistence region. If the solid branch becomes lower in enthalpy than the continuous crossover state, the $N=30$ hexatic and supersolid-like signatures are variational artifacts; if the continuous branch remains lower, the finite-size explanation is wrong. A complementary check is to compare neural quantum state energies at densities $n=0.068$ to $0.072\\ \\mathrm{\\AA}^{-2}$ against diffusion Monte Carlo energies obtained with crystal-pinned trial functions, since a lower solid energy in that window would shift the claimed coexistence pressure.","tokens_in":22393,"feed_emoji":"🧊","tokens_out":6145,"duration_ms":61326,"temperature":0.7,"pith_summary":"The paper maps the zero-temperature phase diagram of two-dimensional helium-4 using a single translation- and permutation-invariant neural quantum state for both liquid and solid. In fixed-pressure calculations the system shows an abrupt density jump at a critical pressure $P_c \\approx 0.485(5)\\ \\mathrm{K}\\,\\mathrm{\\AA}^{-2}$ for $N=30$ and $0.53(1)\\ \\mathrm{K}\\,\\mathrm{\\AA}^{-2}$ for $N=80$ from a Maxwell construction, identifying a first-order liquid-solid transition. Smaller cells ($N=30$) instead cross over continuously through a region with hexatic order and a small condensate fraction, but larger systems ($N=80$) follow metastable liquid and solid branches, indicating the intermediate phase is a finite-size artifact. The paper concludes that in the thermodynamic limit the transition is first order, from a superfluid to a normal non-Bose-condensed quantum solid, with no hexatic phase.","feed_headline":"Neural quantum states show 2D helium-4 freezes abruptly","feed_subtitle":"Fixed-pressure calculations put the first-order freezing line near $0.53\\ \\mathrm{K}/\\mathrm{\\AA}^2$ and rule out a hexatic phase at zero…","key_machinery":"The load-bearing object is the neural quantum state trial wavefunction $\\Psi_\\theta(R)=\\prod_{i<j}\\exp[\\theta_1\\phi(e^{(K)}_{ij})]\\exp[-\\theta_2/d_{\\mathrm{sin}}(r_i,r_j)^5]$, in which McMillan-type short-range pair correlations are multiplied by a graph-neural-network backflow of message-passed edge features, making the ansatz permutation- and translation-invariant with no baked-in crystal symmetry. Because the same functional form must represent both phases, the optimization, not an imposed trial structure, decides whether the state is liquid or solid. The isobaric ensemble treats box lengths and angle as variational parameters through a unit-coordinate affine transformation, minimizing $G=H+PV$, and the Maxwell construction on $N=80$ metastable branches fixes the coexistence line; hexatic order is probed with $g_6(r)$ and positional order with the structure factor near Bragg peaks, while condensate fraction and Rényi-2 entropy are computed from Monte Carlo estimators.","core_discovery":"The central claim is that at zero temperature, two-dimensional helium-4 with the Aziz HFDHE2 pair potential undergoes a first-order phase transition from a superfluid liquid to a normal crystalline solid, and the apparent intermediate hexatic or supersolid-like behavior seen in small simulation cells is a finite-size effect. This is established by using the same neural quantum state ansatz without explicit symmetry breaking in fixed-pressure variational Monte Carlo; the enthalpy as a function of density develops a constant-pressure coexistence region, and the Maxwell construction gives freezing and melting densities $n_f \\approx 0.0673\\ \\mathrm{\\AA}^{-2}$ and $n_m \\approx 0.0698\\ \\mathrm{\\AA}^{-2}$ (for $N=30$) with $P_c \\approx 0.485(5)\\ \\mathrm{K}\\,\\mathrm{\\AA}^{-2}$. For $N=80$ the liquid and solid branches are metastable across the coexistence region, and the hexatic correlation function decays exponentially in the liquid and saturates in the solid without algebraic decay, arguing against an intervening hexatic phase. The condensate fraction drops sharply on freezing, and the Rényi-2 entanglement entropy decreases abruptly, providing a zero-temperature analog of the entropy drop at a first-order transition.","pith_inferences":["The $N=30$ continuous crossover could be interpreted as effective rounding of a first-order transition in a small box; one could test whether the crossover sharpens with additional variational freedom, which would confirm the finite-size reading.","The same Gibbs-ensemble setup could measure the free-energy barrier between liquid and solid and estimate critical nucleus sizes, quantities the paper does not compute.","A zero-temperature analog of the Kosterlitz-Thouless-Halperin-Nelson-Young scenario would predict algebraic hexatic order; the authors' data rule it out at $T=0$, but a finite-temperature extension of the same method could map the hexatic region above zero temperature.","Entanglement entropy may serve as a sensitive order parameter for first-order quantum transitions in other bosonic or fermionic crystals."],"forward_implications":["The zero-temperature freezing line in two-dimensional helium-4 is first order, with no hexatic phase in the thermodynamic limit.","Small-cell simulations that show smooth melting, hexatic order, or a large condensate fraction near freezing should not be read as bulk supersolidity.","The Rényi-2 entanglement entropy drops sharply at freezing, giving a zero-temperature signature of the first-order transition analogous to the entropy of fusion.","The isobaric neural quantum state method, with cell geometry as variational parameters, can be applied to other quantum crystals or electronic structure problems without imposing lattice symmetry.","Finite-size corrections scale as $N^{-1/2}$ for Bragg peak amplitudes and $N^{-3/2}$ for energies, so thermodynamic-limit extrapolation keeps the critical pressure near $0.53$ to $0.54\\ \\mathrm{K}\\,\\mathrm{\\AA}^{-2}$."],"supporting_citations":[{"why":"Defines the Aziz HFDHE2 pair potential that models the helium-4 interactions throughout the paper; all phase-diagram results depend on this interaction.","marker":"[49]"},{"why":"Earlier path-integral Monte Carlo study of two-dimensional helium-4; supplies the previous freezing/melting densities and critical pressures against which the new results are compared.","marker":"[43]"},{"why":"Introduces neural quantum states, the variational framework on which the trial wavefunction is built.","marker":"[38]"},{"why":"Provides the message-passing neural-network backflow ansatz that lets the same functional form describe both liquid and solid without explicit symmetry breaking.","marker":"[51]"},{"why":"Supplies the McMillan short-range Jastrow factor used in the trial wavefunction and the Monte Carlo estimator for the condensate fraction.","marker":"[50]"},{"why":"Gives the finite-size corrections for energy and condensate density used to extrapolate thermodynamic-limit quantities.","marker":"[63]"},{"why":"Earlier diffusion Monte Carlo study of two-dimensional helium-4 whose trial function explicitly imposed hexatic or crystal symmetry, motivating the need for an unbiased ansatz.","marker":"[37]"}],"fun_headline_variants":["2D helium-4 freezes in first-order transition","Neural quantum states pin down 2D helium-4 freezing","No hexatic phase in 2D helium-4, study finds","2D helium-4 freezing is first-order, not hexatic","Fixed-pressure simulations reveal abrupt 2D freezing"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The argument assumes the variational neural-network state is flexible enough and optimized well enough that the absence of metastable branches in $N=30$ cells reflects the true free-energy landscape rather than the optimizer failing to find the solid.","fun_headline_variants_meta":{"raw":{"variants":["2D helium-4 freezes in first-order transition","Neural quantum states pin down 2D helium-4 freezing","No hexatic phase in 2D helium-4, study finds","2D helium-4 freezing is first-order, not hexatic","Fixed-pressure simulations reveal abrupt 2D freezing"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000204,"raw_usage":{"total_tokens":1396,"prompt_tokens":961,"completion_tokens":435,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":577,"completion_tokens_details":{"reasoning_tokens":349}},"tokens_in":577,"tokens_out":435,"duration_ms":3671,"temperature":1.0,"reasoning_tokens":349,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T21:32:18.567793+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Perform the same isobaric neural quantum state simulation with an explicitly symmetry-broken ansatz, such as adding Gaussian pinning to triangular lattice sites, for $N=30$ in the coexistence region. If the solid branch becomes lower in enthalpy than the continuous crossover state, the $N=30$ hexatic and supersolid-like signatures are variational artifacts; if the continuous branch remains lower, the finite-size explanation is wrong. A complementary check is to compare neural quantum state energies at densities $n=0.068$ to $0.072\\ \\mathrm{\\AA}^{-2}$ against diffusion Monte Carlo energies obtained with crystal-pinned trial functions, since a lower solid energy in that window would shift the claimed coexistence pressure.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines the Aziz HFDHE2 pair potential that models the helium-4 interactions throughout the paper; all phase-diagram results depend on this interaction."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Earlier path-integral Monte Carlo study of two-dimensional helium-4; supplies the previous freezing/melting densities and critical pressures against which the new results are compared."},{"cited_title":"Pescia, J","cited_arxiv_id":null,"evidence_quote":"Provides the message-passing neural-network backflow ansatz that lets the same functional form describe both liquid and solid without explicit symmetry breaking."},{"cited_title":"Boninsegni and S","cited_arxiv_id":null,"evidence_quote":"Supplies the McMillan short-range Jastrow factor used in the trial wavefunction and the Monte Carlo estimator for the condensate fraction."},{"cited_title":"Apaja and M","cited_arxiv_id":null,"evidence_quote":"Earlier diffusion Monte Carlo study of two-dimensional helium-4 whose trial function explicitly imposed hexatic or crystal symmetry, motivating the need for an unbiased ansatz."}],"review_version":1}