{"id":"f903cd36-e8ee-4f6e-9457-998eaa907940","arxiv_id":"2412.14953","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"Finite-size scaling of a machine-learned potential trained on PBE density functional theory shows hydrogen's liquid-liquid transition is genuinely first-order, with a critical point at 1250 ± 50 K and 155 to 160 GPa.","lead":"By simulating hydrogen under extreme pressure with a fast machine-learned model trained on quantum chemistry data, this study finds that the switch between two liquid forms of hydrogen is a true sharp phase transition, not a smooth crossover. It places the endpoint of that transition near 1200 to 1300 K and 155 to 160 GPa, lower than earlier estimates and close to the melting line, which matters for understanding giant planet interiors.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The first-order proof rests on unquantified tanh-fit extrapolations in Fig. 7; bootstrap uncertainties on αD and its 1/N intercept are needed before the finite intercept at 1125 K can be trusted.","rationale":"I read the strongest claim as a claim about the model PBE-hydrogen, not about real hydrogen: classical protons with PBE undergo a first-order LLPT whose critical point is at 1250±50 K, 155-160 GPa. The paper's own App. C documents real MLIP biases, so the reader's conditional verdict is reasonable. I do not think the most load-bearing weakness is the MLIP bias itself, because the 2048-atom AIMD comparison at 200 GPa (Fig. 12) and the 200-atom EOS comparison (Fig. 2) give genuine, independent support that the MLIP is in the right qualitative regime. The sharper weakness is the inferential step that converts those simulations into a proof of first-order character: Fig. 7 has no error bars, only four system sizes, and is derived from fits whose input data are acknowledged to be in a regime of large fluctuations and rare transitions. The word 'proving' in the paper is stronger than what the figure can support. This is a fixable statistical issue, not a fundamental one, so I keep the reader's CONDITIONAL verdict. The concrete bootstrap test would settle it directly; if the intercept uncertainty includes zero, the conclusion should be downgraded until more sampling or a stronger method (e.g., free-energy barriers) is supplied.","tokens_in":16827,"tokens_out":8808,"duration_ms":82486,"concrete_test":"Block-bootstrap the density time series for each state point in Fig. 6 (block length >= 20 ps to exceed autocorrelation), refit Eq. (1) on each resample, and report the distribution of αD and of the linear-in-1/N intercept of αD/N at T=1125, 1200, 1300, and 1500 K. If the 95% confidence interval for the 1125 K intercept includes zero, the first-order claim is not supported by the presented data; if it excludes zero, the claim gains the missing statistical backing. As a secondary model-bias check, retrain an independent MACE or DeePMD potential on the same 96-atom data and repeat the FSS at 1125 K; agreement in the intercept would show the order is not specific to the NequIP fit.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that PBE-hydrogen has a true first-order LLPT at 1125 K and a critical point at 1250±50 K is carried by Fig. 7, the 1/N extrapolation of αD obtained by fitting Eq. (1) to NPT density-pressure data. The paper quotes no uncertainties on α, D, or the fitted intercept, yet the supporting runs are in a regime the authors themselves call poorly estimated: only five density switches in 200 ps for N=2048, and density fluctuations of the same order as the 2-3% order-parameter jump for N<1000. With four system sizes (400-2048) and five free parameters per fit, α and D are strongly correlated; the 'clearly finite' intercept could be an artifact of the tanh functional form and unweighted least-squares fitting. If the intercept is consistent with zero within bootstrap error, the first-order conclusion is not established. The MLIP bias flagged by the reader (4 GPa EOS offset, 1.46× quadrupole correlation) would shift where the transition sits, but only after this statistical premise is secured can model fidelity be assessed.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies the liquid-liquid phase transition in hydrogen modeled with the PBE exchange-correlation functional and classical protons (PBE-hydrogen). It trains a NequIP equivariant neural network potential on 96-atom PBE configurations, validates it against 200-atom and 2048-atom AIMD data, and then performs long NPT molecular dynamics with the MLIP for systems of 200 to 2048 atoms. The authors fit the density-pressure equation of state to a tanh form (Eq. 1), extract the maximum susceptibility alpha*D (Eq. 2), and plot alpha*D/N versus 1/N. From the finite extrapolated intercept at T=1125 K and the vanishing intercept at 1500 K, together with the system-size dependence of energy fluctuations, they conclude that PBE-hydrogen has a genuine first-order LLPT ending at a critical point at 1250 +/- 50 K and 155-160 GPa, substantially lower than prior estimates and close to the melting line.","tokens_in":16868,"tokens_out":5633,"duration_ms":42928,"significance":"If the conclusion holds, this is an important step: it demonstrates that finite-size scaling can be applied to ab initio-quality hydrogen simulations through machine-learned potentials, and it would help settle a long-standing ambiguity between a first-order transition and a smooth crossover. The paper's strengths are its use of a standard FSS framework, direct comparisons to external AIMD data (200-atom VASP runs and Karasiev's 2048-atom runs), explicit discussion of MLIP validation pitfalls, and public training data. However, the central first-order and critical-point claims currently rest on an unquantified tanh-fit extrapolation and on the fidelity of a single MLIP in exactly the phase region where the documented model biases are largest. The significance is therefore conditional on the requested robustness checks.","major_comments":[{"comment":"The proof that the transition is first-order rests on the claim that lim_{N->infinity} alpha*D/N is finite at T=1125 K. The paper reports no error bars on alpha, D, or the extrapolated intercept, and no goodness-of-fit for the tanh fits. With only four system sizes (N=400, 768, 1200, 2048) and five fitted parameters (D, alpha, kappa, P_t, rho_t), alpha and D are strongly correlated, and unweighted least-squares fits of this functional form can produce a spurious finite intercept. Moreover, the N=2048 trajectory at 175 GPa shows only five density switches in 200 ps (Fig. 4), and Appendix C states that statistical errors are not well estimated in the critical region because of critical slowing down; both facts undermine the reliability of the fitted density at the transition. The authors should provide bootstrap or block-ensemble uncertainties on alpha*D/N, report chi-squared per degree of freedom for each fit, and show the extrapolated intercept with confidence intervals. If the intercept at 1125 K is consistent with zero within uncertainty, the first-order conclusion is not established.","section":"IV A, Eq. (1)-(2), Fig. 7"},{"comment":"All evidence for the first-order character and the critical point comes from the NequIP model, not from direct DFT trajectories, so the attribution to PBE-hydrogen requires that the model biases do not change the transition order or shift the critical point. The paper documents a roughly 4 GPa pressure offset in the molecular phase (Fig. 2), stronger molecular ordering with shorter bonds (Fig. 13), and a quadrupole-quadrupole correlation 1.46 times larger than the PBE reference (Fig. 14). These biases are concentrated in the molecular liquid, precisely the phase whose relative stability determines the transition. The statement in Section IV C that MLIP imperfections are ruled out as a main source of error is therefore too strong. I request a sensitivity analysis: for example, retrain with different loss weights or cutoffs, or compare the FSS intercept and critical pressure between several MLIPs; alternatively, run direct AIMD FSS at the smallest sizes to anchor the MLIP trend. Without such checks, the quoted error 1250 +/- 50 K should not be interpreted as an uncertainty on PBE-hydrogen.","section":"IV C and App. C"},{"comment":"The critical point estimate T_c=1250 +/- 50 K is inferred from the energy fluctuation variance c (Eq. 3) by the following criterion: at T=1200 K, c increases linearly with system size, while at T=1300 K the maxima for the largest systems overlap. This is a qualitative distinction; no quantitative FSS analysis of c is performed, and the authors state that the joint distribution data are too sparse for a scaling collapse. The text does not explain how 1250 +/- 50 is obtained from the four simulated temperatures (1125, 1200, 1300, 1500 K), nor how the uncertainty is estimated. The authors should state the criterion used to interpolate between 1200 K and 1300 K and provide a quantitative measure, such as the ratio c(N)/N as a function of N and T or a fit to a scaling form. This step is load-bearing for the central claim of a critical point at 1250 +/- 50 K.","section":"IV A, Fig. 8"}],"minor_comments":[{"comment":"The caption says the EOS is shown at four temperatures between 1100 K and 1500 K, but the figure legend lists 1100, 1200, 1300, and 1400 K; the text should be corrected.","section":"Fig. 2 caption"},{"comment":"Equation (3) defines c as the variance of the total potential energy divided by L^3, i.e., per unit volume, while the Fig. 8 caption describes c as 'per atom' and the main text says 'per unit volume'; these differ by a factor N. Please make the normalization consistent.","section":"Eq. (3) and Fig. 8"},{"comment":"The sentence 'For system sizes varying from N=200 to N=2048' conflicts with the figure caption, which says the fits use the four largest sizes and treat N=200 as an extra check; the main text should state both facts clearly.","section":"IV A, Fig. 7"},{"comment":"The sentence 'The statistical error is smaller than symbol size, but is not well estimated in the critical region due to critical slowing down' is internally contradictory; please report how error bars were computed and where they are unreliable.","section":"App. C, Fig. 12"},{"comment":"The models are 'available on demand'; for reproducibility, the trained NequIP parameters should be deposited with a persistent identifier alongside the training data.","section":"Data availability"},{"comment":"The bonded-pair probability p_B(r) uses fitted parameters C=4.29 A^-1 and r0=0.766 A; please state the fitting procedure and data used, since these parameters feed the molecular analysis in Fig. 5.","section":"App. D, Eq. (D1)"}],"recommendation":"major_revision","confidential_remarks":"The paper is appropriate for this journal and the topic is timely. The referee concerns are about statistical rigor and MLIP fidelity rather than a fundamental flaw in the approach; both are addressable by additional analysis. I would not recommend rejection at this stage, but the requested uncertainty quantification and sensitivity checks are necessary before the central claims can be accepted."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should know this paper for one reason: it is the first attempt to settle the order of the PBE-hydrogen liquid-liquid transition with finite-size scaling rather than single-size hysteresis or kinks. That is the right tool, and the authors use it about as carefully as the computational cost allows. They train a NequIP potential on 96-atom PBE data, validate against independent 200-atom and 2048-atom AIMD runs, and show that the density susceptibility extrapolated as αD/N vs 1/N stays finite at 1125 K and vanishes at 1500 K. The energy-fluctuation analysis in Fig. 8 is a genuinely independent check, and it points the same way. That is real progress, and the authors deserve credit for being explicit about the MLIP's known biases: a 4 GPa pressure offset in the molecular phase, stronger molecular ordering, and a 1.46× quadrupole correlation. Now the soft spots, in proportion. The central quantitative claim—the critical point at 1250±50 K, 155–160 GPa—is less solid than the prose suggests. Fig. 7 is a linear extrapolation of four fitted αD values, with no error bars, no goodness-of-fit, and five free parameters per temperature. The paper itself admits statistical errors are poorly estimated in the critical region. A bootstrap on the tanh fits is needed before 'clearly finite' is fully earned. That said, the stress-test note slightly overstates the load: even if Fig. 7's intercept were zero within error, the energy fluctuations in Fig. 8 and the joint distributions in Fig. 9 still support a first-order transition at 1125 K. The extrapolated critical point is the fragile piece, and the MLIP biases are not propagated into the ±50 K. Fig. 12 shows the model's density vs temperature curve is close to Karasiev's AIMD but not identical; a bias shift of the critical pressure by a few GPa would not surprise me. The other real soft spot is reproducibility: the model and all trajectories are 'available on demand,' which means not actually available. The training set is public, but the fitted weights are not, and no analysis scripts are deposited. For a paper whose method is FSS, that is a needless handicap. Who is this for? Anyone working on hydrogen phase diagrams, MLIP validation, or weak first-order transitions in fluids. It deserves a serious referee, not a desk rejection. My recommendation: send it out, but require the authors to add bootstrap uncertainties to the αD/N extrapolation, give an error bar that includes the MLIP biases, and deposit the model and data. If those are addressed, this could be a useful reference point for years.","headline":"First finite-size scaling study of the PBE-hydrogen LLPT; the qualitative first-order conclusion is plausible, but the 1250±50 K critical point rests on unquantified fits and MLIP biases.","tokens_in":853,"tokens_out":918,"would_cite":true,"duration_ms":25504,"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 finite-size scaling shows that hydrogen's molecular-to-atomic liquid transition is a genuine first-order phase change, with a critical point at 1250 ± 50 K and 155–160 GPa.","keywords":["liquid-liquid phase transition","hydrogen","machine-learned potential","NequIP","finite-size scaling","PBE functional","critical point","molecular dissociation"],"falsifier":"A direct PBE ab initio molecular dynamics run — no machine-learned potential — at 1125 K and about 166 GPa with 1200 or 2048 atoms, long enough to observe several density switches, would settle the order of the transition: a bimodal density histogram whose $\\alpha D/N$ extrapolation stays finite confirms first-order behavior, while a unimodal histogram whose width shrinks with system size would refute it. Alternatively, retraining the NequIP model with the molecular-phase pressure offset corrected and checking whether the critical point moves by more than 50 K would test the quoted uncertainty.","tokens_in":16423,"feed_emoji":"💧","tokens_out":7778,"duration_ms":61086,"temperature":0.7,"pith_summary":"The paper tries to settle whether the transition from a molecular, insulating liquid of H2 to an atomic, conducting liquid in high-pressure hydrogen is a real first-order phase change or merely a smooth crossover, and where that transition ends. It studies a deliberately simplified model — 'PBE-hydrogen', classical protons governed by the Perdew-Burke-Ernzerhof density functional — so the question is well posed. Using a machine-learned NequIP potential trained on 96-atom PBE data, the authors run systems up to 2048 atoms and apply finite-size scaling. They conclude that at low temperature the transition is genuinely first order, and that its critical point lies at 1250 ± 50 K and 155–160 GPa, considerably lower than earlier estimates. That matters because it puts the critical point close to the melting line, where the liquid-liquid transition may be accessible only as a metastable phenomenon.","feed_headline":"Hydrogen's liquid-liquid transition is first-order, simulations find","feed_subtitle":"New finite-size scaling puts the critical point near 1250 K and 155–160 GPa, close to the melting line.","key_machinery":"The machine is a NequIP potential, an E(3)-equivariant graph neural network trained on 54,000 PBE configurations of 96 hydrogen atoms, which reproduces the DFT forces and energies well enough to run 200–2048-atom NPT trajectories for hundreds of picoseconds. The argument runs on finite-size scaling: the density susceptibility $\\chi_N=\\alpha D$ from a $\\tanh$ fit to the equation of state should grow linearly with $N$ for a first-order transition, and the extrapolated $\\alpha D/N$ distinguishes a true discontinuity from a size-dependent crossover. A second diagnostic is the potential-energy variance per unit volume, $c=(\\langle E^2\\rangle-\\langle E\\rangle^2)/L^3$, whose scaling with $N$ marks the transition and brackets the critical point.","core_discovery":"The central claim is that in PBE-hydrogen the liquid-liquid phase transition is a weak first-order transition at temperatures around 1050–1300 K, with a critical point at 1200–1300 K and 155–160 GPa. The decisive evidence is finite-size scaling of the density susceptibility: fitting the density-pressure isotherm to $\\rho(P)=D\\tanh(\\alpha(P-P_t))+\\kappa(P-P_t)+\\rho_t$ and computing $\\chi_N=\\alpha D$ gives an $\\alpha D/N$ that extrapolates to a clearly finite value as $N\\to\\infty$ at 1125 K, proving a density discontinuity in the thermodynamic limit, while it vanishes at 1500 K, where the system is a crossover. The energy variance $c=(\\langle E^2\\rangle-\\langle E\\rangle^2)/L^3$ grows with system size at 1125 and 1200 K but not at 1300 or 1500 K, bracketing the critical point between 1200 and 1300 K. The paper therefore places the PBE-hydrogen critical point at 1250 ± 50 K, about 250–750 K below earlier estimates, and finds that the transition line meets the melting line near 170 GPa.","pith_inferences":["If the machine-learned potential's roughly 4 GPa pressure offset in the molecular phase and its 1.46-times-larger quadrupole-quadrupole interaction are systematic rather than random, the true PBE critical point could shift beyond the quoted ±50 K; the paper does not propagate these biases into the uncertainty.","The same finite-size scaling protocol could be applied to a potential trained on quantum Monte Carlo energies instead of PBE, directly testing whether the LLPT and its critical point survive in a more accurate model of hydrogen with quantum protons.","If the critical point is as close to melting as this study suggests, static experimental probes of the LLPT may need to access metastable liquid states, much as in the proposed liquid-liquid transition of water.","The tanh form of the equation-of-state fit assumes the transition is in the 3D Ising universality class; near the critical point, corrections to scaling from the lack of particle-hole symmetry could alter the inferred critical temperature, so the 50 K error bar may underestimate model uncertainty."],"forward_implications":["In PBE-hydrogen the molecular-to-atomic liquid transition is first-order below roughly 1250 K, so a density discontinuity survives in the thermodynamic limit at temperatures such as 1125 K.","The critical point is at 1250 ± 50 K and 155–160 GPa, well below the 1500–2000 K range of earlier estimates, placing it near the melting line.","At 1500 K the same model shows no system-size dependence, so the molecular-to-atomic change there is a smooth crossover, not a transition.","Because the density jump is only 2–3% of the total density, simulations with fewer than about a thousand atoms cannot reliably distinguish the transition from a crossover.","The LLPT line likely meets the melting line around 170 GPa, implying a triple point among molecular solid, molecular liquid, and atomic liquid."],"supporting_citations":[{"why":"Supplies the earlier AIMD estimate of the PBE-hydrogen LLPT and melting line against which this paper's 1250 K critical point is compared.","marker":"[13]"},{"why":"Provided the 2048-atom AIMD trajectories used to validate the machine-learned potential and previously argued for the LLPT in PBE hydrogen.","marker":"[21]"},{"why":"Introduced the NequIP architecture, the machine-learned potential used for all large-scale simulations in this work.","marker":"[30]"},{"why":"Supplied the 54,000-configuration PBE database on which the model was trained and a QMC-trained MLIP with a different melting line.","marker":"[23]"},{"why":"Provides the finite-size scaling method for first-order transitions via the magnetization-density analogy used in the $\\alpha D/N$ analysis.","marker":"[24]"},{"why":"Is the textbook finite-size scaling reference that the paper follows for the order-parameter analysis.","marker":"[27]"},{"why":"Supplies the joint density-energy fluctuation distributions used to argue the critical point lies in the 3D Ising universality class.","marker":"[26]"},{"why":"Gives the revised finite-size scaling for fluid critical points, used when discussing corrections from the lack of particle-hole symmetry.","marker":"[38]"}],"fun_headline_variants":["Hydrogen's liquid-liquid critical point sits near 1250 K","Simulations prove hydrogen's liquid-liquid transition is first-order","Machine-learned potential maps hydrogen's liquid-liquid critical point","Critical point of hydrogen's liquid-liquid transition found near melting","Finite-size scaling pins hydrogen's liquid-liquid transition as first-order"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the NequIP potential trained on 96-atom PBE configurations represents the free-energy surface of PBE-hydrogen accurately enough at 200–2048 atoms that its predicted transition order and critical point are unbiased at the quoted level, despite known biases in pressure, molecular bond length, and molecular orientation correlations.","fun_headline_variants_meta":{"raw":{"variants":["Hydrogen's liquid-liquid critical point sits near 1250 K","Simulations prove hydrogen's liquid-liquid transition is first-order","Machine-learned potential maps hydrogen's liquid-liquid critical point","Critical point of hydrogen's liquid-liquid transition found near melting","Finite-size scaling pins hydrogen's liquid-liquid transition as first-order"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00084,"raw_usage":{"total_tokens":3675,"prompt_tokens":973,"completion_tokens":2702,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":589,"completion_tokens_details":{"reasoning_tokens":2617}},"tokens_in":589,"tokens_out":2702,"duration_ms":15625,"temperature":1.0,"reasoning_tokens":2617,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T11:46:13.725906+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A direct PBE ab initio molecular dynamics run — no machine-learned potential — at 1125 K and about 166 GPa with 1200 or 2048 atoms, long enough to observe several density switches, would settle the order of the transition: a bimodal density histogram whose $\\alpha D/N$ extrapolation stays finite confirms first-order behavior, while a unimodal histogram whose width shrinks with system size would refute it. Alternatively, retraining the NequIP model with the molecular-phase pressure offset corrected and checking whether the critical point moves by more than 50 K would test the quoted uncertainty.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provided the 2048-atom AIMD trajectories used to validate the machine-learned potential and previously argued for the LLPT in PBE hydrogen."},{"cited_title":"The current study used 54,000 config- urations with the energies, forces and stresses computed using the PBE functional","cited_arxiv_id":null,"evidence_quote":"Supplied the 54,000-configuration PBE database on which the model was trained and a QMC-trained MLIP with a different melting line."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the finite-size scaling method for first-order transitions via the magnetization-density analogy used in the $\\alpha D/N$ analysis."}],"review_version":1}