{"id":"9e037c27-273c-4a33-8a52-663a792db9bd","arxiv_id":"2411.17348","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A maximum relative entropy bias can force molecular dynamics simulations to reproduce a target radial distribution function and thereby induce crystallization into target crystalline polymorphs.","lead":"The authors present a maximum entropy algorithm that biases molecular dynamics simulations to match a target radial distribution function, and demonstrate it can reshape one water model to match another and drive liquid water and liquid TiO2 to crystallize into specific polymorphs. The method could help prepare atomic structures for simulations and help interpret X-ray scattering data when the atomic arrangement is unknown.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Eq. 29's lambda-update has the wrong sign: the printed rule is gradient ascent on Gamma and would push the RDF away from the target, so the reported simulations cannot follow from the equations as written.","rationale":"The paper's central claim is that a maximum-relative-entropy bias linear in the RDF reproduces a target RDF and can drive liquid-to-solid transitions. For that to hold, the adaptive update of the Lagrange multipliers must actually minimize the objective in Eq. 28. As printed, Eq. 29 moves lambda in the direction of increasing Gamma, not decreasing it, because dGamma/dlambda_b = g_target_b - <g_b> and the update adds a positive multiple of this derivative. That is internally inconsistent with the stated gradient-descent procedure and with the sign convention in Eqs. 12-14, where positive lambda_hat suppresses high g values. The liquid-water demonstration in Sec. 3.1 is quantitative and likely generated by a working implementation, so the most plausible resolution is a sign typo in Eq. 29 rather than a fundamentally broken method. But the discrepancy is load-bearing because the paper, as written, cannot be reproduced without resolving it. The statement in Sec. 2.3 that Gamma is not convex because RDF bins are correlated is also incorrect: the second derivative of Gamma is a covariance matrix and is positive semidefinite, so Gamma is convex. The reader's concern about RDF degeneracy and validation of the crystallized polymorphs remains legitimate and should be addressed with replicate runs and quantitative structural analysis, but it is secondary to the equation-level sign issue. I therefore keep the reader's conditional verdict: the method is promising and likely correct in implementation, but the manuscript needs a corrected update rule and verification before the central claim can be accepted.","tokens_in":19829,"tokens_out":10334,"duration_ms":104754,"concrete_test":"Download the provided implementation at https://collaborating.tuhh.de/m-29/software/maxentrdf and locate the update of lambda_b. Check whether the code computes Delta_b = g_target_b - <g_b> and sets lambda_b += gamma kappa_b Delta_b / sum|Delta| or lambda_b -= gamma kappa_b Delta_b / sum|Delta|. Then run a one-bin toy bias (single pair distance histogram, target peak above the unbiased value) and record whether lambda_b increases or decreases while <g_b> approaches the target. If lambda_b increases and g_b moves away, Eq. 29 is implemented and the central algorithm fails; if lambda_b decreases, correct the displayed sign in Eq. 29 and re-verify the TIP3P-to-TIP4P/2005 water transfer with the corrected update.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The most load-bearing problem is a sign error in the core lambda-update. Eq. 28 defines Gamma(lambda) = ln integral P0 e^{-sum lambda_b g_b} dR + sum lambda_b g_target_b, so dGamma/dlambda_b = g_target_b - <g_b>. Minimizing Gamma by gradient descent therefore requires lambda'_b = lambda_b - gamma kappa_b (g_target_b - <g_b>) / sum |g_target - <g>|. Eq. 29 prints a plus sign, which is gradient ascent. Because the bias enters the Boltzmann factor as exp(-beta sum lambda_hat_b g_b) with lambda_hat_b = lambda_b/beta, a positive update in a bin where the target exceeds the current RDF makes lambda_b more positive and suppresses g_b, moving away from the target. The reported water RDF transfer and crystallization runs therefore cannot be reproduced from Eq. 29 as written. Either the implementation uses the opposite sign (a typo that must be corrected) or the stated optimization procedure is inconsistent with the maximum-entropy derivation. This is more basic than the polymorph-identification concern: it affects every reported result.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a maximum relative entropy biasing scheme for MD simulations in which a histogram RDF enters as a linear bias potential V(R) = sum_b lambda_hat_b g_b(R), with Lagrange multipliers updated on the fly. The authors validate the approach by biasing TIP3P water to reproduce the TIP4P/2005 RDF and by biasing liquid water and liquid TiO2 to crystallize into ice Ih, rutile, and anatase, respectively. They also discuss applications to force-field refinement, WAXS interpretation, and machine-learning potential training.","tokens_in":19990,"tokens_out":9343,"duration_ms":85628,"significance":"If the central result holds, the method gives a practical route to translate measured or simulated RDFs into a driving force that can generate crystalline polymorphs, which would be valuable for structure preparation and interpretation of scattering data. The paper reports a quantitative liquid-state transfer (RDF MAE 0.058 to 0.016, ADF MAE 0.065 to 0.023), provides a LAMMPS implementation, and demonstrates distinct polymorph outcomes from the same liquid TiO2 setup. However, the load-bearing sign error in Eq. (29), the internal sign inconsistency between Eqs. (16) and (20), and the reliance on single hand-tuned trajectories for the crystallization claims leave the central claim in need of correction and stronger validation.","major_comments":[{"comment":"The printed update is gradient ascent, not descent. From Eq. (28), dGamma/dlambda_b = g_target_b - <g_b(R)>, so minimizing Gamma requires lambda'_b = lambda_b - gamma kappa_b (g_target_b - <g_b(R)>) / sum |g_target - <g>|. Eq. (29) has a plus sign, which will increase lambda_b in bins where the target exceeds the current RDF; since the bias enters the Boltzmann factor as exp(-beta U - sum lambda_b g_b), this suppresses those bins and moves the ensemble away from the target. This sign affects every reported simulation, so it must be a typographical error in the manuscript or a fundamental inconsistency; the authors should correct Eq. (29) and state explicitly which sign is used in the LAMMPS code.","section":"Sec. 2.3, Eq. (29)"},{"comment":"The force convention is internally inconsistent: Eq. (16) defines the bias force as F_i = grad_i V(R), but Eq. (20) computes hat F_ij from Theta(r_b + Delta/2 - r_ij) - Theta(r_b - Delta/2 - r_ij), which corresponds to -grad_i V(R) given Eq. (17). The correct physical bias force is -grad_i V(R) when V is added to the potential energy. Please reconcile the sign convention and Eq. (16), otherwise readers cannot reproduce the forces from the printed equations.","section":"Sec. 2.2, Eqs. (16) and (20)"},{"comment":"The claim that the biased simulations crystallize specifically into ice Ih, rutile, or anatase is supported only by visual snapshots, averaged bond-order order parameters, and RDF matching, each from a single trajectory. Because the paper itself notes in Sec. 1 that RDF inversion is ill-posed, a quantitative polymorph identification (e.g., fraction of atoms with the target coordination, comparison of the full orientationally resolved structure factor against the reference crystal, or a finite-temperature order-parameter distribution) is needed to substantiate the central crystallization claim. Please add such an analysis or temper the conclusion to RDF-compatible crystalline order.","section":"Secs. 3.2 and 3.3"},{"comment":"All demonstrations are single runs with hand-tuned update parameters (gamma = 10, 2.5, 5.0, 1.25; kappa_b = 1 or 0.25), and no system sizes, thermostat settings, or statistical uncertainties are reported. For a method paper, this is insufficient to establish robustness; at minimum, report system sizes and run-to-run variability, and ideally a small sensitivity scan over gamma and kappa_b.","section":"Secs. 3.1-3.3"}],"minor_comments":[{"comment":"The statement that Gamma(lambda) is not convex because RDF bin intensities are correlated is incorrect: the log-partition function is convex in lambda regardless of correlations, and bin correlations affect the conditioning or rank of the Hessian, not convexity. Please correct this theoretical aside.","section":"Sec. 2.3"},{"comment":"The notation switches between lambda_b and hat lambda_b within the same equation; since hat lambda_b = lambda_b / beta, the first line should use hat lambda_b consistently or define a new symbol.","section":"Sec. 2, Eq. (13)"},{"comment":"The liquid-water transfer test lacks details about system size, equilibration protocol, and the number of independent samples used to compute the MAE; please provide these to make the quantitative comparison reproducible.","section":"Sec. 3.1"},{"comment":"The temperature increase under strong bias is mentioned but not quantified; reporting the temperature drift and thermostat response would help readers assess the robustness of the NPT crystallization runs.","section":"Secs. 3.2 and 3.3"},{"comment":"In the sentence stating that an equilibration period following the Lagrangian parameter update was unnecessary, the word 'Lagrangian' should be 'Lagrange.'","section":"Sec. 3.3"}],"recommendation":"major_revision","confidential_remarks":"The sign error in Eq. (29) is likely a typo, because the reported MAE decreases; if the code has the opposite sign, the central liquid and crystallization results may be reproducible. In that case, the main remaining hurdles are the single-trajectory validation and the incorrect statement about convexity. The manuscript fits the journal's scope, and the code availability is a positive feature."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: the RDF-transfer machinery is real, the water test is quantitative, and the TiO2 polymorph demonstration including metastable anatase is genuinely new. But as written the paper has a sign error in the central update equation, and the crystallization sections are proof-of-principle single runs, not controlled nucleation studies.\n\nWhat is new: refs. 37–39 already contain max-relative-entropy RDF biasing. The new pieces are the LAMMPS implementation with WAXS form-factor weighting and the demonstration that biasing a liquid with a crystal RDF can drive crystallization into ice Ih, rutile, and anatase. The anatase result matters because it targets a metastable polymorph, something simulated annealing and replica exchange would not do directly. The code is posted. Credit where due.\n\nThe liquid-liquid test is the strongest part. TIP3P biased to a TIP4P/2005 target reduces RDF MAE from 0.058 to 0.016 and improves the ADF MAE from 0.065 to 0.023. Since the target comes from a different water model, this is a genuine transfer result, not just self-consistent fitting. Good.\n\nNow the soft spots, in order of importance.\n\n1. Eq. 29 has the wrong sign. With Gamma as defined in Eq. 28, dGamma/dlambda_b = g_target_b - <g_b>. Minimizing Gamma by gradient descent requires lambda'_b = lambda_b - gamma kappa_b (g_target_b - <g_b>)/norm. The printed equation has a plus sign. Given the bias enters as exp(-beta sum lambda_hat_b g_b), a positive update when target exceeds current makes lambda more positive and suppresses that bin, pushing the RDF away from the target. Either the code uses the opposite sign (a typo) or the simulations did not follow the printed update. The authors need to correct this explicitly, because every reported result depends on this step.\n\n2. Crystallization claims are under-supported. Each is a single trajectory, with gamma and kappa hand-tuned, and the target RDFs are generated with the same potential as the liquid. The bias is deliberately designed to force the target RDF, so calling this \"nucleation\" is stronger than the evidence. The bond-order snapshots show the intended polymorphs but with visible defects, and there are no replicate runs or error bars. The paper itself notes that RDF inversion is ill-posed (Sec. 1), so the authors should at least acknowledge that the bias plus original potential must resolve that degeneracy in the systems studied; they do not prove it in general.\n\n3. Minor: the code is not versioned with input data, and parameter sensitivity is not explored.\n\nWho this is for: people who want to generate initial solid structures, feed ML potential training, or interpret WAXS-derived RDFs. It deserves a serious referee. I would send it back with a request to fix the sign, clarify the code's actual update rule, and either add replicate crystallization runs or explicitly label those results as single-run demonstrations.","headline":"Useful paper: sound MaxEnt RDF biasing with a real LAMMPS implementation, but Eq. 29 has a sign error as printed, and the crystallization claims are single-trajectory demonstrations, not nucleation studies.","tokens_in":20601,"tokens_out":3185,"would_cite":true,"duration_ms":34430,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A linear bias built from a measured RDF can drive a liquid to crystallize into the structure behind that RDF.","keywords":["maximum relative entropy","radial distribution function","molecular dynamics","crystallization","nucleation","WAXS","bias potential","polymorph selection"],"falsifier":"Take a target RDF that is shared by two different atomic arrangements, for example a crystal and a defective or amorphous structure with the same $g(r)$, and run the same biased protocol twice with different random seeds; if the final structure changes with random seed or does not match the intended polymorph, the claim that the RDF defines the crystallized state fails. A simpler check is to vary only the bias step size $\\gamma$ or the virial weight $\\kappa_b$ and see whether the same target RDF still selects the same polymorph.","tokens_in":19537,"feed_emoji":"💎","tokens_out":7814,"duration_ms":68549,"temperature":0.7,"pith_summary":"This paper tries to establish that a radial distribution function (RDF), which is routinely obtained from X-ray scattering but is hard to invert into atomic coordinates, can be turned directly into a driving force in a molecular dynamics simulation. The authors derive a linear bias potential $V(R)=\\sum_b \\hat{\\lambda}_b g_b(R)$ from the principle of maximum relative entropy, a rule that picks the least-biased distribution consistent with a set of measured averages, so the biased ensemble reproduces the target RDF while perturbing the original interaction potential as little as possible. They then show the bias works in practice: it transfers the liquid structure of one water model onto another, and it induces liquid-to-solid transitions into specific crystalline states, crystallizing water into ice Ih and liquid TiO2 into rutile or anatase (alternate crystal structures of the same material) depending only on the target RDF. If the method is right, an experimental RDF can act as a structure generator, supplying initial solid configurations that are otherwise hard to prepare, including metastable crystalline forms that annealing-based methods cannot reach.","feed_headline":"A single radial distribution curve can drive liquids to crystallize","feed_subtitle":"Maximum-entropy biasing reproduces target structures from scattering data, including metastable polymorphs like anatase.","key_machinery":"The load-bearing object is the maximum-relative-entropy bias $V(R)=\\sum_b\\hat{\\lambda}_b g_b(R)$, a sum over RDF bins with Lagrange multipliers learned on the fly. Its force per pair, $\\hat{F}_{ij}=\\frac{1}{\\Delta}\\left(\\frac{\\hat{\\lambda}_{\\hat{b}}}{\\delta V_{\\hat{b}}}-\\frac{\\hat{\\lambda}_{\\hat{b}-1}}{\\delta V_{\\hat{b}-1}}\\right)$ (or the analogous expression with $\\hat{b}+1$), depends only on the two bins neighboring each interatomic distance, which makes the bias cheap to compute. Because the bias acts through a linear tilt of the free-energy surface $F(g)=F_0(g)+k_B T \\lambda g$, it both lowers the crystallization barrier and can make a metastable polymorph the preferred state. The virial contribution of the bias is tracked separately, and the gradient-descent update of the multipliers uses the mismatch $\\tilde{g}^{\\mathrm{target}}_b-\\langle g_b\\rangle$, normalized per step.","core_discovery":"The central claim is that the information in a one-dimensional RDF, combined with the existing force field, is enough to determine the crystalline state the simulation will reach. The algorithm replaces the RDF histogram's delta functions with differentiable rectangular kernels, biases the Boltzmann ensemble with the linear potential above, and updates the Lagrange multipliers $\\hat{\\lambda}_b$ during the run by gradient descent on the maximum-entropy objective $\\Gamma(\\lambda)$, with an optional virial-weighting factor to prevent excessive volume changes. Demonstrations show the TIP3P water model reproducing the RDF of TIP4P/2005, with the mean absolute error dropping from 0.058 to 0.016 while the angular distribution error drops from 0.065 to 0.023; liquid water at 360 K and 1 atm crystallizing into hexagonal ice Ih; and liquid TiO2 crystallizing into rutile or the metastable anatase polymorph depending on which RDF is used as target. The resulting crystals contain visible defects, so the achieved structures resemble, rather than exactly equal, the ideal target lattices.","pith_inferences":["Editorial inference: if the RDF target is not unique, the algorithm might reproducibly drive the simulation into an unintended polymorph or an amorphous state whose RDF matches equally well; testing on deliberately degenerate RDF targets would map how often this happens.","Editorial inference: the observed nucleation from a biased liquid suggests the bias trajectory itself samples transition states; configurations harvested near the steep drop in RDF error could seed umbrella sampling or other free-energy calculations without choosing reaction coordinates first.","Editorial inference: the same maximum-relative-entropy construction should work with other ensemble observables, such as bond-order parameters or the structure factor $S(q)$ directly, since the derivation only requires an observable expressed as an ensemble average.","Editorial inference: feeding experimental WAXS-derived RDFs of unknown or amorphous materials into this bias could generate candidate atomic configurations for machine-learned potentials, effectively turning scattering data into training data."],"forward_implications":["Liquid interaction models can be improved by biasing them with an RDF from a more accurate model or from experiment; in the water example the RDF error fell by roughly a factor of four.","Crystalline structures, including metastable ones, can be generated from a liquid without inserting a seed or choosing collective variables; the target RDF alone selects the polymorph.","The bias can stabilize a crystalline phase at temperatures and pressures where it is not the stable phase in the unbiased model, as in water crystallizing at 360 K.","The virial of the bias changes system volume; the optional per-bin weighting $\\kappa_b$ lets the user suppress excessive contraction of the simulation box.","Because the final crystals show defects, the method is a route to near-target, not perfect, crystalline states."],"supporting_citations":[{"why":"Supplies the maximum-entropy biasing formulation and the objective function minimized here.","marker":"[30]"},{"why":"Provides the review of maximum relative entropy for combining simulations with experimental data that frames the method.","marker":"[31]"},{"why":"Describes hybrid reverse Monte Carlo combined with molecular dynamics, the prior approach whose variance constraint and coupling parameter this work avoids.","marker":"[28]"},{"why":"Demonstrates a minimal-bias method for matching experimental data in simulations, a direct predecessor for RDF-targeted biasing.","marker":"[38]"},{"why":"Describes the molecular dynamics engine used for the implementation and all demonstration runs.","marker":"[45]"},{"why":"Provides the virtual X-ray diffraction code and atomic scattering factor approximation used to weight WAXS-derived RDFs.","marker":"[46]"},{"why":"Defines the four-site water model whose RDF is the target in the liquid-transfer test.","marker":"[50]"},{"why":"Defines the water potential used to generate the ice Ih target RDF and to run the crystallization simulation.","marker":"[53]"},{"why":"Defines the TiO2 interatomic potential used to generate the rutile and anatase target RDFs.","marker":"[61]"}],"fun_headline_variants":["One RDF curve can force liquids to crystallize","Single radial distribution dictates crystal phase","Maximum entropy uses RDF to drive nucleation","RDF biasing turns liquids into crystalline solids"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that a one-dimensional target radial distribution, together with the original force field, contains enough information to pick out the intended crystalline phase; the paper notes that RDF inversion is ill-posed and its own final crystals show defects, but it does not prove the bias resolves that ambiguity in general.","fun_headline_variants_meta":{"raw":{"variants":["One RDF curve can force liquids to crystallize","Single radial distribution dictates crystal phase","Maximum entropy uses RDF to drive nucleation","RDF biasing turns liquids into crystalline solids"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000566,"raw_usage":{"total_tokens":2710,"prompt_tokens":1000,"completion_tokens":1710,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":616,"completion_tokens_details":{"reasoning_tokens":1663}},"tokens_in":616,"tokens_out":1710,"duration_ms":12392,"temperature":1.0,"reasoning_tokens":1663,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T12:13:00.904031+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a target RDF that is shared by two different atomic arrangements, for example a crystal and a defective or amorphous structure with the same $g(r)$, and run the same biased protocol twice with different random seeds; if the final structure changes with random seed or does not match the intended polymorph, the claim that the RDF defines the crystallized state fails. A simpler check is to vary only the bias step size $\\gamma$ or the virial weight $\\kappa_b$ and see whether the same target RDF still selects the same polymorph.","supporting_citations":[{"cited_title":"W.; Chodera, J","cited_arxiv_id":null,"evidence_quote":"Supplies the maximum-entropy biasing formulation and the objective function minimized here."},{"cited_title":"Using the Maximum Entropy Principle to Combine Simulations and Solution Experiments","cited_arxiv_id":null,"evidence_quote":"Provides the review of maximum relative entropy for combining simulations with experimental data that frames the method."},{"cited_title":"J.-M.; Coasne, B","cited_arxiv_id":null,"evidence_quote":"Describes hybrid reverse Monte Carlo combined with molecular dynamics, the prior approach whose variance constraint and coupling parameter this work avoids."},{"cited_title":"D.; Voth, G","cited_arxiv_id":null,"evidence_quote":"Demonstrates a minimal-bias method for matching experimental data in simulations, a direct predecessor for RDF-targeted biasing."},{"cited_title":"P.; Sichani, M","cited_arxiv_id":null,"evidence_quote":"Provides the virtual X-ray diffraction code and atomic scattering factor approximation used to weight WAXS-derived RDFs."},{"cited_title":"Molecular Dynamics Simulation of the Structural and Physical Properties of the Four Polymorphs of TiO2","cited_arxiv_id":null,"evidence_quote":"Defines the TiO2 interatomic potential used to generate the rutile and anatase target RDFs."}],"review_version":1}