{"id":"f65c60a5-09f2-4eb7-a46b-48f554c8ab43","arxiv_id":"1908.02448","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"Neural networks trained on molecular dynamics data predict equations of state from pair potentials and invert radial distribution functions into pair potentials more accurately than simple Boltzmann inversion.","lead":"The authors trained neural networks to predict pressure from pair potentials and to recover pair potentials from radial distribution functions in simple fluids. The networks beat low-density Boltzmann inversion, and including force information in the loss function makes the recovered potentials smoother and more accurate.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Potential-level leakage in the train/test split may inflate the reported held-out accuracy: the paper does not state whether simulations of the same potential were kept in the same split, and with 891 data points from 657 potentials a random data-point split would almost certainly leak potentials.","rationale":"The reader's CONDITIONAL verdict is appropriate, but I want to harden the condition. My primary concern is not only that the spline family is unrepresentative, though that remains a secondary worry given the 47% discard rate, but that the train/test split may not be disjoint at the potential level. The text in Sec. 2.2 says only that 790 of 891 data points were used for training/validation and 101 for testing; it never states that all simulations of the same U(r) were grouped into the same split. Since an average of 891/657 ≈ 1.36 retained data points exist per potential, a random split by data point would put many identical potentials on both sides. For the UN/c EOS network, which takes the tabulated U(r) as input, the model would then be evaluated on potentials it has already seen during training, making the reported test-set LMRAE ≈ 0.11 a measure of interpolation over density for known potentials rather than generalization to new ones. For the CN/f potential network, the input g(r) differs at each density, so the leakage is less direct, but the network can still memorize the common underlying U across densities, and the claimed transferability, supported by a single test example in Fig. 9, would be weakened. The 4-fold cross-validation used to select architectures has the same risk if folds are not potential-disjoint. This is a concrete, checkable omission rather than an accusation: if the authors did split by potential, the concern is void and the paper's empirical support remains strong. I would keep the verdict at CONDITIONAL, now explicitly conditioned on demonstrating a potential-disjoint split or re-running the benchmarks on one. The paper's own conclusion already concedes limited coverage and lower structural accuracy than IBI, which is honest but does not address the split issue.","tokens_in":14713,"tokens_out":7290,"duration_ms":75786,"concrete_test":"Check the released repository (gitlab.rlp.net/anikouba/boltzmann) for potential IDs and compute the overlap between the 790 training/validation entries and the 101 test entries. If any potential appears in both sets, re-train UN/c and CN/f from scratch on a potential-disjoint split (all densities of one potential in the same fold, e.g., 90% of potentials for training and 10% for testing) and recompute the test-set LMRAE, δ, and the R^2 of the combined g-to-U-to-P pipeline. If these metrics degrade substantially (e.g., test LMRAE doubles or the combined-pipeline R^2 drops below the current 0.461), the reported generalization is inflated by potential leakage.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that the NNs generalize to new pair potentials depends on the test set containing truly unseen potentials. Section 2.2 reports 657 generated potentials, 1678 simulations, and 891 retained data points, with 790 used for training/validation and 101 for testing. If this split is made at the level of individual simulations rather than potentials, then many test points will share their U(r) with training points at a different density. For the EOS networks the input is U(r) plus rho, so the network can recognize a potential it has already learned and only interpolate in rho; for the g-to-U networks the underlying U seen in training can be memorized even though g(r) differs with density. The central generalization claims would then be unsupported, and the reported test errors (LMRAE ≈ 0.11 for UN/c, δ ≈ 0.03 kBT for CN/f) would not transfer to real soft-matter applications. The same issue affects the 4-fold cross-validation used for architecture selection, since potential overlap across folds would bias model choice. The paper should state whether splits were disjoint by potential; if not, the final benchmarks need to be redone. This concern is distinct from but compounds the reader's point about the 47% discarded simulations.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript introduces neural-network surrogates for two complementary tasks in soft-matter physics: predicting the equation of state from an isotropic pair potential U(r) via learned virial coefficients, and recovering (effective) pair potentials from radial distribution functions g(r). Training data come from 1678 MD simulations on 657 randomly generated spline potentials at four densities, reduced to 891 usable data points after discarding non-equilibrated or heterogeneous runs. The authors compare dense, convolutional, and U-Net architectures, input representations (U, clipped U, exp[-beta U]), and auxiliary inputs such as forces and low-density pressure estimates. They report that a U-Net with clipped potentials predicting B2 through B5 gives near-unity R^2 for pressure, and that a convolutional network with force information in the loss recovers potentials with MAE width delta about 0.03 kBT, far better than Boltzmann inversion. A composite g-to-U-to-P pipeline is also tested, with R^2 about 0.46 on the test set.","tokens_in":15020,"tokens_out":2517,"duration_ms":30417,"significance":"If the reported accuracy and transferability hold, the paper makes a useful practical contribution: it provides fast, differentiable surrogates for the virial equation of state and for inverse coarse-graining, with code released, physically motivated output representations, and a genuine end-to-end out-of-sample check (g to U to P). The systematic ablation of input representations, force information, and loss terms is a strength, and the explicit comparison against the low-density analytic second-virial estimate places the gains in context. The main uncertainty is not methodological novelty but whether the benchmark is trustworthy under a properly disjoint train/test split; the potential-level leakage concern, if real, would invalidate the quantitative claims in Sections 3.1 and 3.2.","major_comments":[{"comment":"The paper does not state whether the 790/101 train/test split and the 4-fold cross-validation used for architecture selection are disjoint at the level of pair potentials. Because each potential is simulated at up to four densities, a random split of the 891 data points would place different densities of the same U(r) in both training and test sets. For the U(r)-to-P networks this would let the network interpolate in density for potentials it has already seen; for the g(r)-to-U networks the target potential itself would be memorized. Since the central generalization claims in Sections 3.1 and 3.2 depend on unseen potentials, the authors must either state explicitly that no potential appears in both training and test partitions, or redo the split and all reported benchmarks (including the cross-validation in Sec. 2.1) with potential-disjoint folds.","section":"Sec. 2.2, data split"},{"comment":"About 47% of the 1678 simulations were discarded because they did not reach equilibrium or formed heterogeneous structures, with the exclusion criteria deferred to the ESI. This filtering is not a neutral preprocessing step: it removes exactly the states where the underlying physics is most complex, so the training distribution may be systematically easier than the distribution of real soft-matter systems to which the networks are advertised. The authors should quantify the excluded cases (e.g., by density, potential shape, and pressure range) and demonstrate that the reported held-out errors are not an artifact of this screening; without this, the transferability claims in the abstract and conclusions are not supported.","section":"Sec. 2.2, data filtering"},{"comment":"The composite pipeline from g(r) through the predicted potential to pressure achieves R^2=0.461 on the test set, and this is the only end-to-end benchmark that exercises both networks together. The paper should report this limitation prominently in the abstract or conclusions and explain the outlier at P approximately 80 kBT/sigma^3: is it within the training density/potential range, and does it correspond to a regime where the discarded simulations would have been relevant? As written, the central 'much more accurate' claim is based on the direct U-to-P task, but a reader applying the method to experimental g(r) will face the composite accuracy, so the discrepancy matters for the stated use case.","section":"Sec. 3.2, Fig. 11"}],"minor_comments":[{"comment":"In the caption, the LMSE value for case (b) is printed as 5.7 x 10^3, which is likely a typo for 5.7 x 10^-3; please correct it, as the comparison with the other cases suggests the exponent is wrong.","section":"Sec. 3.2, Fig. 10 caption"},{"comment":"Reference 18 lists 'J. Chem. Phys. 2018, DETC2018, 241706'; the volume and article identifier appear inconsistent with the journal's format and should be checked.","section":"References"},{"comment":"The inset of Fig. 7 and the axis labels in Fig. 1 would benefit from stating that r is in units of sigma explicitly on the axes, rather than only in the text, for clarity.","section":"Sec. 2.2, Fig. 1"},{"comment":"The text says the training-set LMRAE for UN/c is about 0.06 and the test-set about 0.11, but Fig. 3 shows only train/validation values; stating the test value in the figure caption or in the text where the network is chosen would help the reader connect the figures.","section":"Sec. 3.1, Fig. 4"}],"recommendation":"major_revision","confidential_remarks":"The decisive issue is the train/test split. If the split is not disjoint by potential, the paper's headline numbers are not evidence of generalization and the benchmarks need to be recomputed. The filtering of 47% of simulations compounds this. I would ask the authors to state the split policy explicitly and, if needed, rerun the final benchmarks on potential-disjoint partitions before acceptance."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague —\n\nQuick take on Berressem/Nikoubashman (arXiv:1908.02448): the paper does useful, honest work on two ML tasks — mapping pair potentials to equations of state and inverting g(r) to effective potentials — but the headline accuracy numbers are undercut by an unstated train/test split policy that likely leaks potentials between training and test. That needs to be resolved before taking the numbers at face value.\n\nWhat is genuinely new here: the virial-coefficient intermediate representation (U -> B2...B5 -> P) is a clean idea and clearly helps extrapolate in density; the force-aware loss for potential inversion is a small but real improvement in smoothness and accuracy over plain Boltzmann inversion, with δ ≈ 0.03 kBT on the test set versus 0.5 kBT for U0(r). The systematic comparison of clipped vs exponential inputs and the ablation of force info is more thorough than most papers in this niche. They ship code and example scripts, which is good.\n\nSoft spots, in order of severity.\n\nFirst, the split. Section 2.2 says 657 potentials, 1678 simulations, 891 retained points, and 790/101 train/test split. It never says the split is disjoint by potential. Since each potential is simulated at up to four densities, a random point-wise split almost certainly puts the same potential in both train and test at different densities. For the EOS net the input includes U(r) and rho, so the network can memorize a training potential and merely interpolate in rho. Same for the g-to-U net. The stress-test note is right: without a potential-disjoint split, the LMRAE ≈ 0.11 and δ ≈ 0.03 numbers are not a valid test of generalization to new potentials. This is a fixable issue but it's load-bearing.\n\nSecond, the data filtering. 47% of simulations discarded, with criteria in the ESI. We only have the paper, so I can't judge whether the remaining set is representative, but the authors should quantify how many of the discarded cases are tied to specific potential shapes. If the hard cases correlate with features that matter in real soft matter, the trained nets will be optimistic.\n\nThird, the end-to-end pipeline (g to U to P) drops to R^2=0.461 on the test set. The authors are honest about it, but it means the practical usefulness for coarse-graining is still limited.\n\nFourth, no quantitative comparison with IBI or RMC. The conclusion mentions they expect lower accuracy than iterative methods, but a direct benchmark on the same test potentials would make the 'fast initial estimate' claim much stronger.\n\nWho is it for: anyone working on ML-based coarse-graining or inverse design in soft matter. The force-aware loss idea and the virial-coefficient output representation are worth discussing. I'd bring it to a reading group as a good case study in data-split pitfalls.\n\nRecommendation: it deserves serious peer review, but only after the authors clarify—or redo—the train/test split by potential and report how many potentials actually survive. If the split was point-wise, the central generalization claims need to be recomputed. As is, I would not trust the reported test errors.","headline":"Useful ML-for-soft-matter study with a real data-split leak that needs fixing before the accuracy claims can be trusted.","tokens_in":15513,"tokens_out":4023,"would_cite":true,"duration_ms":41132,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":["05.20.Jj","61.20.Ja","64.10.+h"],"model":"deepseek-v4-flash","headline":"Trained neural networks can predict equations of state and recover effective pair potentials from radial distribution functions with errors an order of magnitude below standard analytic estimates.","keywords":["neural networks","equation of state","virial coefficients","pair potentials","radial distribution function","coarse-graining","Boltzmann inversion","molecular dynamics simulation"],"falsifier":"Take the trained pressure and potential networks and evaluate them on a held-out family of potentials that the spline generator cannot produce, for example long-ranged Yukawa or SALR potentials with a repulsive barrier, and compare pressure and potential errors to the reported values of about 0.11 mean relative error and 0.03 kBT; a substantial error increase would show the training family was not representative. A cheaper check is to re-run the benchmark including the 787 discarded simulations and see whether test errors rise.","tokens_in":14520,"feed_emoji":"⚛️","tokens_out":7699,"duration_ms":78982,"temperature":0.7,"pith_summary":"This paper tries to show that a single trained neural network can replace two standard statistical-mechanics approximations: the low-density virial estimate for equations of state, and Boltzmann inversion for recovering pair potentials from structure. For the forward direction, a U-Net is trained to turn a pair potential into density-independent virial coefficients B2 through B5, and the pressure follows from the virial expansion; on held-out potentials this reaches a mean relative pressure error around 0.11, where the analytic second-virial estimate is off by roughly an order of magnitude. For the inverse direction, a convolutional network maps a measured radial distribution function to the underlying potential, and including force information in the training loss cuts the typical potential error from about 0.5 kBT under Boltzmann inversion to about 0.03 kBT while making the output smoother and more transferable across densities. If correct, this offers a fast, one-shot route to effective potentials for coarse-graining and inverse design, without iterative simulation loops.","feed_headline":"Neural nets beat Boltzmann inversion on fluid potentials","feed_subtitle":"One network learns virial coefficients for pressure; another recovers pair potentials from structure at 0.03 kBT error.","key_machinery":"The load-bearing construction is the virial-coefficient layer: instead of mapping U(r) directly to pressure, the network outputs density-independent coefficients B2 through B5, and fixed, parameter-free layers compute βP = ρ + $B2ρ^{2}$ + $B3ρ^{3}$ + $B4ρ^{4}$ + $B5ρ^{5}$. This forces the density dependence to come from the virial expansion rather than from the network, which is why the model extrapolates better than a direct U-to-P map. For the inverse problem, the crucial mechanism is a loss augmented with force information: the network trains not only on U(r) values but on their finite-difference derivatives, through terms like the discretized Laplacian and difference-correlation losses, so the predicted potential is constrained to vary smoothly. Input clipping, capping U at 20 kBT, and, for the inverse side, using the Boltzmann estimate U0(r) as input are supporting representations that keep the network's search space bounded.","core_discovery":"The central claim is that machine-learned maps can carry the physics that the analytic low-density limit misses. Networks that first predict the virial coefficients {B2,...,B5} from a pair potential and then form the pressure through the virial expansion reproduce simulation pressures with R²≈1.00 on both training and test data, whereas the second-virial estimate has R²≈0.172 and misses negative pressures entirely. On the inverse problem, a convolutional network fed g(r) at a known density, with the potential derivative included in the loss, recovers U(r) with an exponential error distribution of width δ≈0.03 kBT on the test set, compared with δ≈0.5 kBT for U0 = -kBT ln g(r); the recovered potentials reproduce g(r) in fresh simulations and remain nearly independent of the density at which g(r) was measured. The paper also shows the two networks can be chained: potentials predicted from g(r) feed into the pressure network, giving R²≈0.971 on training data and R²≈0.461 on the test set.","pith_inferences":["If the training family were expanded to include long-ranged or bounded potentials, for example screened Coulomb or cluster-forming potentials with repulsive barriers, the same two-stage architecture would plausibly extend to the soft-matter systems the introduction motivates; the current spline family has no such features, so this is an open test rather than a paper claim.","The paper's differing verdict on force information, helpful for recovering potentials but unnecessary or harmful for predicting pressure, suggests the force input matters most when the network must learn derivatives; a shared network with force inputs on both tasks could be tested against the separate designs.","Because 47% of simulations were discarded before training, a useful stress test would be to keep the discarded configurations and compare errors; if the network fails preferentially on those, the reported test-set accuracy overstates real-world robustness.","The end-to-end g(r)-to-pressure chain's drop from R²≈0.971 on training data to 0.461 on the test set hints that error accumulates through the two networks, and a combined loss trained on the full chain might close much of that gap."],"forward_implications":["Pressure predictions from the chosen U-Net stay accurate on densities not seen in training, because the learned B2 through B5 are density-independent and only the polynomial in density is evaluated.","The inverse convolutional network gives potentials that are smoother and closer to the target than Boltzmann inversion, with the predicted potential varying little between g(r) curves measured at different densities.","Predicted potentials can serve as initial guesses for iterative refinement methods such as Iterative Boltzmann Inversion, replacing a poor U0 start and reducing the number of simulation cycles needed.","Chaining the two networks provides an end-to-end route from structure g(r) to pressure, with near-simulation accuracy on training data and correct trends on the test set.","Because pressure comes from learned virial coefficients, the approach inherits the virial expansion's limitation: it is most reliable where the series converges, and the reported accuracy applies to the densities and potential shapes in the training family."],"supporting_citations":[{"why":"supplies the U-Net block architecture used in the pressure-prediction networks.","marker":"[46]"},{"why":"gives the analytic second virial coefficient formula that defines the low-density baseline P0 the pressure networks are compared against.","marker":"[52]"},{"why":"defines the WCA hard-core repulsion added to about half of the training potentials.","marker":"[49]"},{"why":"provides the GPU molecular dynamics engine used to generate the training pressures and g(r).","marker":"[50]"},{"why":"documents the strong-scaling molecular dynamics implementation used for the same data generation.","marker":"[51]"},{"why":"describes iterative Boltzmann inversion, the standard coarse-graining method the inverse networks are positioned against.","marker":"[26]"},{"why":"states the uniqueness result that justifies learning the potential from the pair correlation function.","marker":"[37]"}],"fun_headline_variants":["Neural nets beat analytic virial and Boltzmann inversion","Nets predict pressure and recover potentials from g(r) at 0.03 kBT","ML maps g(r) to pair potentials better than Boltzmann inversion","Two neural networks outdo standard fluid physics estimates"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The accuracy claims depend on the 657 randomly generated spline potentials that survived the simulations being representative of real isotropic pair potentials; if the discarded simulations or unseen potential shapes are systematically harder, the reported test-set errors will not carry over to applications.","fun_headline_variants_meta":{"raw":{"variants":["Neural nets beat analytic virial and Boltzmann inversion","Nets predict pressure and recover potentials from g(r) at 0.03 kBT","ML maps g(r) to pair potentials better than Boltzmann inversion","Two neural networks outdo standard fluid physics estimates"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000681,"raw_usage":{"total_tokens":3078,"prompt_tokens":916,"completion_tokens":2162,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":532,"completion_tokens_details":{"reasoning_tokens":2090}},"tokens_in":532,"tokens_out":2162,"duration_ms":20588,"temperature":1.0,"reasoning_tokens":2090,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T14:43:22.961735+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take the trained pressure and potential networks and evaluate them on a held-out family of potentials that the spline generator cannot produce, for example long-ranged Yukawa or SALR potentials with a repulsive barrier, and compare pressure and potential errors to the reported values of about 0.11 mean relative error and 0.03 kBT; a substantial error increase would show the training family was not representative. A cheaper check is to re-run the benchmark including the 787 discarded simulations and see whether test errors rise.","supporting_citations":[],"review_version":1}