{"id":"427175bb-439b-46cc-a518-98d9048f0e8c","arxiv_id":"2411.17191","paper_version":4,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A neural network potential generator that samples unstable short-distance structures and screens them by interatomic distance yields potentials stable for 20 ns MD simulations of propylene glycol and polyethylene glycol.","lead":"This paper describes an active-learning workflow that builds neural network potentials stable enough for 20-nanosecond molecular dynamics simulations of organic liquids and polymers. The central trick is deliberately adding distorted structures with very short interatomic distances to the training data, which keeps long simulations from collapsing.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 70%-volume compression may not sample the actual collapse configuration: PG failed at an O–H distance of ~0.81 Å, but the data enrichment shown in Fig. 8 is reported at 0.87–0.9 Å.","rationale":"The reader's CONDITIONAL verdict is appropriate. My concern sharpens the reader's weakest assumption (adequacy of sampling short contacts) into a specific quantitative statement: the published data do not show that the actual ~0.81 Å failure geometry was ever added to the training set. If it was not, Section 3.1.3's causal narrative overreaches. The check is cheap: the distribution is already computed for Fig. 8; reporting the low-distance tail and the minimum resolves it. This is not a rejection: the 2D-vs-3D control suggests the 3D screening genuinely helps, and the property predictions are externally benchmarked. A secondary gap is that the >10,000-atom (150-mer) stability claim is reported for 'the trained model' without the multi-seed protocol used for the 4-mer (Section 3.2.3), so the headline robustness claim is under-supported; this could be fixed by running the other three seed models on the 150-mer system. Both issues strengthen, rather than overturn, the need for CONDITIONAL acceptance with requested clarifications.","tokens_in":25482,"tokens_out":9133,"duration_ms":81309,"concrete_test":"Compute, for the PG iteration-11 set, the minimum intermolecular O–H distance and the fraction of 3D-screened labeled structures with O–H ≤ 0.85 Å and O–H ≤ 0.81 Å; do the same for the raw NNP-NEMD compressed trajectories (70% volume). If the fraction at ≤0.85 Å is zero, the training data never sampled the collapse geometry and the mechanism as stated is not supported. Then, as a control, run 20 ns NNP-MD with a model trained on the iteration-11 NEMD data but with the O–H < 0.90 Å structures removed (keeping an equal-size random NEMD subset); if stability persists, the short-distance data are not the operative factor.","verdict_should_be":"UNCHANGED","load_bearing_attack":"In Section 3.1.3, the PG collapse at ~18 ps is traced to an O–H contact of ~0.81 Å producing an unphysical 6 eV/Å force on the H atom (Fig. 6 and surrounding text). The proposed remedy is to add NNP-NEMD structures with short O–H distances, screened by the 3D method. But Figure 8, the only quantitative distribution shown, reports enrichment in the 0.87–0.90 Å bin; no counts below ~0.85 Å are given. If the 500 labeled 3D-screened structures contain no (or very few) O–H distances near 0.81 Å, then the training set never actually included the failure geometry. The observed stability would then not directly validate the paper's causal claim that 'insufficient data points with short O–H distances' caused the collapse and that adding such points fixed it. Stability could instead come from an improved repulsive wall at 0.87–0.90 Å that prevents the trajectory from ever reaching 0.81 Å, or from the presence of NEMD data in general; the 2D-vs-3D comparison does not fully separate these because both add NEMD structures. This matters because the method's general recipe—compress to 70% volume and screen for the manually chosen pair—would fail for systems whose critical short contact is rarer or at a different distance. The paper itself notes in the Conclusions that selecting the element pair is a manual challenge; the distance-range question is the quantitative version of that limitation.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper describes an automatic active-learning (AL) generator for constructing neural network potentials (NNPs) for long nanosecond-scale molecular dynamics simulations. The generator iterates over NNP training, NNP-MD sampling, a two-step screening (query-by-committee ensemble followed by structural-feature-based selection), DFT labeling, and retraining. To address stability failures, the authors add an iteration that samples compressed structures via NNP-NEMD (compression to 70% volume) and uses a 3D structural-feature screening that explicitly incorporates a target interatomic distance (O-H for propylene glycol, H-H for polyethylene glycol). The method is demonstrated on liquid PG and PEG: after this added iteration, four independently seeded NNPs all complete 20-ns production runs (up to 10,530 atoms for a 150-mer PEG), without the collapses observed with earlier iterations. The predicted densities, thermal expansion coefficients, and isothermal compressibilities are within about 3-9% of experiment for PG; the predicted PG self-diffusion coefficient is 5.1 +/- 0.4 versus the experimental 2.6 x 10^-7 cm^2/s. The paper claims that the cause of instability was insufficient data with short O-H/H-H distances and that the NEMD+3D screening supplies this missing data.","tokens_in":25773,"tokens_out":11164,"duration_ms":98733,"significance":"If the stability mechanism is as claimed, this is a practically useful methodology contribution to automated NNP construction for organic molecular liquids and polymers. The work provides a reproducible pipeline built on open-source tools (DeePMD-kit, LAMMPS, CPMD, Quantum ESPRESSO), repeated stability tests with four random seeds per system, a direct comparison of 2D vs 3D screening, and validation of physical properties against independent experimental data. These strengths go beyond a single fitting exercise. The impact is, however, conditional: the central causal claim depends on the assumption that the 70%-compression sampling adequately covers the critical short-contact distance (0.81 A for PG), which the reported histogram does not clearly support, and the paper contains an internal ambiguity about whether intramolecular or intermolecular O-H distances are the screened and analyzed quantity. The overstatement of the PG self-diffusion agreement also needs correction.","major_comments":[{"comment":"The collapse is traced to an O-H distance 'between the molecules' of approximately 0.81 A (Section 3.1.3, text near Fig. 6), but Figure 8 reports the distribution of intramolecular O-H distances. Please clarify whether the 3D screening target was the global minimum O-H distance over all atom pairs (including intermolecular) or only intramolecular pairs. If the latter, the added training data may not contain the intermolecular failure geometry at all, and the claim that short O-H data fixed the collapse is unsupported. If the former, the figure label and the surrounding text need correction. In either case, report the count of labeled structures with O-H distances at or below 0.85 A and the minimum O-H distance in the iteration-11 set, so that the reader can verify whether the failure geometry was actually sampled.","section":"3.1.3, Figure 8 and text near Figure 6"},{"comment":"The PG self-diffusion coefficient reported in Table 2 is 5.1 +/- 0.4 x 10^-7 cm^2/s versus the experimental value of 2.6 x 10^-7 cm^2/s, i.e., a factor-of-two overestimate. The Abstract and Section 3.1.4 describe this as 'excellent agreement,' which is not supported by the data. Please either temper the wording or provide a quantitative discussion of the expected accuracy of BLYP-D2 for this property, along with any finite-size or block-averaging caveats. The same overstatement appears in the Conclusions.","section":"3.1.4, Table 2, and Abstract"},{"comment":"The text states that the observed stability trend is 'directly correlated with the number of short O-H distance structures in the data set,' but the only quantitative evidence in Figure 8 is for the 0.87-0.90 A bin, and no counts below 0.85 A are provided. Please report the full distribution of the minimum O-H distance for both the 2D- and 3D-screened iteration-11 labeled sets, including the number of structures below a physically appropriate threshold (e.g., 0.85 A) and the minimum value in each set. This is necessary to distinguish the proposed mechanism (adding the failure geometry) from an alternative mechanism (strengthening the repulsive wall at 0.87-0.90 A so that the trajectory never reaches 0.81 A).","section":"3.1.3, Figure 5 and text"}],"minor_comments":[{"comment":"There is a typo in 'prop ylene glycol'; it should read 'propylene glycol.'","section":"Abstract"},{"comment":"The 3D structural-feature screening is described as incorporating 'normalized minimum O-H distance values,' but the normalization procedure is not specified. Please state the exact normalization used.","section":"Section 2, Screening"},{"comment":"The acceptable maximum-model-deviation range for the model ensemble-based screening is stated for iterations 1-10 (0.05-0.15 eV/A) but not for iteration 11. Please state the range used in iteration 11.","section":"Section 3.1.1 and Table 1"},{"comment":"The comment that model 3 in Figure 5(b) shows much lower performance than in Figure 5(a) due to 'the order of training data loading' is an important sensitivity caveat. It would be helpful to state explicitly in the main text that the 3D-screened data removed this sensitivity, rather than only in the discussion of that figure.","section":"Section 3.1.3, Figure 5"},{"comment":"The NVE energy-conservation check is performed for 1 ns only. Please state whether longer NVE runs were attempted and, if not, note that this is a necessary but not sufficient stability check for the 20-ns production scale.","section":"Supporting Information, Figure S6"}],"recommendation":"major_revision","confidential_remarks":"The paper is within the journal's scope and the proposed NEMD+3D-screening approach is a potentially useful advance for automated NNP construction. The main technical issues--the intramolecular/intermolecular distance ambiguity, the lack of reported data near the 0.81 A failure distance, and the overstatement of the PG self-diffusion agreement--are fixable with additional analysis and rewording. I therefore recommend major revision rather than rejection. Note also that the paper does not provide the generator code; making it available would strengthen reproducibility, but this is not a blocking issue for this journal."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know. The actual novelty is the 3D structural-feature screening—adding a chosen interatomic distance as a third axis to the 2D feature projection—combined with NNP-NEMD sampling of compressed boxes. On two liquids, PG and PEG, the authors show that this materially stabilizes 20 ns NNP-MD runs across four seeds each. The stability result is credible and useful. The second thing is that the property agreement is oversold: for PG, the self-diffusion prediction is 5.1 ± 0.4 × 10⁻⁷ cm²/s against an experimental 2.6, a factor of two, which the text calls \"good agreement\" and the abstract calls \"excellent.\" That is not excellent.\n\nWhat the paper does well: repeated independent seeds per system, a clear mechanistic diagnosis of collapse through short O–H contacts, property benchmarks against experiment rather than only train/test fits, energy conservation in NVE, and a fully specified workflow that others could reproduce if the generator were released. The 3D screening idea is genuinely absent from the cited 2D approaches, so this is a real if incremental contribution.\n\nThe soft spots, in proportion. The causal claim—that missing short O–H data caused the collapse and adding such data fixed it—is not fully supported by the data shown. The collapse happened at an O–H distance of roughly 0.81 Å, but Figure 8 reports enrichment mainly in the 0.87–0.90 Å bin, with no counts shown below 0.85 Å. We do not know whether the 500 labeled 3D-screened structures contain even one 0.81 Å O–H contact. The stability improvement could instead come from an improved repulsive wall at 0.87–0.90 Å that prevents the trajectory from ever reaching 0.81 Å, or simply from having extra NEMD data in general. The 2D-versus-3D comparison does not isolate the mechanism because both arms add NEMD structures. The authors concede that choosing the element pair is a manual step; the distance-range question is the quantitative version of that limitation. A serious referee should ask for a histogram of minimum O–H distances in the 500 labeled structures, or at least a count of contacts below 0.85 Å.\n\nMinor but worth noting: the post-labeling force filter threshold is never specified, and the compression-to-70% heuristic and QBC range are free parameters. No code or data is released, which limits direct verification though not the validity of the stability demonstration itself.\n\nBottom line: this is a solid engineering contribution with a mechanistic story that needs one additional quantitative check. It deserves a serious referee, and the referee should press for the missing short-distance counts before the causal narrative is accepted.","headline":"A practical AL recipe for stabilizing NNP-MD by deliberately sampling short interatomic distances, with a load-bearing but not fully resolved distance-matching question and an overstated diffusion agreement.","tokens_in":26392,"tokens_out":2194,"would_cite":false,"duration_ms":19843,"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":"By deliberately training neural network potentials on compressed structures with very short interatomic distances, the authors achieve stable 20-nanosecond molecular dynamics simulations of liquids and polymers with more than 10,000 atoms.","keywords":["neural network potential","active learning","molecular dynamics stability","interatomic distance screening","nonequilibrium molecular dynamics","propylene glycol","polyethylene glycol","potential energy surface"],"falsifier":"Take a liquid system generated by this method, observe a 20-ns run that collapses despite the short O–H (or H–H) enrichment, and check whether the collapse is preceded by a different geometry, such as a C–C compression, a collective hydrogen-bond rearrangement, or a multi-atom mode that the 70% compression did not generate. Concrete evidence of any such non-contact collapse would show that the sampled short contacts are not the only missing region of the potential energy surface.","tokens_in":1705,"feed_emoji":"🧪","tokens_out":2752,"duration_ms":70883,"temperature":0.7,"pith_summary":"This paper proposes an automatic active-learning generator for neural network potentials (NNPs) that is designed to keep molecular dynamics simulations stable for tens of nanoseconds, not just picoseconds. The central claim is that long-simulation crashes come from poorly sampled regions of the potential energy surface, specifically configurations with very short interatomic distances, and that these regions can be deliberately filled in. The authors combine nonequilibrium simulations that rapidly compress the simulation box to 70% of its volume, generating structures with short O–H (propylene glycol) or H–H (polyethylene glycol) contacts, with a three-dimensional structural-feature screening step that selects labeled structures enriched in those short distances. On liquid propylene glycol and polyethylene glycol systems, the resulting potentials run stably for 20 ns on systems of more than 10,000 atoms, and predicted densities, self-diffusion coefficients, thermal expansion coefficients, and compressibilities agree with experiment at roughly the level of tuned classical force fields.","feed_headline":"Neural-net MD hits 20 nanoseconds with squeezed-box training","feed_subtitle":"An active-learning generator adds compressed structures with very short atom distances, curing the crashes that end long simulations.","key_machinery":"The load-bearing mechanism is an active-learning loop with a two-stage filter. Candidate structures are collected from NNP-driven molecular dynamics; a four-model committee computes the maximum standard deviation of atomic forces as an uncertainty score to reject both well-represented and catastrophically unrealistic frames; then a structural-feature screen projects NNP descriptor-layer features into a low-dimensional space, keeping diverse frames. For the final stability-fixing iteration, the feature space is extended from two dimensions to three by appending the normalized minimum interatomic distance of a user-chosen element pair, and the sampling is replaced by a nonequilibrium compression of the box to 70% of its original volume. The added distance coordinate ensures the labeled set contains many frames in the 0.87–0.9 Å O–H range that ordinary thermal sampling misses, which is what stabilizes the long runs.","core_discovery":"The central discovery is that the instability of long NNP-MD trajectories is not primarily a matter of average force accuracy but of missing data in specific short-distance regions of the potential energy surface. In the authors' runs, collapse begins with unphysical forces (about 6 eV/Å on a hydrogen atom) when an intermolecular O–H distance reaches roughly 0.81 Å, or when H–H distances become extremely short, because the potential has not been trained there. Their generator addresses this with an extra active-learning iteration whose sampling is a nonequilibrium compression of the box and whose screening selects structures by both descriptor-feature diversity and a normalized minimum interatomic distance (O–H for PG, H–H for PEG). With this additional data, all four independently seeded potentials finish 20-ns production runs without a rise in the maximum model deviation, whereas the previous iteration collapsed in at least one trajectory. The paper claims this makes the workflow a general, largely automatic route to stable NNPs for organic materials, and transferable to inorganic systems.","pith_inferences":["The choice of which element pair to watch is still manual; the same logic could be automated by scanning all element pairs for the shortest under-sampled contacts in collapsed trajectories, removing a user bottleneck the authors acknowledge.","If the method generalizes, “squeeze then mine short contacts” could become a standard pre-flight test for any NNP before committing to expensive long simulations, because the 70% compression is an inexpensive probe of potential-energy-surface holes.","A possible limitation not fully addressed is that compression to a fixed volume explores only one family of unstable configurations; collective rearrangements or slow diffusional modes could in principle trigger collapse without producing ultra-short pair contacts.","The density overestimate for PEG chain lengths not present in the training set suggests that transferability across chain lengths may need additional training data, even though the overall experimental trend is reproduced."],"forward_implications":["NNPs trained on small cells (130–155 atoms) can be made stable for production molecular dynamics on much larger cells (up to 10,530 atoms), once short-distance contacts are represented.","For the tested liquids, the predicted density and self-diffusion coefficient match experiment to within a few percent, comparable to or better than classical force fields including one tuned specifically for propylene glycol.","The 3D screening, not merely adding compressed structures, is decisive: the same iteration with 2D screening still produced collapses in some seeds, while 3D screening did not.","Stability correlates with the count of short-distance training data rather than with lower force RMSE, reinforcing the view that dataset design matters more than architecture tweaks for long simulations.","The generator's workflow is applicable to other chemical systems including inorganic ones, according to the authors."],"supporting_citations":[{"why":"Supplies the query-by-committee maximum model deviation criterion used to select candidate structures for labeling.","marker":"[17]"},{"why":"Supplies the nonequilibrium MD approach for generating compressed structures with short interatomic distances.","marker":"[41]"},{"why":"Establishes long and hot MD stability as a key robustness issue for neural network potentials and motivates the hole-filling strategy.","marker":"[11]"},{"why":"Provides the benchmark evidence that force accuracy alone does not determine MD stability, supporting the paper's dataset-design focus.","marker":"[29]"},{"why":"Supplies the active-learning philosophy of sampling chemical space with minimal redundant labeling, which underlies the iterative loop.","marker":"[14]"},{"why":"Provides the structural-feature-based screening idea that the paper extends to NNP descriptor features and a 3D distance coordinate.","marker":"[18]"},{"why":"Supports the use of structural representations to choose diverse training frames for neural network potentials.","marker":"[44]"},{"why":"Illustrates a manual data-selection bottleneck for polymer NNP development, which the proposed generator aims to automate.","marker":"[30]"},{"why":"Provides the DeepPot-SE NNP architecture used for training and evaluation in the demonstrated applications.","marker":"[5]"}],"fun_headline_variants":["Squeezed-box active learning yields stable 20-ns neural-net MD","Short-distance training data prevents neural-net MD crashes","Active learning with squeezed boxes gives stable 20-ns MD","Neural-net MD stability via active learning on short-range regions","Squeezed-box sampling extends neural-net MD to 20 ns"],"cache_read_input_tokens":28416,"weakest_assumption_plain":"The argument rests on the premise that the crashes that end long NNP-MD runs are caused by under-represented short contacts of one chosen element pair, and that rapidly compressing the box to 70% of its volume produces enough of those contacts to patch the potential.","fun_headline_variants_meta":{"raw":{"variants":["Squeezed-box active learning yields stable 20-ns neural-net MD","Short-distance training data prevents neural-net MD crashes","Active learning with squeezed boxes gives stable 20-ns MD","Neural-net MD stability via active learning on short-range regions","Squeezed-box sampling extends neural-net MD to 20 ns"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000549,"raw_usage":{"total_tokens":2673,"prompt_tokens":1051,"completion_tokens":1622,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":667,"completion_tokens_details":{"reasoning_tokens":1535}},"tokens_in":667,"tokens_out":1622,"duration_ms":11941,"temperature":1.0,"reasoning_tokens":1535,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T12:24:46.630881+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a liquid system generated by this method, observe a 20-ns run that collapses despite the short O–H (or H–H) enrichment, and check whether the collapse is preceded by a different geometry, such as a C–C compression, a collective hydrogen-bond rearrangement, or a multi-atom mode that the 70% compression did not generate. Concrete evidence of any such non-contact collapse would show that the sampled short contacts are not the only missing region of the potential energy surface.","supporting_citations":[],"review_version":1}