acceptodds
Under review as a conference paper at ICLR 2027

Entanglement-aware Flow Matching for Polymer Dynamics

Abstract

Learning long-timescale dynamics in dense polymer systems is important for understanding material behavior, but remains difficult because chain motion is governed by entanglement, a collective topological constraint that existing learned dynamics models often fail to preserve during generation. We formulate windowed Conditional Flow Matching (wCFM) for force-free polymer dynamics, training continuous-time velocity fields on trajectory windows sampled across multiple temporal scales. We further introduce orthF, a geometry-aware regularization that penalizes predicted motion across local inter-chain boundaries. We evaluate orthF on molecular dynamics trajectories of dense polymer systems across three dynamics models, showing that its benefit is not limited to a single architecture. Across 9 polymer datasets,orthF improves aggregate displacement accuracy for all three models, yielding generated trajectories that better match ground-truth distributions and structural and topology-related observables. Once trained, our approach generates a 50ns backbone trajectory in 13s on a single GPU, compared with 43h for all-atom molecular dynamics on 64 CPUs, corresponding to an approx 1.2x 10^4 wall-clock sampling-time ratio. These results show that explicit geometric inductive bias is a simple and effective way to improve learned dynamics in entangled molecular systems.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Molecular Dynamics

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_2J7HOpenReview ↗
Rating
3(+1)
Confidence
3(+1)
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper addresses dense polymer dynamics by accounting for the constraint that polymer chains cannot freely pass through one another. It directly learns the motion of polymer backbones and introduces a local geometric regularizer to discourage predicted motion that could lead to inter-chain crossings, thereby improving the stability of long-horizon trajectory generation.

Critical strengths
  1. The paper focuses on dense polymer systems and designs a method around their distinctive physical properties. The motivation is clear and well justified, and the problem has practical significance.

  2. The method is simple, intuitive, and applicable across different models. orthF translates local inter-chain geometry into a regularizer on velocity directions without requiring expensive global topological quantities to be explicitly computed during training. It can be applied to wCFM, TFM, and MLCGMD.

  3. Beyond displacement distributions, the paper evaluates multiple additional metrics, providing a more comprehensive assessment and stronger empirical support.

Critical weaknesses
  1. The connection between the proposed regularizer and the prevention of chain crossings is largely heuristic and lacks direct validation. orthF uses the direction from an atom toward the mean position of neighboring atoms on other chains as its constraint direction. However, this direction does not necessarily correspond to the actual collision normal between chain segments. The current results do not establish whether the regularizer reduces chain-crossing events or merely restricts motion directions more generally.

  2. It remains unclear whether the reduced topological drift reflects the benefits of the proposed design or simply results from orthF suppressing physically valid large-amplitude motion. The MSD of orthF is lower than that of the baseline, and both are lower than the MD reference. Thus, the smaller structural and topological drift may partly arise from reduced motion rather than more accurate modeling of entanglement.

  3. Although the paper aims to learn polymer dynamics under entanglement constraints, orthF is only a soft regularizer on local velocity directions. It guarantees neither the absence of inter-chain crossings nor the preservation of global topology.

Questions for authors
  1. Do the benefits of orthF persist when the baseline and orthF are matched in motion amplitude? Do the generated trajectories reproduce the correct temporal behavior of the physical system?

  2. What are the respective contributions of the geometric input features and the regularization loss?

  3. Can chain-crossing events be measured directly, rather than assessed solely through indirect statistical metrics?

2 comments
AuthorsOpenReview ↗

Rebuttal response

We thank the reviewer for recognizing the relevance of entanglement-aware polymer dynamics and the transferability of orthF across different dynamics models. Below, we address the concerns with methodological clarifications and supporting experimental evidence.

W1–W3 — Soft crossing constraint and motion suppression.

We thank the reviewer for highlighting the physical interpretation of orthF. We agree that it is a soft local geometric prior rather than an exact non-crossing constraint. Nevertheless, we evaluate its effectiveness using established topology-related descriptors and Z1+ primitive-path diagnostics. On 3hb, orthF reduces final Cross-ACN by 61.5% (49.48→19.07; MD: 5.91) and Self-ACN by 50.5% (284.94→141.03; MD: 54.62; Table 4). Although atom MSD decreases relative to Base (273.5→234.3 Ų; MD: 322.1 Ų), orthF also improves displacement-distribution WSD across all three dynamics families (Table 2) and bond-length and dihedral distributions (Table 3). These results demonstrate improved agreement with MD across complementary observables, rather than relying solely on reduced crossing-related statistics.

Q1 — Controlled motion amplitude and temporal behaviour.

We thank the reviewer for raising this concern. Motion amplitude alone is insufficient to characterize entangled polymer dynamics: velocities with identical magnitudes can produce different trajectories depending on their directions and temporal correlations. Importantly, orthF penalizes cosine alignment with local inter-chain directions rather than velocity magnitude, so it does not explicitly encourage slower motion. Nevertheless, we additionally evaluate motion amplitudes at lag times of 1–4 ns:

Lag timeMDBaseorthF
1 ns0.01130.00680.0052
2 ns0.00670.00580.0047
3 ns0.00510.00590.0043
4 ns0.00400.01160.0038

At longer lags, particularly 4 ns, orthF maintains motion amplitudes closer to MD, whereas Base deviates substantially. Together with improved displacement-distribution agreement (Table 2), structural fidelity (Table 3), and topology-related descriptors (Table 4), this supports the benefit of direction-aware regularization beyond scalar motion amplitude alone. We acknowledge that a motion-matched control would be required to exclude an indirect contribution from reduced mobility conclusively.

Q2 — Input features versus the regularization loss.

We appreciate the opportunity to clarify this experimental control. Both Base and +orthF receive identical geometric input features, including local inter-chain directions and backbone tangents; they differ only in the orthF regularization loss. Thus, the comparison isolates the contribution of the loss under the same geometric conditioning. Adding orthF improves mean displacement WSD across wCFM (0.053→0.047), TFM (0.149→0.035), and MLCGMD (0.628→0.618; Table 2), while also improving final Cross-ACN on 3hb (49.48→19.07; Table 4). We will make this distinction explicit in the revision.

Q3 — Direct measurement of crossing events.

We thank the reviewer for this suggestion. We evaluate inter-chain geometry and entanglement-related structure using established measures, including Cross-ACN, Self-ACN, writhe, and Gauss linking (Table 4; Appendix F.2), together with Z1+ primitive-path analysis (Fig. 4; Table 16). These quantitatively characterize the resulting polymer configurations. Individual crossing events between discretely sampled frames cannot be uniquely reconstructed without assumptions about intermediate chain motion. While these measurements are not explicit event counts, they provide direct evidence of improved agreement with MD in established topology-related descriptors.

We appreciate the reviewer's constructive feedback, which has helped clarify the physical interpretation and experimental validation of orthF. The complementary displacement, structural, and topology-related results reinforce the benefits of geometric regularization for entangled polymer dynamics.

Thanks for the author's further illustration and additional experimental results. Most of my concerns have been solved. I regard geometric prior as an effective stepforward. I will raise my score to 3 accordingly.

Reviewer_shhfOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper targets learning long-timescale dynamics in dense polymer systems. The author identified the problem of entanglement in polymer dynamics simulations. The author proposed windowed conditional flow matching sampled across multiple temporal scales and introduced a geometry-aware regularization to regulate local inter-chain boundaries.

Critical strengths
  1. The problem is clear, and the representation is easy to follow.
  2. The method is relatively simple to integrate.
Critical weaknesses
  1. The speed acceleration claim is somehow awkward: "Once trained, our approach generates a 50 ns backbone trajectory in 13 s on a single GPU, compared with 43 h for all-atom molecular dynamics on 64 CPUs". The classical MD is supposed to output the full trajectory, and the proposed method only outputs the trajectory with a trained time-stride (10 ps). The output trajectory is not the same; the coarse-grained trajectory could not replace the fine-grained trajectory completely, so such a comparison has no practical contributions.
  2. The motivation for "multiple temporal scales" for training CFM is unclear. The multiple temporal scales design is claimed major contribution, but it has no clear motivation, intuition, and its effectiveness is not verified with an ablation study. Its rationale is also not clear.
  3. The benefit of "multi-scale window sampling": exposing the model to both short-time local motion and longer-time conformational change is not verified either theoretically or empirically.
  4. The interpolated state is simply linear interpolation. However, the dynamics simulation is nonlinear; I suppose that this may pass through configurations that are not physically plausible and may create overlaps or distort local geometry. This should be clarified.
  5. The orthF penalty is not a direct measure of crossing. In my understanding, the loss of orthF only penalized squared cosine alignment between an atom’s velocity and the mean direction to nearby atoms on other chains. I suppose this still omits crossing-like motion that is not aligned with the mean direction.
  6. Overall, the design is rather incremental: a motivation-unclear but incremental design with "multiple temporal scales", and an indirect training loss penalty.
  7. The evaluation is rather weak; the author only used LAMMPS for the datasets. Ab initio level datasets should be considered at least for some small-scale molecules.
  8. The baselines are insufficient, I do not know the purpose of comparing with MLP nd Transformer. Besides, MD-oriented ML methods (TFM and MLCGMD) are outdated. Why did the author not compare with the methods mentioned by the author, such as MDGen, ConfRover, and Star-MD?
Questions for authors
  1. Since the learned model generates a backbone trajectory at a 10 ps stride, whereas the all-atom MD reference produces a fine-grained trajectory, what practical tasks is the reported speed comparison intended to represent?

  2. What is the motivation for training wCFM across multiple temporal scales? Could the authors explain why this design should improve learning, and provide an ablation comparing multi-scale window sampling with single-scale sampling?

  3. The paper states that multi-scale windows expose the model to both short-time local motion and longer-time conformational change. What evidence supports this claim? Could the authors provide empirical analyses of performance across timescales or observables?

  4. Since the flow-matching states are constructed by linear interpolation between MD endpoints, how do the authors address the possibility that these interpolants may be physically implausible, contain overlaps, or distort local geometry?

  5. OrthF penalizes squared cosine alignment between an atom’s velocity and the mean direction to nearby atoms on other chains. How do the authors validate that this directional penalty reduces crossing-like motion, given that it may miss motions not aligned with the mean direction and may also penalize physically valid motion?

  6. Could the authors clarify the methodological novelty of the overall design? In particular, what do multi-scale window sampling and the orthF regularizer contribute beyond existing dynamics-learning approaches, and which experiments isolate each component’s effect?

  7. Could the authors justify the dataset scope and discuss how the findings might generalize beyond the LAMMPS-generated datasets?

  8. Could the authors clarify the purpose of the MLP and Transformer comparisons and justify the selection of TFM and MLCGMD as baselines? Would comparisons with relevant methods such as MDGen, ConfRover, or Star-MD be appropriate under matched evaluation conditions?

5 comments
AuthorsOpenReview ↗

Rebuttal response

We thank the reviewer for recognizing the relevance of the problem and the simplicity of our approach. Below, we address all concerns through methodological clarifications, ablation studies, and experimental evidence.

W1 / Q1 — Scope of the speed comparison.

We thank the reviewer for raising this distinction. We model long-timescale polymer backbone dynamics, not full atomistic trajectories at MD timesteps. Backbone extraction retains the original coordinates of selected atoms without introducing pseudo-atoms (Section 3.1). For this defined task, generating a 50 ns trajectory at 10 ps sampling intervals takes 13 s on one GPU, compared with 43 h for all-atom MD on 64 CPUs (Table 6). This is a post-training sampling-time comparison for the target backbone representation, not a claim of equivalent atomistic outputs. Training takes 3.1 GPU-hours, excluding the cost of generating training data.

W2–W3 / Q2–Q3 — Motivation and ablation of multi-scale windows.

Polymer relaxation spans local motion and slower collective changes. Multiple window lengths supervise both within one velocity field. Tables 13–14 compare five fixed-window settings with mixed-window training across nine systems. With orthF, mixed-window training achieves the lowest mean trajectory MSE (0.0156 vs. 0.0244 for the best fixed window), while the optimal fixed window varies across polymers.

W4 / Q4 — Linear interpolation and physical configurations.

We thank the reviewer for raising the physical interpretation of interpolation. Linear interpolation is a standard construction in conditional flow matching (Lipman et al., 2023). Interpolated states are auxiliary training samples, not physical trajectories. Under the squared-error objective, the optimal velocity field is the conditional expectation of target velocities given the interpolated state and time. Thus, the learned field need not follow individual straight-line paths. In wCFM (Eqs. 2–4), endpoints come from windows of the same MD trajectory and we evaluate the physical fidelity on generated trajectories.

W5 / Q5 — Directional penalty and crossing-like motion.

We appreciate the reviewer's observation that the local inter-chain direction is not an exact collision normal. orthF is a soft bias, not a non-crossing guarantee, and retains motion outside the penalized direction. Its effectiveness is evaluated empirically: on 3hb, final Cross-ACN decreases from 49.48 to 19.07 and Self-ACN from 284.94 to 141.03 (Table 4), while displacement WSD improves across all three model families (Table 2). orthF also achieves lower MSE than hard projection orthV in all three families (Table 5).

W6 / Q6 — Methodological contribution.

We thank the reviewer for the opportunity to clarify our methodological contribution. We target dense, entangled multi-chain dynamics rather than single-molecule conformations. We combine a nine-system polymer benchmark and chemistry-aware backbone representation with multi-scale flow matching and a geometric regularizer applicable across three dynamics formulations. Window ablations (Tables 13–14) evaluate temporal sampling. Base/+orthF (Table 2) and soft-versus-hard comparisons (Table 5) assess regularization. Beyond displacement accuracy, we demonstrate improved bond-length and backbone-dihedral distributions (Table 3) and topology-related agreement with MD (Table 4).

W7 / Q7 — Dataset scope and generalization.

We thank the reviewer for raising the dataset scope. Our benchmark comprises nine chemically distinct dense polymer systems with approximately 2,900–8,900 backbone atoms each, generated from all-atom MD. Unlike small-molecule benchmarks, our systems capture collective constraints among long chains. Our method uses backbone coordinates, connectivity, and chain identities, not LAMMPS-specific features. It can therefore, in principle, be applied to trajectories generated using other simulation engines or force fields, although such transfer has not yet been evaluated.

W8 / Q8 — Neural backbones and molecular-dynamics baselines.

We thank the reviewer for this suggestion. The MLP, Transformer, GAT, and other comparisons are neural-network architecture ablations within wCFM. We selected TFM for its related flow-matching formulation and MLCGMD for its direct relevance to polymer dynamics, enabling evaluation of orthF across different temporal-learning approaches. MDGen jointly generates trajectory windows, while ConfRover and STAR-MD employ autoregressive diffusion, potentially requiring repeated sampling and accumulating rollout errors. Moreover, their published implementations primarily target single-protein dynamics with protein-specific representations, requiring substantial adaptation for dense multi-chain polymers. We will clarify this baseline rationale in the revision.

We appreciate the reviewer's constructive feedback, which has helped clarify the motivation, evaluation scope, and methodological contributions.

Thank you for the detailed response and the revised experiments. My questions about the scope of the speed comparison and the motivation and ablation of multi-scale windows are mostly addressed. However, several concerns remain:

  • The mixed-window ablation does not directly demonstrate improvements in both short-time and long-time observables. Moreover, mixed-window training is unstable without orthF, suggesting that its benefit depends on the regularizer.
  • The absence of matched comparisons with more recent trajectory-generation methods, including those discussed in the “Continuous-time trajectory learning” section, limits my ability to assess the method’s relative contribution. The authors compare against TFM, which was developed for clinical time-series modeling rather than molecular dynamics, while citing adaptation challenges as a reason not to compare with more directly relevant molecular trajectory models. This baseline-selection rationale is not sufficiently convincing.
  • The results remain confined to one simulation engine and force field under the studied conditions. The statement that the method could, in principle, be applied to trajectories from other engines or force fields does not provide evidence of such generalization.
  • While reviewing the additional ablations, I noticed that Table 9 reports an MLP rollout MSE of 1.100 ± 6.140 for 3hp. The unusually large standard deviation raises concerns about the reliability and reporting of the results.

Given these remaining concerns, I maintain my score.

AuthorsOpenReview ↗

Addressing Reviewers concerns

We thank the reviewer for the follow-up and for acknowledging that we have largely addressed the speed comparison and multi-scale ablations. We clarify the remaining concerns below.

1. Multi-scale window sampling.

We thank the reviewer for further examining the window-size ablation. Tables 13–14 directly compare mixed- and fixed-window training under identical orthF regularisation. Across nine systems, mixed-window training achieves lower mean trajectory MSE than the best fixed-window setting (0.0156 vs. 0.0244), demonstrating the benefit of combining temporal windows for trajectory prediction. This comparison directly isolates the effect of window mixing with orthF held fixed. The instability of mixed-window Base does not invalidate this result; rather, it provides complementary evidence that multi-scale supervision and geometric regularisation interact in the proposed framework.

2. Baseline selection.

We thank the reviewer for highlighting MDGen, ConfRover, and STAR-MD. We selected these baselines to represent complementary trajectory-learning formulations under the same polymer data and rollout protocol. Although TFM was originally introduced for clinical time series, its domain-agnostic trajectory-level flow-matching formulation provides a relevant methodological comparison to wCFM. MLCGMD, meanwhile, represents autoregressive time-coarsened dynamics and was developed specifically for polymer systems.

The recent methods address related but distinct settings: ConfRover and STAR-MD generate conformations autoregressively using diffusion-based approaches, whereas MDGen jointly generates trajectories using a flow-based model. Their published implementations target peptides and individual proteins with protein-specific representations and modelling components, requiring adaptation to Cartesian backbone dynamics in dense, interacting polymer systems. Our goal is to evaluate whether orthF provides a useful geometric inductive bias across different dynamics formulations. TFM and MLCGMD therefore enable controlled comparisons under matched conditions, with Base/+orthF ablations isolating the contribution of regularisation (Tables 2 and 5).

3. Dataset scope and generalisation.

We thank the reviewer for raising the question of generalisation. Our contribution targets long-timescale dynamics and entanglement-related observables in dense multi-chain polymer systems, rather than transfer across simulation engines or force fields. To this end, we construct a benchmark of nine chemically distinct polymer systems under a consistent all-atom MD protocol, enabling evaluation of collective inter-chain dynamics across different molecular structures.

Importantly, our formulation operates on backbone coordinates, connectivity, and chain identities, without LAMMPS-specific inputs. Thus, it is not intrinsically restricted to a particular simulation engine, although cross-engine and cross-force-field transfer have not been evaluated. These constitute separate generalisation questions and do not undermine the demonstrated results within the studied physical setting.

4. MLP variance in Table 9.

We thank the reviewer for the careful reading. The large value on 3hp reflects a heavy-tailed error distribution in which some MLP rollouts diverge, which inflates both the mean and the standard deviation. This entry belongs to the architecture ablation only: the graph-based backbones show no comparable spread on any system, and all main results use the GAT backbone. The behaviour is consistent with prior observations that unstructured MLP dynamics models are sub-optimal for particle systems (Rubanova et al., 2022) and can diverge in rollout (Greydanus et al., 2019).

We appreciate the reviewer's continued consideration and hope these clarifications address the remaining concerns about the interpretation and scope of our experimental evidence.

Reviewer_ZpCqOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes an entanglement-aware framework for learning polymer dynamics from MD trajectories. The authors introduce wCFM, which trains a continuous-time velocity field using trajectory windows sampled at multiple temporal scales. To account for the geometric constraints imposed by neighboring polymer chains, they further propose orthF, a regularization term that penalizes predicted velocity components aligned with local inter-chain directions. The framework also uses a chemistry-aware backbone representation extracted from molecular graphs.

Critical strengths
  1. An important and practically relevant problem. Learning long-timescale dynamics in dense polymer systems is challenging because intermolecular interactions and topological constraints make direct MD simulation computationally expensive.

  2. Evaluation across multiple model families and polymer systems.

  3. Multiple evaluation perspectives. In addition to displacement-distribution accuracy, the paper evaluates bond lengths, backbone dihedrals, radius of gyration, mean-squared displacement, and topology-related observables.

Critical weaknesses
  1. The proposed orthF regularization does not directly characterize or prevent topological crossings. My primary concern is the physical validity of the proposed entanglement-aware regularization. In Eq. (8), orthF penalizes the cosine alignment between an atom's predicted velocity and the average direction toward nearby atoms belonging to other chains. However, this quantity is not equivalent to a non-crossing constraint or a direct measurement of topological entanglement. Although the authors acknowledge that orthF does not guarantee topological conservation, the central claim that it improves entanglement-aware dynamics still requires more direct evidence. In particular, the paper does not report explicit crossing-event statistics demonstrating that the proposed constraint actually reduces unphysical chain crossings.

  2. The observed improvements may partly arise from suppressing polymer motion rather than preserving entanglement. This is an important confounding factor that is not sufficiently addressed. Table 4 shows that orthF produces lower mean-squared displacement than the unregularized baseline, while both models are already under-diffusive compared with MD. Specifically, the atom MSD decreases from 273.5 to 234.3, moving further away from the MD reference of 322.1. This suggests that some improvements in structural stability may result from restricting molecular motion rather than correctly modeling the underlying dynamics.

  3. Insufficient ablation to isolate the contribution of the proposed regularization. The proposed framework introduces both geometry-dependent input features and an additional velocity regularization loss. However, the experiments do not clearly disentangle their respective contributions. The improved performance could arise from exposing the model to local inter-chain geometric information rather than from the particular orthF objective. Although Table 5 compares orthF against a hard velocity projection, this comparison is insufficient to establish that the proposed regularization is more effective than simpler alternatives.

  4. The methodological novelty of wCFM and orthF is relatively limited. The proposed wCFM objective is a fairly direct adaptation of Conditional Flow Matching to trajectory windows sampled at different temporal scales. The use of linear interpolation between observed endpoint configurations and velocity regression follows the established flow-matching formulation. Existing work, including Trajectory Flow Matching and MDGen, has already explored learning molecular or temporal trajectories using generative modeling. Related work on learning long-time polymer dynamics, such as ML-GLE, also provides relevant context regarding physically meaningful temporal extrapolation and diffusion.

Questions for authors
  1. Can the authors directly validate that orthF reduces unphysical inter-chain crossings?

  2. How can the authors distinguish entanglement preservation from motion suppression?

  3. What is the independent contribution of geometric conditioning versus the orthF loss?

  4. How does the proposed method capture physically meaningful long-timescale dynamics?

1 comment
AuthorsOpenReview ↗

Rebuttal Response

We thank the reviewer for recognizing the importance of this problem and for the constructive feedback. Below, we address all concerns with methodological clarifications and supporting experimental evidence.

W1 / Q1 — Validation of inter-chain crossings.

We thank the reviewer for highlighting the importance of validating inter-chain crossings. We agree that orthF is a soft geometric regularizer rather than an explicit non-crossing constraint. We evaluate its effect using established topology-related descriptors of polymer chains (Table 4; Appendix F.2) and complementary Z1+ primitive-path analysis (Fig. 4; Table 16). On 3hb, orthF reduces final Cross-ACN from 49.48 to 19.07 (MD: 5.91) and Self-ACN from 284.94 to 141.03 (MD: 54.62). Final-frame writhe and absolute Gauss linking also move closer to MD. Although individual crossing events between sampled frames cannot be uniquely identified, these established descriptors provide direct quantitative evidence that orthF improves entanglement-related geometric fidelity.

W2 / Q2 — Motion suppression versus entanglement-related fidelity.

We thank the reviewer for raising the potential motion-suppression confound. We acknowledge that orthF reduces MSD relative to Base, but MSD alone is insufficient to characterize entangled polymer dynamics. Importantly, orthF penalizes velocity alignment with local inter-chain directions rather than velocity magnitude, so it does not explicitly encourage slower motion. Empirically, orthF improves displacement-distribution WSD across all three model families (Table 2) and reduces excess bond-length growth relative to MD by approximately 49% at frame 97 (0.231 Å for Base vs. 0.118 Å for orthF; Table 16). Together with improved backbone-dihedral distributions (Table 3) and topology-related descriptors (Table 4), these results support improved geometric fidelity beyond displacement magnitude alone, although reduced mobility may contribute indirectly.

W3 / Q3 — Geometric conditioning versus the orthF loss.

We thank the reviewer for highlighting the need to isolate the contribution of orthF. Both Base and +orthF receive identical molecular coordinates, connectivity, and local geometric features derived from the current configuration, including inter-chain directions and backbone tangents. No explicit entanglement labels or topological descriptors are provided as inputs. The only difference is the additional orthF loss, so their comparison isolates the effect of regularization under identical geometric conditioning. Adding orthF improves mean displacement WSD across wCFM (0.053→0.047), TFM (0.149→0.035), and MLCGMD (0.628→0.618; Table 2). Furthermore, orthF achieves lower trajectory MSE than both Base and the hard velocity projection orthV across all three families (Table 5). Thus, the improvements cannot be attributed solely to access to geometric inputs, which are identical in both variants.

W4 / Q4 — Novelty and long-timescale dynamics.

We appreciate the opportunity to clarify our methodological contribution. We agree that wCFM builds on established conditional flow-matching principles rather than introducing an entirely new generative objective. Our contribution is extending trajectory-based generative learning to dense, entangled multi-chain polymer systems, where collective inter-chain constraints govern long-timescale relaxation. Specifically, wCFM trains a continuous-time velocity field across multiple temporal windows, while orthF introduces a lightweight geometric regularizer that incorporates local inter-chain constraints without explicitly computing topological quantities during training. Together, these components enable efficient generation of long-timescale polymer trajectories while improving agreement with MD in displacement distributions, structural fidelity, and topology-related observables across nine polymer systems and three dynamics formulations (Tables 2–4). The window-size ablations (Tables 13–14) and 50 ns extrapolation experiment (Table 12) further support their effectiveness across temporal scales. Thus, our novelty lies in demonstrating that multi-scale generative trajectory learning combined with transferable geometric regularization can improve modeling of entanglement-related polymer dynamics over extended timescales.

We appreciate the reviewer's feedback, which has helped clarify our contributions and strengthen their empirical justification. We hope our responses address the concerns and encourage reconsideration of the assessment.

Comments on the paper

1 comment
Area_Chair_NJNgOpenReview ↗

Requesting reviewer engagement

Dear Reviewers,

Thank you for your reviewing service. The authors have provided detailed rebuttals to the comments. Please confirm that you have read them, state which of your concerns are or are not addressed, and whether that changes your recommendations. Thank you for your continued engagement.

Thank you,

AC

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.