- Rating
- 2(+1)
- Confidence
- 3
- Informativeness
- 1
- Presentation
- 2
- Soundness
- 2(+1)
Read review
Paper summary
This paper studies trajectory accuracy, invariant preservation, and dissipative-rate fidelity in learned PDE dynamics. It compares direct direct one-step predictors with Trajectory Flow. It then introduces GENERIC Flow to incorporate prescribed conservative and dissipative structure, with additional functional-change supervision for dissipative rates. Experiments report stronger trajectory prediction for Trajectory Flow, improved invariant preservation at the expense of trajectory accuracy for GENERIC Flow, and partial recovery of dissipative fidelity through rate supervision.
Critical strengths
None
Critical weaknesses
-
The claimed advantage of continuous-time generators is insufficiently justified: The paper claims “We analyse why continuous-time generators achieve stronger long-horizon fidelity than the direct one-step predictors.” However, the argument in Sec. 4 does not establish this advantage. It provides error-propagation expansions but does not establish that generators yield smaller errors or weaker amplification than direct predictors. Importantly, a direct one-step predictors is not inherently restricted to one-step supervision, it can be trained on unrolled trajectories, with losses on intermediate and terminal predictions. Yet the local-error and Jacobian diagnostics in Table 1 cover only the basic direct predictors. The conclusion should be restricted to the tested configurations, not presented as an explanation of a broader model-class advantage.
-
The motivation and contribution are unclear: The paper moves from general claims about continuous-time generators to Trajectory Flow without sufficiently justifying why this specific training formulation should represent the broader model class. The subsequent GENERIC extension addresses physical constraints, but its connection to the preceding temporal-representation analysis remains underdeveloped. Moreover, the reported improvements in invariant preservation come at the expense of trajectory. These observations demonstrate trade-offs within the tested framework, not inherent incompatibility between the objectives.
Questions for authors
-
Clarify the central contribution: Please clarify the central contribution and the connection between continuous-time generators and trajectory, invariant, and dissipative-rate fidelity? The paper first argues for continuous-time generators, then selects Trajectory Flow, introduces GENERIC constraints, and reports the resulting fidelity trade-offs. However, the unifying research question remains unclear: how does the temporal-representation analysis motivate this particular model and its structural extension? Please clarify whether the primary contribution is a general explanation of dynamical fidelity, a physics-constrained modelling method, or an empirical investigation of trade-offs, and distinguish conclusions about continuous-time generators broadly from observations specific to the chosen framework.
-
Abnormal data in experimental results: Please clarify the reliability and consistency of the reported results: Table 2 reports catastrophic FNO rollout errors, including an actual nRMSE@100 of on Advection; The same CNN Trajectory Flow baseline has different nRMSE@100 values in Tables 2 and 3 after unit conversion, for example, versus on Heat. Please reconcile these differences and identify the runs and evaluation scripts underlying both tables.
-
Readability and proofreading: Table 2 uses text substantially smaller than the 10-point size specified in the ICLR template, making the numerical results difficult to read. Page 10 retains template instructions, such as “This section is required and does not count toward the page limit”. Additionally, the formula in the “Skew operator” box of Figure 4 contains an unmatched closing parenthesis. Additionally, please clarify whether Figure 4 was directly generated by AI tools.
7 comments
Response to reviewer kUds continued..
Weakness 1 continued..
As expected, multistep training gives up some one-step accuracy at reference states, since it optimises the error along the model's own rollout rather than the one-step error. To measure this effect directly, we also evaluate every model at the state it actually reaches during deployment. Starting from the true initial state, each model is rolled out for 100 steps on its own predictions. At the state where it ends up, we measure the rollout error, the local defect (one step of the model compared with one step of the exact PDE solver from the same state), and the Jacobian amplification.
| Backbone | Model | Rollout MSE | Local defect | Median | P95 | Frac. |
|---|---|---|---|---|---|---|
| CNN | One-step | 0.9977 | 1.0399 | 21.2% | ||
| CNN | Multistep | 0.9993 | 1.0210 | 43.6% | ||
| CNN | One-step + Jacobian | 0.9974 | 1.0157 | 17.1% | ||
| CNN | Multistep + Jacobian | 0.9986 | 1.0060 | 31.0% | ||
| CNN | TFM | 0.9959 | 1.0089 | 14.5% | ||
| FNO | One-step | 1.1211 | 2.1097 | 98.6% | ||
| FNO | Multistep | 1.3243 | 2.8286 | 100.0% | ||
| FNO | One-step + Jacobian | 1.0213 | 1.3268 | 89.5% | ||
| FNO | Multistep + Jacobian | 1.1388 | 1.4370 | 100.0% | ||
| FNO | TFM | 0.9993 | 1.2018 | 25.5% |
Three KdV trajectories diverge, so the exact solver cannot be restarted from these states. We keep them rather than drop them.
On these states, the local defect ranks the models in the same order as their long-horizon accuracy. For CNN the order is identical, and for FNO the only swap is between Multistep and Multistep + Jacobian, whose values are nearly equal. This shows how multistep training helps the direct predictors: it lowers the local defect on the states the model actually visits. Lower amplification alone is not sufficient; for example, FNO + Jacobian still diverges on three KdV trajectories. Both observations are consistent with Eq. 2, where long-horizon error depends on the two quantities together. In both settings, TFM has the lowest local defect of all models.
Both diagnostics are now included in the revised version (App. C), with a pointer from Section 4.1 (l. 220).
Weakness 2: The motivation and contribution are unclear: The paper moves from general claims about continuous-time generators to Trajectory Flow without sufficiently justifying why this specific training formulation should represent the broader model class. The subsequent GENERIC extension addresses physical constraints, but its connection to the preceding temporal-representation analysis remains underdeveloped. Moreover, the reported improvements in invariant preservation come at the expense of trajectory accuracy.
As described in the general comment, the unifying research question of the paper is how different temporal parameterisations and structural constraints affect learned PDE surrogates under autoregressive deployment, in both long-horizon trajectory accuracy and physical consistency. In the revised version, this question and the three connected questions below are stated explicitly in the Introduction (l. 050-062):
- How does temporal parameterisation affect long-horizon trajectory accuracy? (Sec. 4.1, Sec. 6.2)
- Does better trajectory accuracy also imply physical consistency? (Sec. 4.2, Sec. 5 and Sec. 6.3)
- Does preserving invariants and enforcing admissible dissipation also give the correct dissipation rate? (Sec. 6.4)
Why Trajectory Flow : The distinction we study in Section~4 is between directly learning the finite-step map and learning a continuous-time vector field , whose numerical integration gives the transition. TFM is very close to Neural ODEs in this respect. It learns the same kind of vector field and integrates it at inference. The only difference is in training, where TFM learns the vector field directly from observed transitions without differentiating through the solver (Related Work, l. 108-115; Sec. 3.2). So TFM is not an arbitrary choice. Its continuous-time formulation is exactly the property being studied. We also chose TFM because, within this model class, it is a strong representative. Continuous Flow Operators (CFO; Hou et al., 2026), which we cite in the Related Work (l. 114-115), compare TFM-based training against Neural ODEs and PDE-Refiner and report the strongest results among them.
Response to reviewer kUds Continued...
Weakness 2 Continued: These observations demonstrate trade-offs within the tested framework, not inherent incompatibility between the objectives.
We agree, and we would like to clarify that we never claim the objectives are mathematically incompatible. This is clearly not the case. The reference PDE solutions themselves have accurate trajectories, preserve the invariants, and dissipate at the correct rate, so all three objectives can hold at the same time. Our claim is that trajectory accuracy, invariant preservation and dissipative-rate fidelity are distinct notions of fidelity, and improving one does not necessarily improve the others (Introduction, l. 035-042; Conclusion, l. 478-480).
The question we study is therefore whether learned surrogates can satisfy these objectives together, and how this can be achieved. We study this systematically. Unrestricted TFM achieves strong trajectory accuracy but violates invariants (Fig.1). Adding hard GENERIC structure restores invariant preservation at some cost in trajectory accuracy (Table 3). Explicit rate supervision then partly recovers the dissipation rate (Sec. 6.4; Fig. 3). This is similar in spirit to APEBench (Koehler et al., 2024), which systematically studies how different emulator architectures and training choices handle different PDE dynamics. Our study does the same along the axes of temporal parameterisation and physical structure.
-
Felix Koehler, Simon Niedermayr, Rüdiger Westermann, and Nils Thuerey. APEBench: A Benchmark for Autoregressive Neural Emulators of PDEs. NeurIPS Datasets and Benchmarks Track, 2024.
-
Xianglong Hou, Xinquan Huang, and Paris Perdikaris. CFO: Learning Continuous-Time PDE Dynamics via Flow-Matched Neural Operators. ICLR, 2026.
Response to reviewer kUds
Thank you for your thoughtful review.
Weakness 1
The claimed advantage of continuous-time generators is insufficiently justified. Sec. 4 provides error-propagation expansions but does not establish that generators yield smaller errors or weaker amplification than direct predictors.
We would like to clarify the nature of the argument in Section 4 and why the error-propagation framework is important.
TFM and direct one-step predictors use very different parameterisations and training methodologies. TFM learns a vector field that is integrated numerically (Sec. 3.2), while a one-step predictor directly learns the finite-step map (Sec. 3.1). So we cannot directly compare their internal parameters to understand the effect of the time parameterisation.
However, since both models are deployed in an autoregressive setting, the error-propagation framework (Eq. 2) gives a standard way to compare them through the quantities that are responsible for long-horizon error. The framework identifies two terms: the local transition defect δk and the input Jacobian Jk of one deployed step. The input Jacobian is the Jacobian with respect to the state, not to be confused with the parameter Jacobian. Long-horizon error depends jointly on both terms (l. 194-195).
Next, we measure both terms directly on matched test states in Table 1. TFM has a much lower local defect for both backbones. It also has fewer expansive steps.
Finally, we give a mechanistic explanation of why continuous-time vector-field learners can have a better inductive bias (Eq. 3, l. 197-203). In TFM, the finite-step transition and its sensitivity both come from the same learned vector field, so learning the vector field correctly constrains both of them together. A direct one-step predictor only fits the endpoints, which leaves the map and its sensitivity less tightly coupled (l. 204-207).
The purpose of Section 4 is therefore not to prove that continuous-time generators are always better. It is to identify the quantities that govern autoregressive error, measure them, and give an explanation that is consistent with these measurements. In the revised version, we also extend these diagnostics to the strengthened direct predictors (App. C; pointer in l. 220), as discussed in our response to the next point.
Importantly, a direct one-step predictor is not inherently restricted to one-step supervision; it can be trained on unrolled trajectories, with losses on intermediate and terminal predictions. Yet the local-error and Jacobian diagnostics in Table 1 cover only the basic direct predictors.
We agree, and this is also the setting we study in the paper. In Section 3.1, we explicitly consider multistep rollout-aware training (with losses on all intermediate and terminal predictions of an M-step rollout) and input-Jacobian regularisation, to check whether the failures of direct predictors can be fixed purely through stronger training objectives. These strengthened predictors are motivated in the Related Work (l. 097-100), evaluated in Table 2, and discussed in Section 6.2. The strengthened predictors improve some settings, particularly with the CNN backbone, but do not consistently recover the long-horizon accuracy of TFM.
The purpose of Table 1 is to compare the two temporal parameterisations through the two quantities in Eq. 2, which is why it reports the basic formulation of each. The strengthened objectives target exactly these quantities: multistep training targets the error accumulated along the rollout, and Jacobian regularisation targets the input Jacobian.
Following the reviewer's suggestion, we have now measured the same quantities for every training strategy in Table 2.
The conclusion of Table 1 holds for every training strategy: TFM keeps a local defect several orders of magnitude lower than every direct predictor, for both backbones.