acceptodds
Under review as a conference paper at ICLR 2027

CyPhy-STN: Cycle-aware Physics-inspired Message Passing for Multivariate Spatio-Temporal Forecasting

Abstract

Can a single physics-inspired architecture perform consistently across diverse time-series prediction tasks governed by different dynamics? We introduce the Cycle-aware Physics-inspired Spatio-Temporal Network (CyPhy-STN), a unified architecture for multivariate spatio-temporal forecasting. It combines cyclic temporal attention with parallel, mechanism-specific graph propagation inspired by continuity, viscosity, relaxation, advection, and reaction. Learnt channel- and regime-aware routing determines which physical variables enter each mechanism and how messages are propagated, allowing the same architecture to be applied across different domains without the need of a fixed governing equation or an architectural redesign. Across four traffic and atmospheric forecasting benchmarks, CyPhy-STN achieves lower cross-benchmark regret and average rank than state-of-the-art spatio-temporal models with a low parameter count.

Primary area:
learning on time series and dynamical systems
Keywords:
spatio-temporal forecasting, physics-inspired, message passing, gnn

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 4 reviews, above 27% of the others.
Reviewer_ELcxOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

CyPhy-STN jointly forecasts multiple variables at graph nodes. Cyclic attention relates node histories within each variable, followed by graph messages inspired by continuity, diffusion, relaxation, advection, and reaction. Physical roles determine message inputs, and edge gates weight their contributions. Experiments cover two traffic and two atmospheric datasets, with a reported advantage in average rank and relative regret.

Critical strengths
  1. The evaluation covers two traffic and two atmospheric datasets with five seeds and per-variable results. CyPhy-STN keeps the best average rank after low-scale channels are excluded (Table 16), so the ranking is not explained by those channels alone.
  2. The graph pathway helps on traffic. Removing all physics messages worsens normalized error on both datasets (Table 9), and Figure 3 shows approximately 22% and 29% lower last-step flow MAE than the temporal-only variant on PEMS04 and PEMS08.
Critical weaknesses

I am not convinced that the experiments establish the contribution of channel-aware routing. Table 6 broadcasts mechanism outputs to all latent channels on PEMS04, PEMS08, and KnowAir. Physical roles still determine the source channels, but these three datasets do not test selective output routing. Beijing is the only selective case. There, the full model is worse than the temporal-only variant at the default horizon: MAE 51.143 versus 49.561, and normalized MAE 0.621 versus 0.614 (Table 2). Figure 3 does show gains at later horizons. Those gains, however, do not tell us whether selective routing helps compared with broadcasting the same messages. Removing individual operators does not answer that question either.

The definition of routing is also unclear. The abstract says learned routing determines which physical variables enter each mechanism. Sections 3.2.3 and A.1 instead describe specified roles and target-channel maps, alongside learned edge gates. Fixed role assignments are reasonable, but the paper should identify what is actually learned. As written, the claimed novelty and the experiments do not line up.

Questions for authors
  1. Are the source-role assignments, regime groups, and target-channel maps in Table 6 fixed configuration, or does training change any of them? Is “learned routing” intended to refer only to the edge gates in Equation (10)?
  2. On Beijing, how does selective routing compare with broadcasting the same mechanism outputs, with the remaining architecture and training held fixed? Pollutant errors across lead times would show whether routing contributes to the later-horizon gains.
1 comment
AuthorsOpenReview ↗

Statement on Weaknesses and Questions

We want to thank the reviewer for the feedback and want to take the weaknesses as a chance for improvement.

W1 / Q2 – Contribution of selective routing

Routing has two parts, not clearly separated in the paper. We apologize for that.

  • Input routing: which observed variables enter each mechanism. Role-based on all four datasets.
  • Output routing: which latent channels receive the update. Selective only on Beijing.

We did not compare selective vs. broadcast output routing on Beijing. We no longer claim it as a contribution. Input routing is tested. On KnowAir, we varied the transported-quantity role, all else fixed (Tin=Tout=24T_{in} = T_{out} = 24, five seeds):

Role assignmentPooled MAEPM2.5 MAE
PM2.5 only32.69 ± 0.0918.74 ± 0.05
Restricted (8 wind-advected variables)32.97 ± 0.0918.79 ± 0.12
Full (12 variables, paper)31.95 ± 0.0518.92 ± 0.05

Results therefore change in a structured way.

Gap to temporal-only on Beijing. Physics raises the error over the first steps and lowers it later. It overtakes temporal-only from about step 10 (KnowAir) and step 25–30 (Beijing) (Fig. 3). Over 24 steps, this is neutral on KnowAir and negative on Beijing.

W2 / Q1 – What is learned? Only the edge gates?

Yes. Roles, regime groups and target-channel maps (Table 6) are fixed configuration. Training does not change them. Learnt are:

  • the edge gates (Eq. 10), weighting each mechanism per edge,
  • the message parameters: advection networks Φη,g\Phi_{\eta,g} and scales sgs_g, relaxation parameters θτ\theta_\tau and β\beta, reaction networks ΦR,g\Phi_{R,g}.

"Learnt routing" in the abstract was imprecise. The abstract and Sec. 3.2.3 are revised to avoid misunderstandings. We now claim the role-based mechanism is set, not learnt routing.

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

The paper proposes CyPhy-STN, an architecture for joint multivariate spatio-temporal forecasting that applies a different inductive bias to each axis of the problem. A Temporal Cyclic Multi-Head Attention module, which adapts TQNet’s periodic queries to graphs by treating nodes as tokens, models temporal structure separately within each channel. A stack of physics-inspired message-passing layers then models spatial and cross-variable interactions. These layers compute five mechanism-specific messages in parallel: continuity, diffusion/viscosity, relaxation, advection, and reaction. Some are specified analytically and others are learned, and they are combined through per-edge sigmoid gates. A hand-specified routing map assigns observed variables to physical roles (density ρ, velocity v, flux q) and determines which latent channels receive each mechanism’s output. The authors evaluate on two traffic datasets (PEMS04/08) and two air-quality datasets (KnowAir, Beijing), retraining six baselines for joint all-variable prediction. They report the best average rank and lowest geometric and worst-case relative regret across the four datasets, and they include ablations, a long-horizon study, and a Jacobian/Dirichlet-energy analysis of information propagation.

Critical strengths
  1. A clean and reusable abstraction. Separating domain-agnostic mechanisms from dataset-specific role assignments is a nice idea. In principle, it lets one architecture cover systems with quite different governing dynamics without committing to a single PDE. Evaluating the mechanisms in parallel with learned edge-wise gates, rather than splitting operators sequentially as in ADR-GNN, is a sensible and simple design choice. The mix of closed-form terms where the physics is known and small MLPs where it is not is also well motivated.
  2. Careful and unusually candid experimental practice. The paper uses five seeds, paired one-sided t-tests, and a same-seed rerun as a noise floor (Table 9), which is more rigor than is typical in this area. Selecting the input length per model on validation data avoids quietly handicapping baselines designed for short or long contexts. The full per-channel tables are useful. I also appreciate that the authors report results that cut against their own story: the temporal-only variant wins on Beijing, and Appendix A.4 shows that the cross-domain ranking shifts once low-scale channels are excluded.
  3. The propagation analysis is genuinely informative. The distance-resolved Jacobian study, together with the Dirichlet-energy ratios, gives a clear mechanistic picture. The physics stack is strongly local and contractive, and the long-range sensitivity of the full model enters through the temporal module. The w/o-temporal ablation backs this up convincingly. Separately, re-training all baselines in a joint all-variable setting is a modest but useful contribution to benchmarking in its own right.
Critical weaknesses

1. The central claim, that physics-inspired message passing contributes meaningfully, is not well supported.

  • On the atmospheric datasets, the physics component does not help. The temporal-only variant beats the full model on KnowAir MAE and on both Beijing metrics, and removing all physics operators improves Beijing σ-NMAE (Table 9: −0.0065).
  • On traffic, nearly all single-operator ablations fall within noise. The one operator with a consistent, significant effect is reaction, which is a learned node-wise MLP and arguably the least "physical" of the five terms.
  • Most importantly, there is no control that replaces the physics module with a generic message-passing layer (e.g., an MPNN or GAT) of matched parameter count. Without it, I cannot tell whether the traffic gains come from the physical structure or simply from adding a graph-propagation stage after a per-channel temporal model.
  • The paper repeatedly calls the mechanisms "interpretable," but it never examines the learned gates γ, for example which mechanisms remain active on which datasets. As it stands, the interpretability claim is asserted rather than demonstrated.

2. The physical modelling is loose, and the routing still requires domain expertise.

  • The relaxation term uses a Greenshields-style equilibrium, veq=max⁡(0,1−ρˉ/2)v_{eq} = \max(0, 1 - \bar{\rho}/2), taken from traffic flow theory. Applying it to temperature, pressure, or the K index is hard to justify physically, and the paper's reinterpretation as "state-dependent damping" concedes as much.
  • Diffusion uses 1/d1/d rather than the 1/d21/d^2 scaling of a discrete second derivative, and the continuity message is not in divergence (flux-difference) form. Each choice may be reasonable on its own, but together they weaken the "physics-inspired" framing.
  • As far as I can tell, the messages are computed from temporal means and trends of the raw inputs (Xˉphys\bar{X}_{phys}, ΔXphys\Delta X_{phys}) rather than from the evolving hidden state. If so, the module behaves more like physics-motivated feature engineering than propagated dynamics, which may also explain the strong locality observed in Section A.2.2.
  • On KnowAir, twelve heterogeneous variables, including total precipitation and boundary-layer height, are all assigned the role of "transported scalar," which is difficult to defend. More broadly, the routing tables are hand-designed per dataset. This sits uneasily with the claim that the method avoids expert-driven adaptation, and the paper includes no sensitivity analysis of the routing choices.

3. The scope of the evaluation is limited, and the cross-domain advantage is fragile.

  • The four datasets amount to two domains, and PEMS04 and PEMS08 are very similar. This limits how far the "cross-domain" conclusion can be generalized.
  • An average rank over four datasets has little statistical power. The margin over STAEformer is 2.00 vs. 2.25, and Appendix A.4 shows the ordering is sensitive to which channels enter the normalized aggregate.
  • All baselines were adapted by the authors. For Air-DualODE and AirDDE, the paper itself states that the adaptations are a poor fit for joint multivariate forecasting. This likely inflates their regret and, in turn, CyPhy-STN's relative standing.
  • Several natural baselines are missing: TQNet itself (on which TC-MHA is directly built), ADR-GNN (the closest related method), and strong channel-aware multivariate models such as iTransformer or PatchTST. STD-MAE is also excluded.
  • Efficiency is reported only as parameter count. Wall-clock training and inference time and memory would give a more complete picture, especially given the per-edge, per-mechanism computation.
Questions for authors
  1. Are ρˉ\bar{\rho} and vˉ\bar{v} computed on raw or normalized values? If they are normalized, what is the intended meaning of veq=1−ρˉ/2v_{eq} = 1 - \bar{\rho}/2 in those units? Relatedly, do the physics messages at layer ll depend on H(l)H^{(l)}, or only on input statistics plus the gate inputs? If the latter, what does stacking LL layers add beyond deeper gating?
  2. Could you add an ablation that replaces the physics module with a generic MPNN of matched size? Could you also compare against (a) broadcasting all mechanisms to all channels and (b) randomly permuted role assignments?
  3. What do the learned gates γ look like per mechanism and per dataset? Are any mechanisms effectively switched off after training?
  4. On the atmospheric datasets, the physics branch hurts the first 10–30 forecast steps and helps afterwards (Figure 3). Do you have an explanation for this, and have you considered horizon-dependent fusion? Relatedly, for PEMS, could you report results at the standard horizons 3, 6, and 12 rather than only the horizon average?
  5. A few points need clarification or correction:
    • The hyperparameter δ in Table 8 is never defined in the main text.
    • Section A.2.3 describes MASE as "Mean Absolute Percentage Error."
    • Equation (5) labels the term "diffusion," while the rest of the paper calls it "viscosity."
    • Table 10 uses three seeds, while the other experiments use five.
    • The notation for the cyclic embedding in Section 3.2 appears to be garbled.
4 comments
AuthorsOpenReview ↗

Statement on Questions and Weaknesses

Thank you to the reviewer for this very thorough feedback on our paper. We read it carefully and consider it for further improvement of the method.

W1a – Physics does not help on atmospheric data

We agree and have revised the central claim. Physics improves traffic. On air quality, its effect at the default horizon is neutral (KnowAir) or negative (Beijing).

  • Horizon: physics raises the error over the first 10–30 steps and lowers it afterwards (Fig. 3).
  • Roles: on KnowAir, PM2.5 as the only transported quantity gives the lowest PM2.5 error (W2d).

We no longer present physics as uniformly beneficial.

W1b – Traffic: single-operator ablations within noise, only reaction significant

Reaction is the least physically constrained term, this is stated in the text. Redundancy in the mechanisms doesnt mean they are ineffective. It means, that they can (partly) make up for each other during training. On PEMS04/08, removing all operators raises σ\sigma-NMAE by 0.0073/0.0061. Single-operator effects sum to only 0.0013/0.0007 (Table 9). Reaction alone does not explain it: removing all costs 0.0060/0.0048 more than removing reaction.

W1c / Q2 – Generic MP control, broadcast, random role permutation

We ran a first control: replacing the physics message functions with a learnt MLP (inputs, gates, routing fixed). Physics wins on traffic (PEMS04: 6.735 vs. 6.922, PEMS08: 5.233 vs. 5.341). The MLP wins on Beijing (49.734 vs. 51.143).

(a) Broadcast. Output routing is selective only on Beijing, the other three datasets already broadcast. We did not run the Beijing comparison. We no longer claim selective output routing.

(b) Random permutations. Not run. Instead, three physically motivated role assignments on KnowAir change results in a structured way (see W2d).

W1d / Q3 – Interpretability, gates γ\gamma

Interpretability. We do not claim it. The wording in introduction and conclusion was wrong. It now reads: "CyPhy-STN decomposes the graph update into a fixed set of physics-structured message types (continuity, diffusion, relaxation, advection and reaction), whose inputs are set by a specified role configuration and whose contributions are weighted by learnt edge-wise gates."

Gates. We now report mean γ\gamma per mechanism and dataset (over layers and five seeds, Appendix).

  • Near zero: continuity on PEMS04 (0.056), diffusion on KnowAir (0.025).
  • All others active (0.17–0.49).
  • Reaction has the largest gate on PEMS04 and KnowAir, second-largest on PEMS08.

Values are end-of-training batch snapshots. A gate is a weight, not a contribution. We use gates as a diagnostic only.

W2a – Greenshields relaxation applied to temperature, pressure, K index

Relaxation is indeed more motivated by traffic. For atmospheric groups, it has no physical counterpart. It damps wind-role inputs towards an equilibrium that decreases with the transported scalars. That's why we call it generic state-dependent damping. Removing it changes σ\sigma-NMAE by +0.0003 ± 0.0026 (KnowAir) and −0.0021 ± 0.0025 (Beijing), within seed noise.

W2b – Diffusion uses 1/d1/d, not 1/d21/d^2, continuity not in divergence form

Diffusion. Summed over incoming edges, the message is ∑j(vj−vi)/dij\sum_j (v_j - v_i)/d_{ij}: graph-Laplacian diffusion with weights 1/dij1/d_{ij}. Any positive weighting is valid on a graph. 1/d21/d^2 assumes a uniform grid. Messages are standardised per mechanism, so the choice only reweights short vs. long edges. 1/d21/d^2 was worse on three of four target channels (five seeds):

Channel1/d1/d1/d21/d^2
PEMS04 flow18.7118.90
Beijing PM2.537.3838.19
Beijing PM1045.8947.01
On Pems08 it was the same.

Continuity. The message is the inflow flux ρˉivˉi/dij\bar\rho_i \bar v_i / d_{ij}. It is not conservative. Traffic and atmosphere are open systems. Messages are gated into a latent state, not integrated as a PDE. We state both choices and call the module physics-inspired, not a discretised PDE.

AuthorsOpenReview ↗

continues the previous ocmment

W2c / Q1 – Messages from input statistics, raw or normalised, why L layers?

Inputs. Correct: messages use temporal statistics of observed variables (Xˉphys\bar X_{phys}, ΔXphys\Delta X_{phys}), not h(ℓ)h^{(\ell)}. This is deliberate. It keeps inputs physically identifiable (e.g. the observed wind or traffic speed). Messages from h(ℓ)h^{(\ell)} would be closer to generic message passing. We now describe the module as gated injection of physics-structured messages. This matches the locality in A.2.2.

Normalisation. RevIN is applied only on the temporal path. The physics statistics are computed on the standard-scaled inputs, i.e. z-scored with the training-set mean and standard deviation per variable. In these units, veq=max⁡(0,1−ρ/2)v_{eq} = \max(0, 1 - \rho/2) is a relative Greenshields-shaped profile. The equilibrium velocity decreases linearly with density and vanishes two standard deviations above the training mean. It is not a calibrated fundamental diagram, this is now stated.

Layers. Messages are fixed across layers. Extra layers add re-weighting through the edge gates (which depend on h(ℓ)h^{(\ell)}) and repeated node updates.

W2d – Twelve KnowAir variables as "transported scalar", routing hand-designed

ρ\rho marks what continuity and advection treat as transported, not a physical density. Assigning all twelve KnowAir variables is coarse: boundary-layer height and precipitation act through dilution and wet deposition, not transport. We compared three assignments (Tin=Tout=24T_{in} = T_{out} = 24, five seeds). Restricted = eight wind-advected variables: PM2.5, 2 m temperature, 2 m dewpoint, temperature at 925/950 hPa, relative humidity at 950/975 hPa, specific humidity at 950 hPa.

Role assignmentPooled MAEPM2.5 MAE
PM2.5 only32.69 ± 0.0918.74 ± 0.05
Restricted (8 variables)32.97 ± 0.0918.79 ± 0.12
Full (12 variables, paper)31.95 ± 0.0518.92 ± 0.05

The narrowest assignment helps the pollutant. More variables help the meteorological channels. This trade-off is reported in the revision.

Expert adaptation. Architecture and code are identical across datasets, roles are configuration, like a graph or time features. They still need prior knowledge, though we think the assignment is easy to make:

  • velocity (vv): describes motion along the graph,
  • transported quantity (ρ\rho): carried by that motion,
  • flux (qq): rate of transport,
  • none: stays on the temporal path.

W3a–b – Two domains only, average rank has little power

W3a. Two domains do not establish general cross-domain applicability, and PEMS04/08 are similar. The claim is narrowed: one architecture, configured by roles, performs consistently on traffic and air quality. More domains are future work.

W3b. We no longer use the rank gap to STAEformer as evidence. The claim rests on consistency: CyPhy-STN has the lowest worst-case regret (2.12%), with fewer parameters than STAEformer, DTRformer and HG-GFNO. The channel sensitivity in A.4 comes from two low-scale KnowAir channels (total precipitation, 950 hPa vorticity). Excluding them collapses the seed std (e.g. CyPhy-STN 0.0896 → 0.0038). Without them, CyPhy-STN stays first by average rank (2.25).

W3c – Baselines adapted by the authors

Adaptation adds uncertainty. But no established models predict all variables at all nodes on these benchmarks. Traffic models predict flow, air-quality models one pollutant. Adapting was the only option. We kept adaptations minimal, documented them, and release the code. Relative regret is measured against the best model per dataset, so weak baselines do not affect it. Air-DualODE and AirDDE affect only the average rank, and only where they win (AirDDE on KnowAir σ\sigma-NMAE).

W3d – Missing baselines: TQNet, ADR-GNN, iTransformer, PatchTST, STD-MAE

  • TQNet: it has no node dimension. Its graph adaptation is basically our temporal-only variant.
  • iTransformer, PatchTST: Both are graph-free. They would run on the flattened N⋅CN \cdot C variables, therefore we did not use them for comparison.
  • ADR-GNN: no official code, we feared that a reimplementation wouldn't be adequately reflecting the papers method.
  • STD-MAE: it uses single-channel masked pre-training. Extending it to joint channels would change the method extensively. We report its published PEMS04/08 flow results with the other univariate specialists.

W3e – Efficiency only as parameter count

Parameter counts alone are incomplete. We did not track memory usage. Runs were queued on a shared cluster with different GPU types, so the logged training times are not comparable. For runs on the same GPU type, training time was roughly in line with the parameter count. We could not repeat controlled timing runs within the discussion period.

AuthorsOpenReview ↗

Answer to Questions

Q4 – Physics hurts first steps on atmospheric data, horizon-dependent fusion, PEMS horizons 3/6/12

Explanation. Not verified, but consistent with two properties. Messages use window means and trends, not the latest observations, so they say little about the next steps. The physics stack acts as a local, smoothing correction (A.2.2). Short lead times need station-specific persistence, which smoothing blurs. Longer lead times gain from spatially consistent evolution.

Traffic horizons. No early cost. Lower flow MAE and σ\sigma-NMAE than temporal-only at every horizon (PEMS04 / PEMS08):

HorizonPEMS04PEMS08
3−2.6%−4.1%
6−4.0%−6.0%
12−6.8%−9.8%
92 (long)≈ −22%≈ −29%

Both metrics give identical relative reductions (σ\sigma-NMAE rescales single-channel MAE). Results are added to the appendix.

Fusion. We haven't considered it yet, but will keep it in mind for future work. Learnt per-horizon mixing of temporal-only and physics representations would exploit this pattern.

Q5 – Minor clarifications

All fixed:

  • δ\delta: added by accident, removed.
  • MASE: we fixed this error.
  • "Viscosity": we renamed this message to be consistent with other papers, there were two leftovers, we renamed to "diffusion".
  • Table 10: three seeds due to limited compute capacitiy
  • Cyclic embedding: we fixed this in the new pdf

Thank you for the detailed responses and for revising several of the paper’s central claims. I appreciate the clarifications. However, the revisions substantially narrow the contributions: physics is no longer presented as generally beneficial, the cross-domain generality of the approach is no longer strongly supported, and the interpretability claims have been withdrawn. Given these significant reductions in scope, along with the remaining limitations in experimental coverage and evaluation, I maintain my weak reject recommendation.

Reviewer_5SxxOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes CyPhy-STN (Cycle-aware Physics-inspired Spatio-Temporal Network), a unified framework for multivariate spatiotemporal forecasting across different physical domains. The architecture combines Temporal Cyclic Multi-Head Attention (TC-MHA), which captures periodic temporal dependencies independently for each physical variable, with physics-inspired graph message passing that models spatial and cross-variable interactions. The latter incorporates five mechanisms; continuity, diffusion/viscosity, relaxation, advection, and reaction—through parallel message propagation, adaptive edge-wise gating, and channel-aware routing. The main contribution is a structured approach that integrates analytical physical principles with learned dynamics, allowing the same architecture to be applied across different domains without specifying a fixed governing PDE.

The method is evaluated on four benchmarks spanning traffic forecasting (PEMS04 and PEMS08) and atmospheric forecasting (KnowAir and Beijing), comparing it against six adapted multivariate forecasting baselines. The model achieves the best average rank and lowest relative regret across the four datasets, with a relatively low parameter count. The results indicate that physics-inspired message passing is particularly beneficial for traffic forecasting and longer prediction horizons, although its advantages are less consistent on atmospheric datasets, where the temporal-only variant sometimes outperforms the full model.

Critical strengths

1. Unified Physics-Inspired Framework Across Multiple Domains

The paper proposes a coherent framework that integrates cycle-aware temporal attention with physics-inspired graph message passing, incorporating continuity, diffusion, relaxation, advection, and reaction mechanisms. Its ability to apply a common architecture to both traffic and atmospheric forecasting is a meaningful contribution, as these domains exhibit different spatiotemporal dynamics. The structured integration of physical inductive biases and data-driven learning provides a potentially useful direction for developing more generalizable spatiotemporal forecasting models.

2. Competitive Performance and Comprehensive Empirical Evaluation

The model demonstrates competitive performance across four datasets spanning two application domains, achieving the best average rank and lowest mean relative regret among the evaluated methods. These results provide evidence of the framework's cross-benchmark robustness and practical effectiveness, even though improvements are not consistent across all datasets.

Critical weaknesses

1. Inconsistent Benefits of Physics-Inspired Modeling The paper's central contribution is the integration of physics-inspired mechanisms to improve multivariate spatiotemporal forecasting. However, the experimental results do not consistently demonstrate their effectiveness. In particular, the temporal-only variant outperforms the full model on both atmospheric datasets (KnowAir and Beijing), while removing certain physical operators also improves performance in some cases.

2. Insufficient Validation of Physical-Process Coupling and Physical Consistency Although CyPhy-STN incorporates continuity, diffusion, relaxation, advection, and reaction mechanisms, these processes are computed independently and subsequently combined through learned edge-wise gating and message aggregation. The paper does not sufficiently demonstrate that this integration captures meaningful interactions among coupled physical processes or preserves consistency with the underlying physical principles. Compared with existing physics-guided approaches such as AirPhyNet and Air-DualODE, the conceptual advantages of the proposed aggregation strategy remain unclear.

3. Dependence on Domain-Specific Physical Knowledge and Unclear Generalizability Although the framework is presented as unified across physical domains, its implementation relies on manually specified physical roles, regime groupings, and channel-routing configurations. The paper does not provide a systematic procedure for determining these configurations in a new application or investigate sensitivity to alternative assignments.

4. Limited Comparisons with Recent State-of-the-Art Transformer- and LLM-Based Models Although the paper includes Transformer-based baselines such as STAEformer and DTRformer, its evaluation does not adequately cover recent state-of-the-art spatiotemporal forecasting architectures, particularly LLM-based approaches such as ST-LLM+, AirFormer. These methods are relevant to assessing the competitiveness of CyPhy-STN.

5. Insufficient Isolation of Methodological Contributions The proposed architecture combines cycle-aware temporal attention, graph-based message passing, multiple physics-inspired operators, and adaptive gating. Although component ablations are provided, they do not sufficiently establish whether the improvements arise specifically from the physics-inspired formulations rather than additional modeling capacity or generic learned graph interactions. Comparisons with parameter-matched conventional graph neural networks or message-passing architectures without physical inductive biases would help clarify the contribution of the proposed mechanisms.

6. Limited Evaluation Under Challenging Physical Conditions The evaluation primarily focuses on aggregate forecasting accuracy under standard benchmark settings. Unlike physics-guided approaches such as AirPhyNet, which investigates sudden pollution changes and sparse observations, CyPhy-STN does not provide comparable targeted evaluations of abrupt changes, extreme events, or missing and sparse observations. Such experiments are important for determining whether the proposed physical inductive biases provide meaningful robustness under challenging conditions.

7. Incomplete Coverage of Related Physics-Guided Spatiotemporal Forecasting Literature The paper does not sufficiently position its methodological contributions relative to existing physics-guided spatiotemporal forecasting approaches in traffic (STDEN, PI-STGnet) and air-quality prediction (AirPhyNet, TransNet). This omission weakens the positioning of the methodological novelty and makes it difficult to assess whether CyPhy-STN introduces fundamentally new physics-guided modeling capabilities.

Questions for authors
  1. The temporal-only variant outperforms the full CyPhy-STN model on both atmospheric datasets, and removing certain physics operators improves performance. How do authors explain these findings?

  2. How does independently computing the five physical operators and combining them through learned gates capture interactions among coupled physical processes? Can the authors provide evidence that the resulting dynamics exhibit meaningful physical consistency rather than generic learned graph interactions?

  3. How are physical roles, regime groups, and channel-routing configurations determined for a new application?

  4. Can the authors provide comparisons with more recent Transformer- and LLM-based forecasting methods?

  5. Given that different baselines use different historical input lengths, can authors provide controlled comparisons using identical input windows to establish whether the reported improvements are attributable to the proposed architecture?

3 comments
AuthorsOpenReview ↗

Statement on Questions and Weaknesses

We want to say thank you for the very structured and extensive remarks, weaknesses and suggestions. We appreciate the feedback!

W1 and Q1 We agree and have narrowed our claims. Physics helps clearly on traffic. On atmospheric data, it helps only at longer horizons. At the default horizon, temporal-only is better on Beijing (both metrics). On KnowAir, they tie in σ\sigma-NMAE (0.506 vs. 0.505); temporal-only is 1.1% better in MAE (32.13 vs. 32.47). We see three reasons.

Forecast horizon. Physics lowers the error only after the first steps. The full model overtakes temporal-only from about step 10 (KnowAir) and step 25–30 (Beijing) (Fig. 3). Over 24 steps, this averages to neutral (KnowAir) and negative (Beijing). Traffic shows the same pattern at larger scale: about 22% (PEMS04) and 29% (PEMS08) lower flow MAE at step 92. Likely reason: messages use the window mean and trend. They carry slow, spatially propagated information. The latest observations dominate the first steps.

Operators. Removing one spatial operator (advection, continuity, diffusion, relaxation) changes atmospheric σ\sigma-NMAE by at most 0.0022 (Table 9). All changes are within seed std, except a small gain without advection on Beijing (−0.0007 ± 0.0006). The one clear effect is reaction on Beijing: removing it lowers σ\sigma-NMAE by 0.0057 ± 0.0023, nearly all of the 0.0065 from removing all operators. Reaction is the only learnt, node-wise term, without spatial propagation. Beijing has only 12 stations, which gives this shared network little data.

Graph size. Beijing's 12 stations are almost fully connected (density 0.36, mean shortest path 1.69; Table 7). Long-range sensitivity comes from the temporal module. The physics stack acts locally (Sec. A.2.2). On such a small, dense graph, the physics stage adds little beyond attention across all nodes. On the sparse traffic graphs (diameter 53 and 25), it adds road-network structure.

W2 and Q2 We do not model coupling between processes explicitly. The five messages are computed independently. They interact only through the shared nonlinear node update and the edge gates. These interactions matter empirically. On PEMS04/08, removing all operators raises σ\sigma-NMAE by 0.0073/0.0061. The single leave-one-out effects sum to only 0.0013/0.0007 (Table 9), six and nine times less. Removing one mechanism is compensated by the others. Removing all is not. This is functional complementarity, not physical coupling. We do not claim the latter. We also do not claim physical consistency (conservation laws, governing equations). This is the main difference to AirPhyNet and Air-DualODE. They embed a diffusion–advection equation and integrate it in continuous time. They are more constrained but tied to one domain. CyPhy-STN trades physical fidelity for applicability: one mechanism library, configured by roles, runs on traffic and air quality. CyPhy-STN outperforms the jointly adapted Air-DualODE on all four datasets (Table 2).

W3 and Q3 CyPhy-STN relies on domain knowledge. The abstract overstated this and is revised. The goal is not to avoid domain knowledge but to reduce it to a reusable, variable-level form:

  • generic ST-GNNs use no physical structure;
  • AirPhyNet and Air-DualODE need a governing equation for one domain;
  • CyPhy-STN needs only a role per variable and reuses one mechanism library.

New application. For a new application, one can easily give each observed variable a role:

  • velocity: describes motion along the graph
  • transported quantity: carried by that motion
  • flux: rate of transport
  • none: stays on the temporal path.

Each velocity field forms one regime group. The cycle length is one day at the sampling interval. Ambiguous roles can be compared on validation data. No governing equation is needed.

Sensitivity. Three role assignments on KnowAir, all else fixed (Tin=Tout=24T_{in} = T_{out} = 24, five seeds):

Role assignmentPooled MAEPM2.5 MAE
PM2.5 only32.69 ± 0.0918.74 ± 0.05
Restricted (8 wind-advected variables)32.97 ± 0.0918.79 ± 0.12
Full (12 variables, paper)31.95 ± 0.0518.92 ± 0.05

The narrowest assignment is best for PM2.5. The full assignment is best overall. The model uses the role knowledge.

AuthorsOpenReview ↗

continues the previous comment

W4 and Q4 Baselines must run on all four datasets. We cover the main families: Transformer (STAEformer, DTRformer, 2025), neural operator (HG-GFNO), physics-guided (Air-DualODE, ICLR 2025; AirDDE, 2026) and linear (DLinear).

AirFormer. Its spatial attention needs station coordinates. PEMS04/08 only provide road distances. On Beijing, most of the 12 stations would fall into one region. We use the more recent Air-DualODE and AirDDE instead.

ST-LLM+. It relies on a large pretrained backbone, a different design from our small models. Tan et al. (NeurIPS 2024) found that removing the LLM from LLM-based forecasters does not hurt and often helps. We consider the comparison with LLMs for future work.

W5 see general statement, bulletpoint 3

W6 These settings matter for physics-guided models. A fair evaluation needs dedicated protocols applied to all baselines. We consider this out of scope and consider it for future work.

W7 All four are added to the related work.

  • AirPhyNet: Air-DualODE, which we compare against, extends its diffusion–advection formulation to open systems.
  • STDEN: predicts flows on road segments as gradients of a node potential. Our benchmarks have node signals. Its neural-ODE template is shared by Air-DualODE.
  • PI-STGnet: its physics is a Greenshields diagram between flow and speed, undefined for air quality. A Greenshields-type equilibrium is part of CyPhy-STN (relaxation).
  • TransNet: applies ADR-GNN's sequential splitting to PM2.5. It acts on a latent state for one target and needs coordinates and wind direction. We also examined the public TransNet code. It differs from the method described in its paper in central points. A comparison against this implementation would not reflect the published method, so we do not include it.

Q5 We considered using the same input length, but the input lengths of the original papers turned out to be the beneficial one for every model. Therefore, we chose those. For more info see the general statement.

Thank you for the detailed rebuttal and revisions. The responses have addressed several of my questions. But, several concerns remain.

  1. Domain-dependent benefits of physics-inspired modeling: The horizon-dependent improvements are interesting but do not fully establish the broader effectiveness of the proposed physical inductive biases.

  2. Insufficient justification of the five-operator decomposition: The additional results support the contribution of the combined physics pathway, but most individual operator removals have limited or inconsistent effects.

  3. Limited evidence for generalizability: Although the framework is demonstrated across traffic and atmospheric forecasting, the evaluation covers only two application domains.

  4. Experimental consistency and presentation issues: The rebuttal shows a KnowAir pooled MAE of 31.95 ± 0.05 for the full role assignment, whereas Table 2 reports 32.474 ± 0.093 for CyPhy-STN. Additionally, the revised manuscript still contains editing errors, such as “TQNet — pending” entry in Table 2.

Overall, the rebuttal and revisions have improved the paper, through more precise claims. However, due to the the remaining concerns, I maintain my original rating of 2 (Weak Rejection).

Reviewer_SNXnOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes CyPhy-STN, an architecture for joint multivariate spatio-temporal forecasting (all nodes and all physical channels are predicted together). The pipeline is RevIN $ \to $ TC-MHA (a temporal cyclic multi-head attention that extends TQNet's cyclic queries to graphs: per channel, nodes are tokens, the query comes from a learnable phase-indexed embedding) $ \to $ projection with static node and input-context embeddings $ \to $ several layers of physics-inspired message passing $ \to $ prediction head $ \to $ inverse RevIN. The message-passing layer computes five mechanisms in parallel on the same graph state (continuity, viscosity/diffusion, relaxation, advection, reaction), combines them with learned per-edge gates, and writes each mechanism's aggregate only to channels selected by a routing map that assigns channels to the roles $ \rho $ (density-like), $ v $ (velocity-like) and $ q $ (flux-like). The stated goal is a single architecture that does not commit to one governing equation and still works across domains.

Experiments use PEMS04/PEMS08 (traffic) and KnowAir (184 stations) and Beijing (12 stations) (air quality), five seeds, and baselines (DLinear, STAEformer, HG-GFNO, Air-DualODE, DTRformer, AirDDE) adapted to joint prediction. CyPhy-STN is best or near-best in $ \sigma $ -NMAE on three datasets (e.g. PEMS04 MAE 6.735 vs 6.780 for HG-GFNO; KnowAir MAE 32.47 vs 34.15–36.81 for external baselines), has the lowest average rank (2.00), mean relative regret (0.72%) and worst-case regret (2.12%), and uses fewer parameters than STAEformer, DTRformer and HG-GFNO. Ablations show that physics operators help on traffic (removing all of them raises $ \sigma $ -NMAE by 0.0073 / 0.0061 on PEMS04 / PEMS08) and that the benefit grows with forecast horizon ( $ \approx $ 22% / 29% at the last step), whereas on the atmospheric data operator effects are mostly within seed noise and the temporal-only variant is even better on Beijing (MAE 49.56 vs 51.14). Specialist univariate models remain stronger on the three targets where they are reported.

Critical strengths
  1. The paper uses five seeds, paired leave-one-out ablations with one-sided t-tests (Table 9), a "worst-case relative regret" to avoid a geometric mean hiding one bad result, a documented procedure for adapting baselines to joint prediction (including their limitations), and a sensitivity analysis showing how low-scale/intermittent channels affect the aggregate ranking (Tables 16–17, where CyPhy-STN still ranks first but with 2.25 instead of 2.00). It also openly reports unfavourable findings: no benefit from physics on the atmospheric data, Beijing worse than the temporal-only variant, and univariate specialists being stronger. This is valuable because the contribution is essentially an empirical question about which inductive biases help where, and it makes the results readable at face value.
  2. On PEMS04/08 the physics path gives a gain beyond seed noise and beyond the temporal path, and the gain grows steadily with horizon ( $ \approx $ 22% / 29% at the final step), with a parameter count below several strong baselines. Together with the Jacobian/over-squashing analysis, this gives a concrete picture of when message passing with physics-shaped messages is helpful (long horizons, strong spatial coupling) and when it is not.
Critical weaknesses
  1. The cross-domain claim is supported only on traffic. Physics modules matter on PEMS04/08 (where only reaction has a consistently detectable individual effect) but not on KnowAir or Beijing, where they are sometimes harmful (Beijing: full model 51.14 vs temporal-only 49.56 MAE). The evidence covers only two domains and four datasets.
  2. Many gaps to baselines lie within seed standard deviation (PEMS08 MAE; KnowAir $ \sigma $-NMAE 0.500/0.505/0.513; PM2.5 target MAE), so the advantage appears mainly in aggregated ranks.
  3. Input lengths differ: $ T_{in} $ is selected per model from {12, 24, 96}; CyPhy-STN uses 96 while STAEformer, DTRformer and AirDDE use 12 or 24. The effect on the comparison is unclear.
  4. Hand-specified priors: channel grouping, the role mapping (Table 6) and cycle lengths (288/24/8) are given manually. This sits uneasily with the abstract's claim that no fixed equation or architectural redesign is needed, and robustness to a wrong role assignment is not tested.
  5. Design claims lack direct evidence: there is no comparison of parallel gated composition against ADR-GNN-style sequential splitting, no ablation of edge gating or routing, and no generic message-passing control.
  6. Practical gaps: only parameter counts are reported (no time or memory); no out-of-distribution or missing-value robustness; the long-horizon study compares only with the temporal-only variant; scalability beyond a few hundred nodes is not discussed; the 70/10/20 vs 50/25/25 splits are not assigned to datasets.
  7. Presentation: the framework figure and result figures are not clear enough.
Questions for authors
  1. Sec. 5.1 says CyPhy-STN is top-two on KnowAir, but Table 2 omits AirDDE while A.4 text states CyPhy-STN is third before channel exclusion (fourth after). Does the claim hold with AirDDE included, and what are the Table 4 ranks under plain MAE?
  2. The abstract says routing is "learnt", but Table 6 shows R is hand-specified and mostly broadcast, so is any part of routing actually learned?
  3. KnowAir σ-NMAE seed std (about 0.09–0.10) appears to come from total precipitation and vorticity (Tables 14 and 17). Can the authors confirm this and report MAE ranks and seed std without these channels?
  4. Can the authors report validation results for all T_in ∈ {12, 24, 96} for every model (Table 5 shows only the selected value) to show how much of the advantage comes from the lookback?
3 comments
AuthorsOpenReview ↗

Statement on Weaknesses and Questions

Thank you for your remarks. In the following we want to address your concerns:

Weakness 1&2

We did not intend to claim the superiority among all datasets but proposed a method that has consistently good results among datasets. For more details, please see our general statement - bullet point 2. This is why we use the average rank and mean regret as main metric.

Weakness 3 and Question 4

see general comment - 5)

Weakness 4

We distinguish between the architecture, which is identical across all datasets, and a small amount of dataset metadata that it takes as configuration.

Cycle lengths. They are not tuned. 288, 24 and 8 each equal one day at the sampling interval (5 min, 1 h, 3 h). Using known periodicities is standard (CycleNet, TQNet). STAEformer and DTRformer encode the same daily period via time-of-day embeddings. A phase-indexed embedding of length WW represents any period dividing WW. On KnowAir, a weekly cycle (W=56W = 56) instead of a daily one (W=8W = 8) leaves PM2.5 MAE unchanged (+0.01±0.20+0.01 \pm 0.20, four paired seeds) and slightly lowers aggregate MAE (−0.96±0.66-0.96 \pm 0.66).

Role assignment. Roles change the configuration, not the architecture. All datasets use the same modules and code. This is analogous to specifying the graph or time features. It does require domain knowledge. The revised abstract now states this.

Robustness. We compared three role assignments on KnowAir, all else fixed (Tin=Tout=24T_{in} = T_{out} = 24, five seeds):

Role assignmentPooled MAEPM2.5 MAE
PM2.5 only32.69 ± 0.0918.74 ± 0.05
Restricted (8 wind-advected variables)32.97 ± 0.0918.79 ± 0.12
Full (12 variables, paper)31.95 ± 0.0518.92 ± 0.05
The narrowest assignment is best for PM2.5. The full assignment is best overall.

Weakness 5

Sequential splitting.

We thought about analysing this, however, with our current design, the comparison wouldn't be meaningful. The messages are computed from input statistics, not the hidden state. Their order therefore does not change them. We removed the parallel-vs-sequential contribution from the paper. A sequential variant with messages computed from the hidden state would be a different model; we consider it future work.

Generic message-passing control.

We replaced the five physics message functions with a learnt MLP. Inputs, edge gates and routing stay fixed.

DatasetPhysics messagesLearnt MLP messages
PEMS046.7356.922
PEMS085.2335.341
Beijing51.14349.734

Physics messages win on traffic. On Beijing, the MLP wins, and temporal-only remains best. Replacing the physics layer with a GCN gives:

DatasetHorizonMAE
PEMS041215.352 ± 0.031
PEMS081218.022 ± 0.026
KnowAir2471.462 ± 0.144
Beijing2457.895 ± 0.504

A GCN does not fit this architecture. A generic MPNN comparison is future work.

Routing.

Input routing (which variables enter each mechanism) is role-based on all datasets. Output routing (which latent channels receive the update) is selective only on Beijing. We did not test it against broadcasting, so we no longer claim it. Input routing is tested by the role-assignment analysis (W4).

Weakness 6 Efficiency. Parameter counts alone are incomplete. We did not track memory usage. Runs were queued on a shared cluster with different GPU types, so the logged training times are not comparable. For runs on the same GPU type, training time was roughly in line with the parameter count. We could not repeat controlled timing runs within the discussion period. We will report wall-clock time and memory on a single GPU type in the final version.

OOD and missing values. This needs its own protocol applied to all baselines. It is out of scope and named as future work.

Long-horizon study. It is an ablation of the physics path. It therefore compares only against the temporal-only variant.

Scalability. Scalability. Each physics layer computes one message per edge and mechanism. Its cost therefore grows linearly with the number of edges ∣E∣|E|, the number of mechanisms MM (five) and the hidden size dd. Larger graphs (e.g. LargeST) are future work.

Splits. Traffic uses 70/10/20, atmospheric data 50/25/25. This is now stated in the paper.

Weakness 7 We added 2 new figures, for more clarity.

AuthorsOpenReview ↗

Answer of posed questions

Question 1 We apologize for this inconsistency. AirDDE was omitted from Table 2 by mistake. It is now added.

  • MAE: CyPhy-STN is best among full models (32.47 ± 0.09 vs. 34.15 ± 0.52 for HG-GFNO), second only to its temporal-only variant. The claim holds.
  • σ-NMAE: CyPhy-STN is third (0.505 ± 0.090), behind AirDDE (0.495 ± 0.097) and STAEformer (0.500 ± 0.102). The gaps (≤ 0.010) are a tenth of the seed std, caused by two low-scale channels (Q3). Section 5.1 is corrected.

Question 2 No. Routing is specified in advance and fixed during training. Learnt are:

  • the edge gates γ\gamma (Eq. 10), conditioned on both endpoint hidden states;
  • the message parameters: advection networks and scales, relaxation parameters and β\beta, reaction networks. "Learnt routing" in the abstract was imprecise. The abstract and Sec. 3.2.3 are revised.

Question 3 Excluding total precipitation and 950 hPa vorticity collapses the std:

Modelσ-NMAE std (all → w/o)
CyPhy-STN0.0896 → 0.0038
STAEformer0.1021 → 0.0109
AirDDE0.0970 → 0.0025
HG-GFNO0.0874 → 0.0024
Temporal only0.0887 → 0.0010

The pooled MAE ranking without both channels is unchanged:

ModelPooled MAE w/o bothRank (all → w/o)
Temporal only35.715 ± 0.1221 → 1
CyPhy-STN36.105 ± 0.3052 → 2
HG-GFNO37.964 ± 0.3403 → 3
Physics GNN only38.838 ± 0.2994 → 4
DLinear39.614 ± 0.2505 → 5
DTRformer40.683 ± 1.7706 → 6
STAEformer40.928 ± 2.1537 → 7
AirDDE42.454 ± 0.6878 → 8
Air-DualODE49.637 ± 0.4099 → 9

The σ-NMAE ranking without both channels is in Appendix A.4 (CyPhy-STN: fourth).

Thank you for the detailed rebuttal and additional experiments. I appreciate the authors' efforts to address my concerns, particularly the new MLP-MPNN comparison, the role-assignment robustness analysis, and the clarification of the cross-domain claims. These additions strengthen the empirical evidence and improve the clarity of the paper.

I still have some reservations regarding the comparisons with strong external baselines. In particular, while I understand the rationale for selecting the best-performing input length for each model, the corresponding comparison across different input lengths remains incomplete, making it difficult to fully assess the impact of this choice on the reported results. Furthermore, the long-horizon experiments primarily compare against the temporal-only variant rather than strong external baselines, leaving the relative performance at longer horizons less clear.

Nevertheless, I acknowledge the additional evidence and the authors' efforts to address the concerns. I am therefore willing to raise my rating from 2 to 3, while maintaining the above reservations.

Comments on the paper

2 comments
AuthorsOpenReview ↗

General comment to all reviewers

We want to thank every reviewer for carefully reading our manuscript and the detailed feedback on our work. We intend to use this feedback to further improve our proposed method.

Reviewers highlighted the experimental rigor (five seeds, paired tests, same-seed noise floor), the open reporting of unfavourable results, and the propagation analysis. We would like to answer the main shared concerns and share our attempt to improve them. We want to make comments to the main points, that were raised by all or several reviewers:

1) What the cross-domain claim means.

We got the impression that we did not properly describe our cross-domain claim. We proposed a methodology which is consistent across do- mains. The top models per domain are highly tailored to their application, especially when using PDE/ODE approaches. We did not intend to claim, that we proposed a universally superior method for every single dataset. We developed an architecture that was designed for the traffic domain yet also achieves good performance on atmospheric datasets. We thus argue that our model architecture achieves consistently good results across both domains.

CyPhy-STN has the lowest mean (0.72%) and worst-case relative regret (2.12%). Every baseline falls behind on at least one dataset: STAEformer by 2.71% on Beijing, HG-GFNO by 3.71% on KnowAir, and the other four by more than 17%. Per domain, CyPhy-STN has the lowest mean regret of the seven compared models both on traffic (0.20%) and on atmospheric data (1.23%). On atmospheric data alone, its own temporal-only variant has a lower regret; over all four datasets, however, the temporal-only variant has a mean relative regret of 2.71% and a worst-case relative regret of 4.78%, clearly higher than the full CyPhy-STN.

We also agree with nkTa that an average rank over four datasets has limited statistical power. Each additional dataset is expensive in our protocol: all seven models have to be adapted and retrained for joint all-variable prediction, and every model and ablation is run with five seeds. We therefore evaluate two structurally different domains with two datasets each, and restrict the cross-domain claim to these domains. For further evidence of the robustness, we provide the average rank when low-scale channels are excluded (Table 16). CyPhy-STN remains first.

2) Where the physics messages help.

First, we would like to highlight, that our messages should not be understood as exact differential equations of the underlying physical system. Our messages are "Physics-inspired" and contain a significant learnable part, which makes them transferable among different datasets with different underlying governing equations. We narrow the claims to the evidence in the paper. On traffic, the physics messages help clearly. At the default horizon they improve all three channels on both datasets, with 2.4–6.4% lower MAE than the temporal-only variant (Tables 12–13). Removing them raises σ-NMAE beyond seed noise and the same-seed rerun (Table 9). The gain in flow MAE grows with the horizon to about 22% (PEMS04) and 29% (PEMS08) at step 92 (Fig. 3). On atmospheric data, the physics messages give no benefit at the default horizon. On Beijing, the full model’s MAE is higher than the temporal-only variant’s on all six pollutants, by between 1% (SO2 , PM10) and 3.5% (CO), and on both aggregate metrics. On KnowAir, the PM2.5 difference (0.3%) is within seed variation (Tables 14–15). At longer horizons, however, the physics messages become beneficial on all datasets. In the long-horizon study of the pollutant targets (Fig. 3), the physics model is worse over the first steps and overtakes the temporal-only variant from about step 10 on KnowAir and step 25–30 on Beijing. In both domains, the benefit of the physics messages thus grows with the forecast horizon.

3) Does the physics structure matter beyond generic graph propagation?

To answer this question, we compare our physics-inspired messages with a GCN instead of the physics-message-passing and with a message passing, where the messages are learned using MLPs. Unfortunately, using a (smoothing) GCN within our architecture gave very inferior results. We assumed it isn't well-suited within our specific architecture and thus do not report the results in our paper:

DatasetHorizonMAE
PEMS041215.352 ± 0.031
PEMS081218.022 ± 0.026
KnowAir2471.462 ± 0.144
Beijing2457.895 ± 0.504
The MLP-MPNN results are added to the paper. The physics messages give lower MAE on both traffic datasets (PEMS04: 6.735 vs. 6.922; PEMS08: 5.233 vs. 5.341). On Beijing the learnt messages are better (49.734 vs. 51.143), and the temporal-only model remains best there.

4) What is learnt in routing, and does selective routing help?

The channel routing is a specified configuration: the physical roles, regime groups and source and target channel maps (Table 6) are fixed per dataset and do not change during training. The learnt parts are the edge gates γi,j\gamma_{i,j} (Eq. 10), conditioned on the hidden states at both endpoints, and the parameters within the messages (advection, relaxation, reaction). “Learnt routing” in the abstract was imprecise. We apologise for the misunderstanding and have corrected it.

5) Input length

During development, we ran pre-experiments to select the input length for each model. Due to limited storage, these runs were deleted before the final experiments. We are currently rerunning the comparison for different input lengths, unfortunately, the runs didn't finish on time. We selected the best-performing configuration per model to avoid disadvantaging any baseline. STAEformer, for example, performs best when input and output lengths are equal.

6) Interpretability

We did not intend to claim that our method is interpretable. The wording in the introduction and conclusion was wrong. We meant the structure of the module: the graph update is decomposed into named, physics-structured message types instead of a single generic aggregation. This is a property of the design, not evidence that the trained model is interpretable. We have corrected the sentence to: “CyPhy-STN decomposes the graph update into a fixed set of physics-structured message types (continuity, diffusion, relaxation, advection and reaction), whose inputs are set by a specified role configuration and whose contributions are weighted by learnt edge-wise gates.”

7) Gate activations

We now report the mean gate activation γ per mechanism and dataset, averaged over layers and five seeds (Appendix). Two mechanisms are close to zero: continuity on PEMS04 (0.056) and diffusion on KnowAir (0.025). All other mechanisms remain active (0.17–0.49). Reaction receives the largest gate on PEMS04 and KnowAir and the second-largest on PEMS08. The values are taken from training-batch snapshots at the end of training. A gate measures the weight of a message, not its contribution, which also depends on the message magnitude. We report the gates as a diagnostic and draw no interpretability claims from them.

We attached a tex diff of our revised version to show our changes to the paper (CyPhy_STN_Revision_Diff.pdf).

We hope our changes and additional information address all your concerns. If so, we would be grateful if you would consider updating your assessment. If not, we are happy to clarify any remaining points.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.