acceptodds
Under review as a conference paper at ICLR 2027

Green-Routed Neural Operators: Physics Determines Where the Network Reads

40 views

Abstract

We identify a mismatch between the physical role of transport fields in many PDEs and their usual role in neural operators: PDEs use them to select read coordinates, whereas neural operators typically treat them only as input values. We address this mismatch with the Green-Routed Neural Operator (GRNO), which uses the governing equation to determine where latent features are sampled. A parameter-free equation adapter evaluates the diagnostic relation and constructs a departure map whose values are the read coordinates. A multiscale encoder-decoder combines centered and routed reads of latent features to learn the complete finite-time update. Across five two- and three-dimensional PDE systems, GRNO achieves the lowest mean final relative error on four under 40-step autoregressive evaluation and remains competitive on Keller–Segel. Fixed-weight route interventions reveal strong dependence on direction and spatial alignment in four systems, with weak dependence in Keller–Segel. In independently trained ablations, GRNO achieves lower mean errors than variants that supply the transport field only as an input feature, substitute a learned displacement for the equation-specified route, or apply the route with a spatial misalignment, across all five systems. It also substantially outperforms directly advecting the physical state and learning the remaining update, indicating that equation-specified read coordinates provide an effective structural prior for long-horizon PDE forecasting.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Neural Operator, Scientific Machine Learning, Partial Differential Equations, Physics-Structured Machine Learning, Long-Horizon Forecasting

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 4 reviews, above 85% of the others.
Reviewer_E34xOpenReview ↗
Rating
3
Confidence
4
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes the Green-Routed Neural Operator (GRNO) for long-horizon PDE forecasting. GRNO uses a parameter-free equation adapter to compute diagnostic carriers through Green inverses and construct departure maps for latent-feature sampling. A multiscale encoder–decoder combines centered and routed reads to predict the complete state update. Across five 2D and 3D PDE systems, GRNO achieves the lowest mean final relative L2L^2 error among the evaluated models on four systems. Fixed-weight route interventions and independently trained ablations examine route dependence and the predictive benefits of equation-specified routing.

Critical strengths
  • A novel and physically grounded perspective: The paper identifies an important mismatch: transport fields determine read coordinates in PDEs, whereas neural networks often treat them only as input features. By connecting Green inverses with characteristic-based routing, GRNO addresses this gap with a clear physical rationale, making the architectural design well motivated.

  • Thorough and balanced experimental analysis: The evaluation covers five PDE systems and combines fixed-weight route interventions with independently trained ablations to examine both route dependence and predictive benefit. I appreciate the explicit reporting and discussion of limited benefits on Keller–Segel. This helps clarify the method’s scope and limitations.

Critical weaknesses
  • Gap between the physical interpretation and the latent-routing architecture: The characteristic-based physical interpretation concerns transport of the physical state, whereas GRNO first encodes the state and then resamples learned multiscale features. In simplified notation, “transport then encode” yields E(q∘Ξ)E(q\circ\Xi), while “encode then transport” yields E(q)∘ΞE(q)\circ\Xi, these operations do not coincide. Although applying physics-informed operations to latent features is not uncommon, the practical impact of this discrepancy in GRNO remains unclear.
Questions for authors
  • Impact of the encoding–transport gap: Could the authors quantify this gap by freezing the trained encoder and comparing “transport then encode” with “encode then transport,” using the same departure map and consistent carrier handling? This could help assess their practical impact.

  • Keller–Segel and compressibility: Could Keller–Segel’s weak routing dependence relate to its carrier being generally non-divergence-free, unlike those in the other benchmarks? My hypothesis is that coordinate resampling emphasizes displacement, whereas density compression must be captured through the learned update.

  • Data efficiency: Does the physical prior enable GRNO to reach comparable accuracy with fewer training trajectories? How does its 64-trajectory budget compare with those in the baselines’ original papers or official implementations?

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

PDEs use transport fields to determine where the evolving state is read. This is a structural mismatch with current neural operator models, which only receive those fields as input values at fixed grid locations. GRNO addresses this mismatch by introducing a parameter-free equation adaptor. This adaptor can be used to evaluate known diagnostic relations and build departure maps. Its values serve as read coordinates for sampling latent features in a multiscale encoder-decoder architecture. The work evaluates 5 different PDE systems under a 40-step rollout.

Critical strengths
  1. The paper identifies a key value-vs-coordinate distinction grounded in PDE structures. It provides a generalized formulation of "diagnostically routed PDEs" and a clean fix to this family of problems.

  2. The paper is exceptionally meticulous in designing its experiments, benchmarking against 5 different systems spanning both periodic and bounded domains, 2D and 3D, and physical and phase space. The ablations and controls on factors like learned route, shifted router, and state advection with residual provide clear mechanism isolation that explains why the approach proposed in this paper works.

Critical weaknesses
  1. The method explicitly requires a fully specified diagnostic relation, and its coefficients and boundary conditions must be provided beforehand. This would limit the method's application to systems whose governing equation is completely unknown or only partially known.

  2. The margins inside several entries of the ablation tables are very small, and they are only the result at one seed. It's hard to tell whether this is a win for the method over alternative designs or just numerical fluctuation.

Questions for authors
  1. Can you report the results in Table 3, but with 3 seeds, matching the baseline comparisons? I want to confirm whether the margins are consistently reproducible across different seeds.

  2. Is there a way for your method to gracefully fall back in cases where the diagnostic relation is only approximate, or when the equation itself is not fully revealed?

1 comment

Thank you for your careful rebuttal, which has addressed my major concerns. I will maintain my current rating.

Reviewer_T7WhOpenReview ↗
Rating
4(+1)
Confidence
4(+1)
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper introduces the Green-Routed Neural Operator (GRNO) for transport-dominated PDEs, where diagnostic fields such as velocity or electric field determine how the solution moves through space. Instead of treating these fields only as values on a fixed grid, GRNO uses the governing diagnostic equation to compute a routing map, then samples latent features at both fixed and back traced locations. It recomputes this route during each rollout step and predicts the full state update with a U-Net-based architecture. Across five benchmark systems, GRNO achieves the best 40-step relative L2 error on four and ties on the fifth. Ablation and factorial experiments further suggest that the main benefit comes from using the diagnostic field to choose sampling locations, rather than simply adding it as another input feature.

Critical strengths

The paper identifies a specific mismatch between transport-dominated PDEs and standard neural operators: diagnostic fields determine not only feature values, but also where those features should be sampled. GRNO addresses this directly by using the governing diagnostic relation to construct physics-based routing. This is a simple, parameter-free modification with a clear connection to semi-Lagrangian methods and potential applicability across architectures.

The method is evaluated on five substantially different systems, including Vlasov–Poisson, with varied diagnostic operators, dimensions, and boundary conditions. GRNO achieves the strongest long-horizon error on most systems, and the matched rollout protocol and ablations provide meaningful evidence that the method is not limited to a single fluid benchmark. This breadth makes the contribution more credible and potentially useful beyond the particular experiments studied.

Critical weaknesses

The factorial and learned-route ablations, which are the main evidence that physics-based routing is beneficial rather than merely correlated with performance, rely on a single seed. Several reported differences appear comparable to the observed seed variation. In addition, the learned-route control is relatively weak, while the state-advection control uses low-order interpolation that may introduce substantial numerical diffusion. Consequently, the experiments do not cleanly distinguish the benefit of specified routing from the benefits of learned latent features or the weaknesses of the competing controls.

NIPS-AR and RieszNO-3D are author adaptations, yet at least one performs very poorly, making it difficult to interpret the baseline comparison. The paper also does not report parameter counts or per-step runtime under the matched protocol. More broadly, the method’s scope is limited to fully observed systems with known, inexpensive diagnostic inverses, and its robustness to approximate/coarsened carrier fields, additional data, or learned transport is not tested. These omissions weaken the claims about generality and practical advantage.

Questions for authors
  1. Can you add at least 2 more seeds to Table 3, at least for the 2×2 factorial and Learned route?
  2. Can you rerun State advection + residual with cubic (or exact spectral) interpolation of the physical state?
  3. What are the parameter counts and per-step forward times of ReViT, Flowers, NIPS-AR, and RieszNO under the matched protocol?
  4. How does GRNO behave when the carrier is computed from a much coarser state (e.g. 64² Poisson solve on a 128² Kolmogorov field)?
Reviewer_MQ9COpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper argues that, in many evolution PDEs, a known transport or diagnostic carrier should determine where a neural operator reads features, rather than being provided only as an input channel. The proposed Green-Routed Neural Operator (GRNO) uses a parameter-free equation adapter to compute the diagnostic carrier and a short-time midpoint departure map. A multiscale encoder-decoder then combines centered and equation-routed latent reads to predict the complete finite-time state increment. The adapter, carrier, and departure map are recomputed from every predicted state during autoregressive rollout.

Experiments cover Kolmogorov flow, dissipative SQG, Vlasov-Poisson, 3D Keller-Segel, and 3D porous convection. With four training rollout steps and 40-step free evaluation, GRNO has the lowest final relative L2 error on four of five systems. Fixed-weight route interventions and independently trained ablations also indicate that the fitted models use the equation-specified route, especially on the four systems where routing materially improves accuracy.

Critical strengths
  1. Clear problem formulation. The distinction between supplying a carrier as values and using it to determine read coordinates is concrete. The Navier-Stokes example and the diagnostically routed PDE definition make the proposed mechanism easy to understand.

  2. Coherent and simple interface. The adapter is parameter-free, separates the physical construction of read coordinates from the learned update, and is recomputed during rollout. The paper documents the diagnostic carrier and boundary rule for each task.

  3. Strong matched accuracy. In Table 1, GRNO reports final errors of 7.80, 2.34, 0.16, 0.37, and 3.34 (all ×10^-2) on Kolmogorov, SQG, Vlasov-Poisson, Keller-Segel, and porous convection, respectively. It is best on four tasks and competitive on Keller-Segel.

  4. Useful mechanism controls. Table 2 changes only the departure displacement while keeping weights fixed; errors increase dramatically on four systems. Table 3 independently compares carrier features, equation routing, learned routing, shifted routing, and state advection. These controls are more informative than an accuracy table alone.

  5. Appropriate caution about system dependence. The paper openly reports that routing has little effect on Keller-Segel and that Centered only is best in the corresponding ablation.

Critical weaknesses
  1. The standard baseline comparison is incomplete. Table 1 includes ReViT, Flowers, NIPS-AR, and RieszNO, but no plain FNO/F-FNO that receives the same state, carrier, context, and time channels. Transolver is cited but also absent. Flowers is a useful learned-warp control, and RieszNO provides a spectral/directional comparison, but neither fully replaces a standard FNO control for the central claim that the equation-specified route matters.

  2. Evidence for an inductive-bias benefit is narrower than the framing suggests. The intervention results show that the fitted GRNO checkpoints depend on their route, while the independently trained ablations show that equation-specified routing can improve prediction within this backbone. However, the interventions can also create rollout states outside the training distribution, so they are evidence of route reliance rather than a fully causal proof of the proposed decomposition. Moreover, the ablations are reported for seed 0 only. Few-shot curves, coefficient or initial-condition OOD tests, or training with an approximate/wrong diagnostic would better support claims about sample efficiency, robustness, and the value of hard-coding the route.

  3. The implementation is tied to regular grids and exact diagnostics. All datasets use Cartesian grids, with bilinear or trilinear off-grid interpolation. The paper does not test nonuniform or unstructured meshes, nor does it test an approximate diagnostic solve or a misspecified carrier. This is a reasonable scope choice, but the scope should be stated explicitly if the contribution is presented as a general neural-operator structural prior.

  4. The headline mechanism is system-dependent. Keller-Segel is the clearest counterexample: fixed-weight interventions barely change the error, and Centered only slightly outperforms the full model in the independently trained comparison. The paper acknowledges this, but the abstract and conclusion should emphasize that routing helps when the equation-specified departure materially changes the finite-time read, rather than implying a uniformly useful mechanism.

Questions for authors
  1. Can you report a plain FNO/F-FNO baseline with the same state, carrier, context, and time-step channels under the K=4, H=40 protocol? Was Transolver tested?

  2. How does GRNO behave with reduced training data, changed coefficients or initial-condition families, or an approximate/wrong diagnostic relation?

  3. Since the fixed-weight interventions can move the rollout away from the training distribution, do retrained zero-route or route-dropout controls give the same conclusion? What physical or timescale property explains the weak routing effect in Keller-Segel?

  4. Is the intended scope Cartesian grids with exact diagnostic relations, or is extension to irregular meshes part of the claim?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.