acceptodds
Under review as a conference paper at ICLR 2027

Composable Spatial Sequence Learning for Steady Flow Prediction

Abstract

There is current interest in neural surrogates for steady flow-field solutions to partial differential equations (PDEs). These surrogates are typically trained as global mappings from geometry and operating conditions to complete solution fields, but extrapolation to domains with substantially greater spatial extent remains challenging. This study investigates whether reformulating the learning problem as repeated local transitions can improve domain-length extrapolation, without increasing the number of model parameters or changing the model architecture. Chip-to-chip (C2C) learning represents a steady velocity-magnitude field as an ordered sequence of narrow slices taken perpendicular to the streamwise direction, termed chips, and learns the geometry-conditioned transitions from upstream chip states to downstream chip states. Repeated composition of such transitions allows the domain length to change through rollout depth rather than model architecture. The proposed C2C-MLP is compared with five non-C2C baselines: a Multilayer Perceptron (MLP), a Physics-Informed Neural Network (PINN), a Graph Neural Operator (GNO), a General Neural Operator Transformer (GNOT), and Transolver, on domains of one, two, and three times the training length and on four out-of-distribution (OOD) geometries. All baseline models are provided with the same initial velocity-magnitude profile and geometric information available to C2C. Across five independent training runs, the C2C-MLP achieves the lowest mean absolute error (MAE) across all three domain lengths; on the 3× domain, its MAE was 0.0679 m/s, a 39.5% reduction relative to the most accurate baseline. OOD performance is geometry-dependent: C2C-MLP achieved the lowest MAEs on the Ahmed body and airfoil cases, while its MAE remained within 5.4% of the best baseline on the periodic-hill and backward-facing step cases. These four geometries are widely used CFD benchmark configurations representing bluff-body, aerodynamic, and separated-flow problems. Overall, the results of this study highlight spatial composition as a promising strategy for surrogate models that must operate on domains extending beyond those represented during training, without retraining or changing the model architecture.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Scientific machine learning, PDE surrogate modeling, flow-field prediction, domain-length extrapolation, spatial compositionality, autoregressive rollout

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_SGHzOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

This study introduces a method to predict steady flow fields across domains with varying streamwise lengths. By dividing the spatial domain into a sequential series of local regions termed "chips," the authors use a trained transition model to estimate a chip's downstream state using its local geometry and upstream conditions. Running this reusable local model sequentially allows for flow field simulation across extended domains. The core premise is that generalizing to longer domains can be done by simply repeating this local transition function, bypassing the need to alter neural network architectures or map entire global geometries to fields. This method is benchmarked against standard global surrogate models across both in-distribution and out-of-distribution shapes. Additionally, the paper analyzes the difference between CFD-forced prediction (which feeds true upstream data into each step) and free rollout (which recursively propagates its own predictions).

Critical strengths

Simple and intuitive problem reformulation: Treating a steady streamwise flow as a spatial dynamical system provides a natural mechanism for predicting domains that are longer than those encountered during training. Useful separation of local and accumulated error: The CFD-forced versus free-rollout analysis distinguishes deficiencies in the local transition from errors introduced through recursive propagation. Transparent modeling choice: The use of a relatively simple transition network helps isolate the effect of the chip-based formulation from the effect of a more sophisticated backbone. Credible and nuanced OOD results: The method performs reasonably well most of the time. Potential relevance to engineering configurations. Many flows possess a dominant direction and reusable local geometric structures, making the proposed decomposition potentially useful.

Critical weaknesses

The main contribution is a decomposition and rollout strategy: The paper does not introduce a fundamentally new neural architecture or physical model. So, the main contribution is somewhat limited. Dependence on a dominant streamwise direction. The ordered chip representation may be unsuitable for flows with strong recirculation, branching transport paths, substantial reverse flow, or global pressure coupling. Velocity magnitude is an insufficient target for establishing physical fidelity: Accurate velocity magnitude does not imply accurate flow direction, mass conservation, pressure, vorticity, or integrated forces. Evaluation on the full velocity vector and pressure would materially strengthen the claims. No explicit enforcement of governing-equation structure: The model does not appear to enforce incompressibility, momentum balance, or interface flux consistency by construction. It is therefore possible for predictions to score well under the selected metric while remaining physically inconsistent. Missing recurrent and marching baselines: Global neural operators are useful comparisons, but recurrent neural operators, autoregressive spatial models, learned marching methods, and sequence or state-space models are closer to the proposed formulation. Sequential inference cost. Inference cost grows with the number of chips, and the reported implementation appears substantially slower than several global surrogates. Limited stability analysis. The forced-versus-free experiment demonstrates error accumulation but does not characterize the stability of repeated transitions or identify conditions under which rollout errors remain bounded.

Questions for authors

Questions that would be worth considering are:

How does the method represent the global or upstream influence associated with elliptic pressure coupling when information is propagated only downstream? Does C2C remain competitive when predicting both velocity components and pressure rather than velocity magnitude alone? How sensitive is performance to chip length, chip overlap, interface-state dimensionality, and interface placement? Can the authors compare C2C against an autoregressive neural operator or recurrent model trained using the same chip decomposition? How does prediction error scale with the number of rollout steps? Is the observed growth approximately linear, multiplicative, or problem dependent? What happens when a recirculation region or substantial reverse flow crosses several chip boundaries? Can the method predict integral engineering quantities such as pressure drop, drag, or lift accurately?

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

This work introduces Chip-to-Chip (C2C) learning for steady 2D flow in channel-like domains. The simulation domain is cut into thin cross-sections perpendicular to the main flow. A single MLP takes the velocity profile at chip i and the wall geometry of the next 10 chips. It predicts the velocity profile at chip i+5. A total of 5 baselines were used to compare. During inference, feeding true CFD states instead of predicted ones removes much of the error, so error accumulation in the rollout dominates. A target-offset study (n = 1, 3, 5, 7, 9) shows n = 1 is best with true states but worst in free rollout.

Critical strengths
  1. Quite an interesting and different approach to rolling out a field. It is a learned marching scheme where the transition rule is reused along the spatial dimension, not the time dimension.
Critical weaknesses
  1. The task is narrow and the operating condition is fixed. Only one inlet velocity and fluid per set. Nothing varies in Re, inlet profile/turbulence. A longer channel mostly repeats the same local physics, which is the easiest case for a local model. Hence, the paper did not show any advantage of using C2C method compared to other baselines. Also, other methods can predict the full-field domain in one step, in which the accumulation of error is minimal.
  2. Only velocity magnitude is predicted, with no direction, pressure, wall shear, Cd/Cl, or separation and reattachment length. Steady separated flow is elliptic, so information also travels upstream. A marching scheme with a 10 mm look-ahead cannot represent that in general. Fig. 4 shows vertical striping and blockiness in the C2C Ahmed-body and backward-facing step results, which break flow continuity and indicate non-physical behavior. The Ahmed wake looks qualitatively worse than Transolver's, even though C2C has the lower MAE.
Questions for authors
  1. How are the vertical stripes in Fig. 4 explained? Is there period-n structure from the interleaved chains, and what does a spectrum of the error along x show?
  2. What happens if a pointwise MLP or small CNN receives the same 11-chip geometry window with no recursion? And what if the baselines are trained on random sub-windows? Which part of C2C is responsible for the gain?
  3. What are the parameter counts of every model? How was baseline hyperparameter tuning done, and why does Transolver train for 200 epochs and the others for 800?
Reviewer_LYWiOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes Chip-to-Chip (C2C) learning, a compositional approach for predicting steady flow fields over spatial domains longer than those encountered during training. The method learns local mappings between successive 1D cross-sections and recursively composes them to predict flow over extended 2D domains. The authors compare C2C-MLP against five ML baselines using 2,100 CFD simulations and four out-of-distribution geometries. The authors also investigate error accumulation during recursive predictions.

Critical strengths

1. Interesting compositional formulation. I like the idea of learning reusable local mappings and composing them to predict solutions over extended domains. This is an elegant approach with potential applications beyond the specific CFD problem considered.

2. Reasonable numerical evaluation. The authors compare against several established ML architectures and investigate error propagation during recursive rollout.

Critical weaknesses

1. Unconvincing problem motivation and practical relevance. My primary concern is that the paper appears to address a problem that is not particularly challenging or important in conventional CFD. Extending flow simulations over longer spatial domains is routinely handled using established numerical approaches, including appropriate periodic or updated boundary conditions. Related approaches involving the composition of flow transformations have also been demonstrated previously (see, for instance, inertial microfluidics). While extrapolation beyond the training domain may be challenging for particular ML architectures, the authors do not establish why solving this specific problem is consequential for computational science. This feels somewhat like a solution in search of a compelling application.

2. Limited physical complexity and computational benefit. All demonstrations involve steady 2D flows, predicting only velocity magnitude rather than the complete velocity-pressure solution. These are computationally inexpensive problems, particularly with modern GPU-accelerated CFD solvers. Indeed, the reported C2C inference time is 2.575 seconds compared with approximately 10–13 seconds for CFD, a relatively modest speedup. To establish practical significance, I would expect demonstrations on realistic 3D flows, where computational costs and physical complexity make surrogate modeling genuinely valuable.

Overall, while I appreciate the compositional learning idea, I am not convinced that the application, physical complexity, or demonstrated computational advantages justify acceptance at ICLR.

Questions for authors

none

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.