acceptodds
Withdrawn from ICLR 2027

PI-NCA: Neural Cellular Automata that conserve by construction

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Current learned PDE emulators can be computationally heavy to simulate and are not always readily scalable. Neural cellular automata (NCAs) are therefore an appealing template: a small local rule shared by every cell mirrors the locality of a differential operator. Used directly, however, an NCA predicts unconstrained state increments; in our evaluation, its mass drifts by tens of units, and on Navier-Stokes it performs worse than predicting no change at all. We introduce Physics-Informed Neural Cellular Automata (PI-NCA), together with its multi-channel variant, MC-PI-NCA, to address this failure structurally. The Physics-Informed Neural Cellular Automaton (PI-NCA) makes the shared rule output a flux rather than an increment and advances the state by the flux's periodic discrete divergence, so the flux-divergence update enforces exact conservation of the field total, regardless of the learned weights. MC-PI-NCA extends the architecture to coupled systems, predicting one flux per field and conserving each field's total independently. Additionally, MC-PI-NCA can simultaneously predict multiple physical variables, such as temperature, pressure, and density, while applying a parameter-free projection to keep the predicted states within physically valid range. We compare both architectures with the Fourier neural operator (FNO), full and matched-size ResNet and U-Net surrogates, the standard NCA, and an identity floor on ten 2-D phenomena, using five to ten seeds and Holm-corrected paired tests. Ablations isolate the effects of the flux head and the projection and identify the regimes in which the conservation prior should not be used. On the shallow-water equations, PI-NCA and MC-PI-NCA are the best-performing models with the lowest relative L2 error while using and fewer parameters than the FNO.On Cahn-Hilliard, PI-NCA has a relative L2 error lower than the best baseline and lower than the FNO; on heat and advection-diffusion, the proposed models are second only to the FNO while using to fewer parameters. PI-NCA outperforms NCA on seven out of the ten evaluated phenomena with fewer parameters and trains faster than the FNO.Overall, PI-NCA demonstrates that conservation can be incorporated into compact NCA as an architectural property rather than a training objective, yielding competitive accuracy with significantly lower parameter and computational costs.

Primary area:
neurosymbolic & hybrid AI systems (physics-informed, logic & formal reasoning, etc.)
Keywords:
Neural Cellular Automata, Physics-Informed Neural Cellular Automata, Multichannel-PINCA, Physics-Informed Machine Learning, PDE Surrogates, Conservation Laws, Mass Conservation, Flux-Based Updates, Shallow-Water Equations, Cahn-Hilliard Equation, Navier-Stokes Equations, Fourier Neural Operator, Neural PDE Solvers, Conservation-Preserving Neural Networks, Parameter Efficiency, Differentiable Physics, Data-Driven PDE Modelling

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_v6f5OpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces PI-NCA, which employs a shared local network to predict fluxes rather than directly predicting unconstrained state increments. Through a fixed, periodic, discrete divergence operator, these fluxes are converted into state updates, ensuring the conservation of the total quantity of each field. An extended variant, MC-PI-NCA, adopts this same field-wise construction approach and is applicable to coupled systems. Furthermore, the paper proposes an optional clip-and-headroom correction mechanism, which enforces a chosen state interval while restoring total mass when feasible. The authors evaluate these models across ten 2D periodic PDE benchmarks, comparing them against NCA, FNO, ResNet, U-Net, low-capacity variants, and an identity baseline. Matched-backbone ablations show that the flux head improves heat but strongly harms a non-conservative reaction system, while the bounded correction improves the Cahn–Hilliard benchmark.

Critical strengths

The paper states a clean architectural guarantee and tests when that prior is appropriate. By keeping the matched-backbone network and altering only the output construction method (A4), the error for heat conduction problems decreases from 0.0339 to 0.00743; conversely, for the non-conservative Nagumo equation, the error increases from 0.0161 to 0.379. This provides compelling evidence that the conservation prior is indeed effective within the local backbone networks investigated.

Critical weaknesses

The experiments fail to clearly establish the incremental value of the NCA parameterization compared to existing conservation-based models. Although the authors mention the FINN model and note that A4 is merely an adjustment internal to the backbone network in the appendix, the main result matrix compares the conservative PI-NCA mostly with unconstrained alternatives. In the absence of a flux-based model built upon the FINN architecture—or similar experiments controlling for flux-output on other backbone networks—it remains unclear what the NCA architecture contributes beyond the mere enforcement of conservation. Furthermore, the bounded correction described in the paper bears a strong resemblance to existing clip-and-assured-sum capacity-redistribution in the equal-weight setting, yet this connection is not discussed. Consequently, the specific added value of this implementation, relative to established conservation techniques, remains unquantified.

Questions for authors
  1. Could you provide a comparison between PI-NCA and a representative FINN-style conservative model, using the same data, training budget, and boundary conditions for one or two relevant tasks? Please report the error, parameter count, activation memory footprint, training time, and throughput.
  2. Could you include an additional set of experiments using per-field mean-centered residuals? This approach preserves the global total for each field without enforcing local flux conservation. These experiments would help disentangle the contributions of local conservation and bound projection, distinguishing them from simpler correction methods.
  3. For the Cahn–Hilliard benchmark, what justifies treating [-1,1] as a hard constraint of the target dynamics rather than a constraint imposed by the clipped reference solver? Does the bound-projection advantage persist with an otherwise consistent reference solver that does not clip, or under a different physically justified potential or bound?
  4. The paper mentions A1 and A4, but the numbering logic is not explained, nor are A2 and A3 defined. Do A2 and A3 refer to related experiments that were omitted from the paper? If not, could you clarify the meaning of these labels and explain why A4 is discussed before A1 in the main text?
Reviewer_MLaMOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The authors develop methods for improving neural cellular automata (NCA) in PDE problems. The two proposed methods are using a flux rather than a residual, and applying a bound to the updates. The effectiveness of the method seems to be mixed, with improvements being concentrated on simpler systems.

Critical strengths
  • For certain systems, the method seems to work quite well (is very parameter efficient and has high accuracy)
Critical weaknesses
  • The proposed physics-informed methods seem to be specific to systems. For example, applying a bound harms performance in systems without a bound, and applying a flux correction harms performance in systems without conservation. Also, the systems where this method works well are the simplest. One of the worst systems is Navier-Stokes, which is probably the most important PDE considered (“on Navier–Stokes, the FNO and the U-Net lead, and the best proposed model ranks sixth”)

  • One of the leading motivations is not well-justified “Learned PDE emulators can be computationally expensive and difficult to scale.”. Neural surrogates are still orders of magnitude faster than numerical solvers (such as in weather for example), and inference is usually not the bottleneck for neural PDE surrogates (the bottleneck is usually stability, generalizability, etc.)

  • This may just be personal preference but the paper reads like an LLM has written it (and probably suggested many of the evaluations/analyses). Beyond stylistic choices, there many things that are just not explained, for example:

    • “Because teacher pre-seeding is disabled for this equation, the evaluation covers the spinodal and early-coarsening transient rather than the slow Ostwald-ripening regime.” What is teacher pre-seeding? And why does it favor one regime over another?
    • The “Reading” column in Table 4. What is going on there? It seems like it is one sentence split across multiple rows, but then each row corresponds to a different experiment?
Questions for authors
  • How is solver error measured? (For example, Line 369, Cahn-Hillard)
Reviewer_CrLKOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates how to structurally preserve the conservation laws of physical fields within compact Neural Cellular Automata (NCA). The authors replace the direct prediction of state increments with a formulation based on flux prediction and discrete divergence, extending this approach to multiple, mutually coupled physical fields that each satisfy their own conservation laws. Furthermore, they propose a parameter-free correction method that ensures both state-range bounds and total-quantity constraints are met, provided the target total quantity is feasible. The method is evaluated across ten categories of 2D dynamical tasks, with comparisons against baselines including NCA, Fourier Neural Operators (FNO), convolutional networks, and identity mappings. Additionally, ablation studies are conducted to analyze the roles of the flux-based structure and the bounded correction mechanism.

Critical strengths

PI-NCA and MC-PI-NCA incorporate a finite-volume flux-conservation mechanism into the local update rules of Neural Cellular Automata (NCA). By utilizing only 4,576 to 10,992 parameters, these models ensure the independent conservation of individual fields—as well as multiple coupled fields—throughout the entire lifecycle of initialization, training, and inference, without the need for additional soft constraints in the rollout loss. Furthermore, the parameter-free "headroom projection" (Proposition 3) simultaneously addresses state truncation and mass recovery. Experimental results on tasks governed by conservation laws (such as shallow-water and Cahn-Hilliard dynamics) demonstrate that embedding conservation and boundedness as structural properties of the local operators allows these models to achieve accuracy comparable to or better than large-scale baselines and to outperform standard NCA in 7 out of 10 test tasks.

Critical weaknesses
  1. The lack of comparative analysis across different backbones limits the ability to attribute performance gains to specific factors. The authors acknowledge that the concept of flux conservation is not novel, and that discrete divergence operators and bounded projections could potentially be applied to other backbones as well. A4 validated the benefits of flux heads on the same NCA backbone, but did not compare ResNet, U-Net, or FNO equipped with the same conserved head and applicable projection. Consequently, the current results are insufficient to determine whether the observed advantages in solving the shallow-water and Cahn–Hilliard equations are contingent upon the NCA backbone, or if other architectures could achieve comparable performance by simply adopting the same constraints.

  2. The incremental value over existing conservation-based methods, as well as the scope of applicability for the experimental conclusions, remain insufficiently addressed. Although the authors discussed structure conservation methods such as FINN and Neural Conservation Laws in related work, the experiments did not include these external conservation baselines. The main comparisons were models without structure-guaranteed conservation (standard NCA, ResNet, U-Net, FNO), so no additional value relative to existing conservation methods was explained. Furthermore, the primary results are based on a 48248^2 grid, and there is no evaluation of out-of-distribution generalization, long-term stability without explicit constraints, or resolution transfer (Appendix F). Existing accuracy tests incorporate truncation-based constraints, and the teacher targets for the Cahn–Hilliard equation also include a [−1,1][-1, 1] truncation, resulting in the fitted target itself containing truncated features; consequently, these results do not directly demonstrate stability in the absence of such constraints, nor do they establish applicability to unconstrained dynamics. Furthermore, due to memory limitations, the full-size ResNet reduced the batch size from 32 to 16, and the number of samples seen (19,200) under a fixed training step of 1,200 was only half that of other models (38,400). These factors affected the completeness and consistency of the experimental comparison.

Questions for authors
  1. Could the authors supplement the comparisons on representative conservation tasks (such as shallow-water, Cahn-Hilliard, and heat) with the following: (a) incorporating external conservation baselines such as FINN or Neural Conservation Laws to evaluate performance relative to existing conservation methods; (b) equipping ResNet-S, U-Net-S (and FNO) with the same flux head (Eq. 2/3) and bound projection (Eq. 5) to distinguish the contributions of the NCA backbone and the conservation output head?

2.Please explain the specific reasons why the experiments on "out-of-distribution generalization," "long-horizon stability with the guard disabled," and "resolution transfer"—as mentioned in Appendix F—could not be completed; if possible, please provide the results for these experiments. Additionally, for the heat, shallow-water, and Cahn-Hilliard cases, could you provide data on how the error evolves over longer rollout horizons—and the proportion of non-divergent trajectories—when the 1.25x divergence guard is disabled?

  1. Under the time steps and test distributions used in practice, what is the actual trigger frequency of the [−1,1][-1, 1] truncation applied by the CH teacher, and what is the resulting change in quality? Furthermore, how is the specific range for the projection determined?

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

Related papers

Discussion (0)

Sign in to comment.