acceptodds
Withdrawn from ICLR 2027

Pseudo-DeepONet: Explicit Neural Operators for PDE Dynamics

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Neural operators for time-dependent partial differential equations are built in both implicit and explicit forms, yet rigorous guarantees that favor one design over the other for broad classes of problems have remained scarce. We introduce the Pseudo-DeepONet (PDON),, an explicit architecture that separates spatial representation from temporal evolution in latent space, and we establish an error analysis showing that this split design can achieve strictly smaller approximation error than standard implicit operator formulations under mild regularity and spectral-decay conditions typical of many commonly studied dissipative and transport-dominated systems. Empirically, a parity-controlled inductive-bias study shows that the factorisation encourages learning the underlying dynamics rather than mere snapshot reconstruction when encoder, decoder, and capacity are matched. Among recurrent state-space temporal backbones, a linear oscillatory core (LinOSS) performs best in our suite, and we supply a theoretical error-bound analysis explaining how its second-order oscillatory structure matches temporal structure in PDE trajectories and lowers approximation error versus alternative cores. We further demonstrate resolution-invariant evaluation via band-limited spectral transfer between discretisations. Across seven widely used PDE benchmarks spanning one-, two-, and three-dimensional spatiotemporal dynamics, PDON mostly improves strong operator-learning baselines. Finally, we report results on the large-scale reanalysis dataset ERA5, showing that the same explicit factorisation carries to complex real-world spatiotemporal fields.

Primary area:
learning on time series and dynamical systems
Keywords:
time series, physical system

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_3bgNOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

DeepONet represents the solution of a PDE as a function of space and time, which is implicit in time. On the other hand, there are also methods that represent time explicitly. This paper tries to verify the usefulness of the latter.

However, in practice, most neural operator research is already explicit in time. As far as I can tell from Figure 2 and Eq. (1), the proposed method is indistinguishable from a variant of the Fourier neural operator that performs rollout internally. The details of the internal network architecture are different, but I cannot see any novelty beyond the architectural change.

Moreover, many neural operators that perform rollout internally already exist, and the authors themselves cite them, for example RONOM and MemNO. It is fine to claim that "All of these are special cases of the PDON skeleton," but this is just taking the union of many prior works, and I do not think it counts as a contribution.

Using LinOSS for the rollout is novel. But, put the other way around, that is the only real modification, and there is a gap between the stated motivation (a framework that unifies existing work) and what the paper actually does.

Critical strengths

None.

Critical weaknesses

The most serious problem is the reported numbers.

  • First, in Table 3, four of the seven numbers for Causal PINN and LNO are identical. The remaining three differ only in the leading digit or the order of magnitude.
  • Table 6 is even more abnormal. For the proposed method, the standard deviation is exactly 0.06 times the mean (up to rounding error). For PINN it is 0.05 times, for Causal PINN 0.065 times, and for MemNO 0.07 times. The other methods also show the same exact pattern as far as I have checked.
  • FNO(d+1)-LinOSS in Table 1 and FNO (implicit) in Table 3 should be very different methods, but their numbers are identical except for some trailing digits.

Also, Lemma 2 is attributed to Yarotsky (2017), but I could not find the corresponding theorem or lemma in that paper.

Checking all the other theorems and lemmas is beyond what can reasonably be expected of a normal review.

I believe the problems above are, by themselves, sufficient to reject this paper.

Questions for authors

Are the weaknesses above caused by a bug in the code, AI hallucination, or fabrication?

Reviewer_XkiuOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes PDON for time-dependent PDEs: spatial snapshot -> evolving vector/tensor -> reshaped spatial field.

Critical strengths
  • The authors provide an honest assessment of where the method works (weakly nonlinear, dissipative, moderate-horizon regimes) and identify where it is expected to struggle (wave-like dynamics, long time horizons, and large-amplitude problems).
Critical weaknesses
  • Limited novelty, and the closest baselines are missing. The paper itself lists POD-DeepONet, L-DeepONet, RONOM, DeepOMamba, and MemNO as special cases of its proposed skeleton. The implicit-vs-explicit contrast is drawn mainly against vanilla DeepONet. Apart from MemNO and LaMO, these closest relatives are absent from Table 3.
Questions for authors
  • Could you report POD-DeepONet, L-DeepONet, RONOM, and DeepOMamba under the same POD encoder? How are mixed inputs packed into the input tensor?
Reviewer_PC3fOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper proposes Pseudo-DeepONet (PDON), an explicit neural operator framework that separates spatial representation from temporal evolution for time-dependent PDEs. The proposed architecture follows an encode–evolve–decode formulation, in which spatial fields are mapped into compact latent representations, advanced through a temporal operator, and reconstructed into physical solution fields. The framework supports multiple spatial representations, including flattened fields, Fourier coefficients, and POD-based representations, as well as different recurrent and state-space temporal models. The authors provide a theoretical analysis of approximation complexity, comparing a factorisable holomorphic function class with a broader joint Sobolev class. They further investigate the LinOSS architecture as an oscillatory temporal backbone and derive conditional trajectory and energy-error bounds. Experiments on seven PDE benchmarks, together with an ERA5 atmospheric forecasting task, demonstrate strong empirical performance, long-horizon prediction capabilities, and cross-resolution evaluation.

Critical strengths

The paper presents a flexible encode–evolve–decode framework that separates spatial representation from latent temporal dynamics. Although related factorised architectures already exist, PDON provides a coherent formulation that accommodates flat, spectral, and POD-based spatial representations together with interchangeable temporal backbones. This modularity enables systematic investigations of the corresponding inductive biases. Besides, the experimental evaluation is relatively extensive, covering seven PDE benchmarks involving transport, diffusion, reaction, and fluid dynamics, together with an ERA5 atmospheric forecasting task. The comparisons include multiple operator-learning architectures and temporal backbones, with additional investigations of computational cost, long-horizon prediction, and resolution transfer.

Critical weaknesses

Limited methodological novelty relative to existing factorised neural operators. The central encode–evolve–decode factorisation is closely related to existing approaches, including POD-DeepONet, Latent DeepONet, RONOM, MemNO, and DeepO-Mamba. Moreover, LinOSS itself is an existing state-space architecture rather than a newly developed temporal operator. Although PDON provides a unified formulation, the proposed architecture appears primarily to integrate established spatial compression and temporal modelling components. The current comparisons do not sufficiently isolate the incremental contribution of PDON against closely related factorised operator-learning methods under matched spatial representations, temporal backbones, and computational budgets. Ambiguities and inconsistencies in the experimental specification raise reproducibility concerns. Several benchmark specifications are inconsistent across the appendices. For instance, Appendix F.1 specifies a Burgers viscosity of (\nu=10^{-3}), whereas Appendix G.2 presents a unit diffusion coefficient; the spatial domain and initial time also differ. Similarly, the forced KdV horizon is specified as (T=4) in Appendix F.2 but (T=5) in Appendix G.2. The evidence for resolution invariance and real-world generalisation remains limited. The claimed resolution invariance is primarily demonstrated through band-limited interpolation and qualitative visualisations, without sufficiently detailed quantitative error comparisons across spatial and temporal discretisations. Such interpolation-based transfer does not, by itself, demonstrate discretisation-independent operator approximation. Furthermore, the ERA5 experiment uses a relatively short three-month dataset, and PDON-LinOSS reports a relative error of 0.04, substantially worse than the reported FourCastNet (0.008) and ClimaX (0.01) results. Although PDON performs well against several general-purpose temporal backbones, these findings do not fully substantiate its claimed competitiveness or robustness in real-world spatiotemporal forecasting. The theoretical approximation advantage is not established under comparable function classes. The central theoretical contribution is weakened by the comparison of approximation bounds over different function classes. Theorem 1 contrasts a minimax lower bound over a broad joint Sobolev class with an upper bound over a restricted factorisable holomorphic subclass. Consequently, the resulting separation does not establish that PDON is more efficient than implicit architectures when both are evaluated on the same structural class. Serious inconsistencies and insufficient manuscript validation. The manuscript contains numerous editorial, mathematical, and experimental inconsistencies that raise concerns about the reliability of the reported methodology. For example, Figure 7 incorrectly presents the oscillatory dynamics with a positive rather than negative stiffness term and reverses the composition order of the encoder and decoder operators. More seriously, the benchmark configurations in Appendices F and G disagree on essential PDE parameters, time horizons, and training dataset sizes. In particular, the forced KdV experiment is reported to use 27,000 training samples in Appendix F.2 but only 1,500 in Appendix G.2.

Questions for authors

refer to weakness

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

Related papers

Discussion (0)

Sign in to comment.