acceptodds
Under review as a conference paper at ICLR 2027

Parametric Koopman Neural Operators for Condition-Dependent PDE Dynamics

Abstract

Neural operators have emerged as powerful surrogates for learning evolution operators underlying PDE dynamics in complex dynamical systems. However, time-varying physical conditions, such as forcing and background fields, induce condition-dependent dynamics, posing the challenge of adapting evolution operators to changing conditions while preserving temporal composition across intervals. We introduce the **Parametric Koopman Neural Operator (PKNO)**, which learns condition-dependent Koopman dynamics in shared observable coordinates. We construct shared hybrid observables over finite field histories, combining prescribed physical quantities with learned supplementary observables to capture temporal context in condition-independent coordinates. On these shared coordinates, we learn a condition-to-propagator map that generates mode-wise spectral Koopman propagators, enabling ordered composition under time-varying physical conditions. We further condition propagator generation on the rolling predicted history and derive a rollout-error bound characterizing how rolling-history sensitivity amplifies local prediction errors. Across four PDE benchmarks, PKNO is competitive under standard fixed conditions and demonstrates substantial gains over conditional baselines in challenging time-varying regimes.

Primary area:
learning on time series and dynamical systems
Keywords:
Neural Operators, Koopman Operator Learning, Parametric Koopman Dynamics, Partial Differential Equations, Condition-dependent Dynamics, Physics-guided Learning, Long-horizon Forecasting

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 2 reviews, above 27% of the others.
Reviewer_9xRfOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces PKNO, a Koopman neural operator for PDEs whose physical conditions, such as forcing, change over time. A shared encoder maps the recent field history to fixed latent coordinates, and a network generates a linear propagator for each Fourier mode from the current condition and a summary of the predicted history. On Navier-Stokes with a time-varying forcing amplitude, PKNO has lower rollout error than conditional FNO and KNO, while on shallow water it is worse than conditional FNO.

Critical strengths

The motivation is clear: many physical systems change their dynamics over time, and most neural operators do not model this change explicitly.

Critical weaknesses

The test of time-varying conditions is limited. The only changing condition is a scalar amplitude in [0.7, 1.3] that scales a fixed forcing pattern, there is no test outside the training range, and on shallow water PKNO has about twice the error of conditional FNO (0.0416 vs. 0.0197, Table 1).

The main design idea is not what is actually used. The paper motivates PKNO by multiplying linear propagators in fixed latent coordinates (Eq. 6), but in practice each propagator also depends on the predicted history, and the history is encoded again at every step (Appendix A.4), so the model works as a standard step-by-step predictor.

The ablation does not show where the gain comes from. The history summary is not removed in Table 2, so the improvement over conditional KNO may come from this extra input rather than from the proposed observables or the conditioned propagator.

The comparison is not strong enough. Only two conditional baselines are used, one of which is built by the authors, closely related methods such as PK-NN (Guo et al., 2025) and CAPE (Takamoto et al., 2023) are not compared, and all results use a single random seed.

The theoretical results add little. Theorem 1 follows from compactness for any continuous map and gives no bound on the dictionary size, and Theorem 2 is a standard error bound whose constant is not measured.

Some results are difficult to interpret. Burgers is a single-step task without rollout, and the fRMSE values in Table 1 (e.g., 2850 for IKNO) are not consistent with the normalized definition in Eq. 13.

Questions for authors

What is the error when the history is encoded once and only the conditioned propagators are applied, as in Eq. 6?

How does PKNO perform without the history summary, and how does conditional KNO perform when it receives the same summary?

How does PKNO perform when the forcing amplitude is outside the training range, or when the forcing pattern or viscosity changes over time?

Why are many fRMSE values in Table 1 much larger than 1?

9 comments
AuthorsOpenReview ↗

Response to Reviewer 9xRf (Part 1 of 4)

Dear Reviewer 9xRf,

We sincerely appreciate your thoughtful and constructive comments. We have carefully considered each of your questions and provide detailed responses below. All changes are in green in the revised PDF.

W6: Some results are difficult to interpret ... & Q4: Why are many fRMSE ...

The large fRMSE values in the original Table 1 arise from normalization by the target energy in the high-frequency band, which can be nearly zero for smooth solutions. The fixed-condition Burgers benchmark evaluates one-step prediction across initial profiles, while multi-step results for Navier–Stokes and shallow water are reported in Table 6 in our response to AC and Table 2 in our response to Reviewer mDV2.

Theoretically, let Us,t(k)U_{s,t}(k) and U^s,t(k)\widehat U_{s,t}(k) be the Fourier coefficients of the target and prediction for sample ss at time tt. We have

fRMSE=∑s,t∑k∈H∣U^s,t(k)−Us,t(k)∣2∑s,t∑k∈H∣Us,t(k)∣2,\mathrm{fRMSE}=\sqrt{\frac{\sum_{s,t}\sum_{k\in H}|\widehat U_{s,t}(k)-U_{s,t}(k)|^2}{\sum_{s,t}\sum_{k\in H}|U_{s,t}(k)|^2}},

where $H=\{\mathbf{k}\in\mathcal K: \|\mathbf{k}\|_2/k_{\max}\geq 2/3\}$, with $k_{\max}=\max_{\mathbf{k}\in\mathcal K} \|\mathbf{k}\|_2$. Our numerical audit confirms that the target energy in this band can be extremely small, accounting for only 3.040×10−163.040\times10^{-16} of the total energy on fixed Burgers. Consequently, even small high-frequency prediction errors can produce large fRMSE values.

In the tables below, we normalize the high-frequency error by the total target energy, giving

RHF=∑s,t∑k∈H∣U^s,t(k)−Us,t(k)∣2∑s,t∑k∣Us,t(k)∣2.R_{\mathrm{HF}}=\sqrt{\frac{\sum_{s,t}\sum_{k\in H}|\widehat U_{s,t}(k)-U_{s,t}(k)|^2}{\sum_{s,t}\sum_k|U_{s,t}(k)|^2}}.

We will replace fRMSE with RHFR_{\mathrm{HF}} in Table 1 and update Eq. (13) in the revised manuscript.


W1: The test of time-varying conditions ... & Q3: How does PKNO perform ...

PKNO achieves the lowest Full L2L_2 among the evaluated methods when the forcing amplitude exceeds the training range, when the spatial forcing pattern varies, and when viscosity varies over time. See Table 7 in our response to AC for the data-generation settings.

Amplitude extrapolation and faster forcing paths. On Navier–Stokes at ν=10−3\nu=10^{-3}, we train on forcing amplitudes qn∈[0.7,1.3]q_n\in[0.7,1.3]. Table 4 in our response to AC reports the tests for qn∈(1.3,1.5]q_n\in(1.3,1.5] and for faster forcing paths with 16 control knots. PKNO's Full L2L_2 is 0.0180 in the extrapolation test and 0.0168 on the faster paths, compared with 0.0189 and 0.0180 for CAPE, the strongest baseline in both settings.

Time-varying spatial forcing. We replace the fixed profile q(t)f0(x,y)q(t)f_0(x,y) with fn(x,y)=∑j=18aj,nϕj(x,y)f_n(x,y)=\sum_{j=1}^{8}a_{j,n}\phi_j(x,y), where the coefficients aj,n∈[−0.5,0.5]a_{j,n}\in[-0.5,0.5] vary over time and the sine and cosine bases use wavevectors (1,1)(1,1), (1,2)(1,2), (2,1)(2,1), and (2,2)(2,2). Table 3 in our response to AC gives the full comparison. On held-out spatial profiles, PKNO reduces Full L2L_2 by 19.7% relative to Cond. FNO and 30.0% relative to PKNO-scalar, which receives only ∣an∣2\\|a_n\\|_2 under matched-norm conditions (Table 1).

Table 1. Rollout and high-frequency errors under time-varying spatial forcing (ν=10−3\nu=10^{-3}).

SplitModelFull L2L_2 ↓RHFR_{\mathrm{HF}} ↓
IDCond. FNO0.02220.0009
Cond. KNO0.02530.0045
PK-NN0.02950.0058
CAPE0.02320.0013
PKNO-scalar0.02400.0047
PKNO0.01950.0038
OODCond. FNO0.03050.0011
Cond. KNO0.03380.0050
PK-NN0.04120.0064
CAPE0.03240.0015
PKNO-scalar0.03500.0052
PKNO0.02450.0041

Time-varying viscosity. The Fourier diffusion term νnΔω^(k)=−νn∥k∥22ω^n(k)\widehat{\nu_n\Delta\omega}(k)=-\nu_n\|k\|_2^2\widehat\omega_n(k) motivates mode-wise conditioning, since changing viscosity affects different frequencies by different amounts. We test νn∈[0.7,1.3]×10−3\nu_n\in[0.7,1.3]\times10^{-3} with either fixed forcing amplitude q=1q=1 or independently varying νn\nu_n and qnq_n. PKNO lowers Full L2L_2 by 13.8% relative to Cond. FNO and 21.7% relative to Cond. KNO in the viscosity-only case, and by 22.9% and 29.3% under joint variation (Table 2).

Table 2. Rollout and high-frequency errors under time-varying viscosity.

ConditionModelFull L2L_2 ↓RHFR_{\mathrm{HF}} ↓
Varying νn\nu_n, fixed q=1q=1Cond. FNO0.02180.0010
Cond. KNO0.02400.0048
PK-NN0.02750.0053
CAPE0.02050.0014
PKNO0.01880.0039
Jointly varying νn,qn\nu_n,q_nCond. FNO0.02750.0010
Cond. KNO0.03000.0052
PK-NN0.03400.0058
CAPE0.02540.0016
PKNO0.02120.0043

Under the updated experimental settings and a consistent training budget across all methods, PKNO achieves a Full L2L_2 of 0.01944 on shallow water, outperforming four baselines. The full comparison is reported in Table 6 in our response to AC. We will add these experiments and protocol details to Sections 4.1–4.2 and Appendix A.


TO CONTINUE

AuthorsOpenReview ↗

Response to Reviewer 9xRf (Part 2 of 4)

W2: The main design idea ... & Q1: What is the error ...

Encoding the initial history once and composing the condition-dependent propagators as in Eq. (6) yields a Full L2L_2 of 0.03585 on the 40-step time-varying Navier–Stokes rollout (ν=10−3\nu=10^{-3}) and 0.0160 on fixed Navier–Stokes. The corresponding errors for complete PKNO (M6) are 0.01581 and 0.0150. See Tables 1 and 2 in our response to AC for the propagation variants and full metrics, and Table 3 below for the accompanying RHFR_{\mathrm{HF}} comparison.

(i) Re-encoding improves rollout accuracy. M1 initializes z0=Eη(h0)z_0=E_\eta(h_0) and composes z^n+1(k)=Kθ(k;cn)z^n(k)\widehat z_{n+1}(k)=K_\theta(k;c_n)\widehat z_n(k). Adding step-wise re-encoding while keeping the condition-only propagator (M3) reduces time-varying Full L2L_2 from 0.03585 to 0.01809 (49.5%). Adding latent-state adaptation without re-encoding (M2) reduces it to 0.02785 (22.3%).

(ii) History adaptation adds a gain at matched re-encoding. With identical re-encoding, adding the rolling history summary Sh(hn)S_h(h_n) (M3 to M6) lowers time-varying Full L2L_2 from 0.01809 to 0.01581 (12.6%) and RHFR_{\mathrm{HF}} from 0.001446 to 0.001305 (9.7%). On fixed Navier–Stokes, Full L2L_2 decreases from 0.0154 to 0.0150.

(iii) Mode-wise propagation improves on the matched nonlinear update. M5 replaces the Koopman matrix with a parameter-matched nonlinear latent update while keeping the encoder, decoder, inputs and step-wise re-encoding identical. Relative to M5, M6 lowers Full L2L_2 by 12.2% (0.01581 vs. 0.0180) and RHFR_{\mathrm{HF}} by 15.8% (0.001305 vs. 0.001550).

Table 3. Rollout and high-frequency errors for the Navier–Stokes ablations (ν=10−3\nu=10^{-3}).

ConditionVariantFull L2L_2 ↓RHFR_{\mathrm{HF}} ↓
Fixed NSM10.01600.0002
M30.01540.0001
M6 (PKNO)0.01500.0001
Time-varying NSM10.03590.0014
M20.02790.0012
M30.01810.0014
M50.01800.0016
M6 (PKNO)0.01580.0013
Cond. KNO0.02480.0021
Cond. KNO+H (M7)0.01850.0012

The practical closed-loop rollout decomposes exactly into the propagator product of Eq. (6) plus step-wise re-encoding corrections. Defining the one-step re-encoding residual as δn+1enc=Eη(hn+1)−KnEη(hn)\delta_{n+1}^{\mathrm{enc}} = E_\eta(h_{n+1}) - K_n E_\eta(h_n) with Kn=Kθ(cn,Sh(hn))K_n = K_\theta(c_n, S_h(h_n)), each closed-loop latent update satisfies zn+1=Knzn+δn+1encz_{n+1} = K_n z_n + \delta_{n+1}^{\mathrm{enc}}. Expanding this recursion over rr steps yields

zn+r=Kn+r−1⋯Knzn+∑j=0r−1Kn+r−1⋯Kn+j+1δn+j+1enc.z_{n+r} = K_{n+r-1}\cdots K_n z_n + \sum_{j=0}^{r-1} K_{n+r-1}\cdots K_{n+j+1}\delta_{n+j+1}^{\mathrm{enc}}.

The first term is the ordered product of the history-adaptive propagators along the realized trajectory, in the shared coordinates of Eq. (6). The second term accounts for the accumulated step-wise re-encoding corrections. We will include the propagation ablations in Section 4.2, incorporate this decomposition as Proposition 1 in Section 3.3, and clarify in Section 3.2 that Eq. (6) defines the shared-coordinate composition core while Section 3.3 describes its closed-loop realization.


TO CONTINUE

AuthorsOpenReview ↗

Response to Reviewer 9xRf (Part 3 of 4)

W3: The ablation ... & Q2: How does PKNO perform ...

As shown in Table 3, on time-varying Navier–Stokes (ν=10−3\nu=10^{-3}), dropping the explicit history summary from PKNO (M3) gives a Full L2L_2 error of 0.01809, which is still 27.1% lower than standard Cond. KNO (0.02480). Conversely, giving Cond. KNO a matched history-summary module (M7) reduces its error to 0.01852, and full PKNO (M6) lowers this by another 14.6% to 0.01581. Thus, while the history summary accounts for part of the improvement, PKNO still holds a clear rollout advantage when both models use the same history input.

Controlled ablations disentangle the explicit history summary from the hybrid observables and condition-dependent propagator. In PKNO, the rolling field history hnh_n enters through two pathways: the shared observable encoder zn=Eη(hn)=[gphys(hn); gηnn(hn)]z_n = E_\eta(h_n) = [g^{\mathrm{phys}}(h_n);\, g^{\mathrm{nn}}_\eta(h_n)] and the context summary sn=Sh(hn)s_n = S_h(h_n) supplied to the propagator generator. Table 3 gives the matched comparisons under the same 40-step rollout protocol. PKNO w/o H (M3) drops sns_n from the generator so that z^n+1(k)=Kθ(k;cn)z^n(k)\widehat z_{n+1}(k)=K_\theta(k;c_n)\widehat z_n(k) while retaining hybrid observables and step-wise re-encoding. Cond. KNO+H (M7) keeps the standard Cond. KNO dictionary and decoder but conditions its propagator on both cnc_n and an independently trained history summary sns_n matching PKNO's architecture, dimension, input history, and update frequency (z^n+1(k)=KθKNO(k;cn,sn)z^n(k)\widehat z_{n+1}(k)=K_\theta^{\mathrm{KNO}}(k;c_n,s_n)\widehat z_n(k)).

(i) PKNO retains gains without the summary. Compared with Cond. KNO, M3 combines hybrid observables with condition-dependent propagation Kθ(k;cn)K_\theta(k;c_n) and reduces Full L2L_2 from 0.02480 to 0.01809 (27.1%). It reduces RHFR_{\mathrm{HF}} from 0.0021230.002123 to 0.0014460.001446 (31.9%).

(ii) Matched history inputs preserve the rollout gain. Adding sns_n within PKNO (M3 to M6) lowers Full L2L_2 from 0.01809 to 0.01581 (12.6%). Adding condition and history-summary inputs jointly to the KNO baseline (M7) lowers its Full L2L_2 from 0.02480 to 0.01852 (25.3%). With matched (cn,sn)(c_n,s_n) inputs, M6 has 14.6% lower Full L2L_2 than M7 (0.01581 vs. 0.01852). M7 has lower RHFR_{\mathrm{HF}} (0.0011720.001172 vs. 0.0013050.001305 for M6). Together with Table 2 in the main paper, the comparisons show that the explicit history summary accounts for part of the gain.

Hybrid observables and history-conditioned propagation act through distinct mathematical pathways in our theoretical framework. Theorem 1 shows that a condition-independent dictionary Ψ\Psi and a continuous coefficient map A(c)A(c) can uniformly approximate one-step target observables (sup⁡c∈CΔt∥[KΔt[c]g1,…,KΔt[c]gp]⊤−A(c)Ψ∥L2(μ;Rp)≤ϵ\sup_{c\in\mathcal C_{\Delta t}}\|[\mathcal K_{\Delta t}[c]g_1,\dots,\mathcal K_{\Delta t}[c]g_p]^\top - A(c)\Psi\|_{L^2(\mu;\mathbb R^p)} \le \epsilon), motivating our separation between shared hybrid observables Eη(h)E_\eta(h) and condition-dependent propagation. For the closed-loop predictor Pc(h)=Dω(Kθ(c,Sh(h))Eη(h))P_c(h) = D_\omega(K_\theta(c,S_h(h))E_\eta(h)), taking the Fréchet derivative with respect to a history perturbation vv (with u=Kθ(c,Sh(h))Eη(h)u=K_\theta(c,S_h(h))E_\eta(h)) separates two distinct mechanisms:

DhPc(h)[v]=DDω(u)[Kθ(c,Sh(h))DhEη(h)[v]⏟observable pathway+(DsKθ(c,Sh(h))[DhSh(h)[v]])Eη(h)⏟propagator-adaptation pathway].D_h P_c(h)[v] = D D_\omega(u)\Big[\underbrace{K_\theta(c,S_h(h))D_h E_\eta(h)[v]}_{\text{observable pathway}} + \underbrace{\big(D_s K_\theta(c,S_h(h))[D_h S_h(h)[v]]\big)E_\eta(h)}_{\text{propagator-adaptation pathway}}\Big].

Accordingly, in the rollout error bound of Theorem 2 (∥xn+r−xn+rpred∥X≤ϵloc∑q=0r−1Γq\|x_{n+r}-x_{n+r}^{\mathrm{pred}}\|_X \le \epsilon_{\mathrm{loc}}\sum_{q=0}^{r-1}\Gamma^q), the sufficient Lipschitz constant decomposes as Γsuff=LDκLE+BzLQ\Gamma_{\mathrm{suff}} = L_D\kappa L_E + B_z L_Q with LQ≤LKLSL_Q \le L_K L_S. Removing sns_n in M3 zeroes out the second pathway (BzLQ=0B_z L_Q=0) while preserving the shared observable pathway (LDκLEL_D\kappa L_E). We will incorporate the M3 and M7 comparisons and this pathway discussion into Section 4.2 and Appendix B of the revised manuscript.


W4: The comparison is not strong enough ...

We added PK-NN and CAPE and used the same hyperparameter-search budget for each model. Table 2 in our response to Reviewer mDV2 reports mean ±\pm standard deviation over three seeds.


TO CONTINUE

Reviewer_mDV2OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes PKNO, a Koopman neural operator for PDEs where the physical condition (for example, forcing strength) changes over time. It uses one shared encoder for all conditions. The encoder combines Fourier coefficients of the field with extra learned channels. Only the Koopman matrix changes with the condition. A small network generates this matrix for each Fourier mode from the condition and a summary of the recent field history. Theorem 1 shows that one set of features that doesn’t depend on the condition is enough to approximate one step of evolution for all conditions. Theorem 2 gives a bound on the rollout error. Experiments cover Burgers, Navier–Stokes at two viscosities and shallow water, plus Navier–Stokes and shallow water with forcing that changes over time. PKNO is best on Burgers, and under time-varying forcing it beats conditional FNO and KNO on Navier–Stokes. But it is only comparable to the baselines on fixed-condition Navier–Stokes and weaker than them on shallow water.

Critical strengths

The paper extends parametric Koopman learning from finite-dimensional systems to PDE fields. It keeps one shared feature space and lets only the Koopman matrix for each Fourier mode change with the condition. That way the steps for different conditions can be applied one after another in the same space. Theorem 1 supports this design, and Theorem 2 bounds the rollout error. PDEs whose conditions change over time are an important and less studied setting.

Critical weaknesses

The experiments do not strongly support the method. Under changing conditions, PKNO only clearly helps on Navier–Stokes. On shallow water it is worse than conditional FNO and conditional KNO. Under fixed conditions it is only best on Burgers, which is a simple single-step task. The changing condition is also just one scalar (forcing strength), and results use one random seed with no error bars. The paper claims a general method, but the evidence supports it on only one problem. More test problems are needed, or the claims should be narrowed. The theory does not match the implementation. The paper’s motivation is that the matrices can be chained in a shared feature space (Eq. 6). But in the actual model, the matrix also depends on the model’s own past predictions, and the field is re-encoded at every step (like memory). So Eq. 6 is not what is actually run. The theory is used to justify the design, but it does not explain why the method works better than simpler conditional models.

Questions for authors
  1. Under fixed conditions, PKNO is not clearly better than the baselines on Navier–Stokes and shallow water. Why? When the condition does not change, what does the condition-dependent matrix add over a fixed KNO?
  2. PKNO uses a different learning rate and gradient clipping than KNO (Table 4). Were the baselines tuned equally? Can you report results over several random seeds?
  3. Please test more problems and other types of conditions, such as boundary conditions or spatially varying fields. Right now only a single scalar forcing is tested.
5 comments
AuthorsOpenReview ↗

Response to Reviewer mDV2 (Part 1 of 4)

Dear Reviewer mDV2,

We express our gratitude for your thorough review and insightful comments on our paper. We respond to each point below. All changes are in green in the revised PDF.

W1: The experiments do not strongly support the method ... & Q1: Under fixed conditions ...

Even when the physical condition cn=mc_n=m is constant, PKNO's propagator Kθ(k;m,sn)K_\theta(\mathbf{k};m,s_n) is conditioned on the rolling history summary sn=Sh(hnroll)s_n=S_h(h_n^{\mathrm{roll}}), allowing it to adapt along each trajectory rather than applying a single static matrix at every step as in a fixed KNO. Under fixed conditions, PKNO's gains are less pronounced on Navier–Stokes and shallow water because constant-forcing vorticity is already a complete Markov state (where specialized autonomous operators like iKNO excel) and truncated Fourier modes are less suited than wavelet bases (WNO) to sharp radial dam-break fronts; under time-varying conditions, however, our rerun benchmarks show that PKNO achieves the lowest mean errors across both Navier–Stokes regimes and shallow water.

Conditioning the propagator on the rolling state history captures latent dynamical context and improves accuracy even when external metadata are fixed. As motivated in Section 3.3, observed state histories carry indirect signatures of local dynamical tendencies and unresolved physical drivers ξn\xi_n in un+1=F(un,cn,ξn)u_{n+1}=F(u_n,c_n,\xi_n)—for instance, in shallow water where the full state is U=(h,hu,hv)U=(h,hu,hv) but only depth hh is observed, the history of hh reflects unobserved momentum (hu,hv)(hu,hv). To isolate this mechanism under fixed conditions (cn=mc_n=m), Table 1 compares three variants: M1 (fixed propagator without step-wise re-encoding), M3 (fixed propagator with step-wise re-encoding), and M6 (PKNO, which adds history-conditioned propagation Kn(k)=Kθ(k;m,Sh(hn))K_n(k)=K_\theta(k;m,S_h(h_n))). Moving from M3 to M6 consistently reduces error on both fixed Navier–Stokes (ν=10−3\nu=10^{-3})—lowering Full L2L_2 by 2.6% (0.0154→0.01500.0154 \to 0.0150)—and fixed shallow water (Full L2L_2 from 0.01220.0122 to 0.011710.01171), confirming that history-conditioned propagation adds predictive value beyond fixed metadata and field re-encoding.

Table 1. Fixed-condition ablation of re-encoding and history-conditioned propagation.

DatasetVariantPropagation mechanismRe-encodingStep L2L_2 ↓Full L2L_2 ↓nRMSE ↓
NS 10−310^{-3}M1K(k;m)K(k;m)No0.01550.01600.0162
M3K(k;m)K(k;m)Yes0.01500.01540.0156
M6 (PKNO)K(k;m,Sh(hn))K(k;m,S_h(h_n))Yes0.01460.01500.0152
SWEM1K(k;m)K(k;m)No0.01270.01280.0129
M3K(k;m)K(k;m)Yes0.01210.01220.0123
M6 (PKNO)K(k;m,Sh(hn))K(k;m,S_h(h_n))Yes0.01160.01170.0117

(i) Comparable fixed-condition NS accuracy. In autonomous 2D Navier–Stokes, the observed vorticity field and constant viscosity/forcing completely determine the evolution, so trajectory history provides only modest refinement over a fixed global operator. In Table 2, mean Full L2L_2 for PKNO and KNO is 0.01530.0153 vs. 0.01540.0154 at ν=10−3\nu=10^{-3} and 0.46570.4657 vs. 0.46580.4658 at ν=10−4\nu=10^{-4}. iKNO achieves the lowest mean Full L2L_2 at ν=10−3\nu=10^{-3} (0.01020.0102), and FNO leads at ν=10−4\nu=10^{-4} (0.46560.4656).

(ii) Mixed results on other fixed tasks. On fixed shallow water, localized radial dam-break discontinuities favor multi-resolution wavelet bases (WNO at 0.00420.0042) over truncated global Fourier modes (PKNO at 0.01200.0120). On fixed Burgers, where one-step operator mapping across varying initial profiles is tested, PKNO lowers mean Full L2L_2 by 33.8% relative to KNO (0.00510.0051 vs. 0.00770.0077).

Under time-varying physical conditions, PKNO achieves the lowest mean rollout error in both Navier–Stokes regimes and on shallow water. Table 2 reports our complete benchmark suite with the updated training setup on time-varying shallow water. PKNO reduces mean Full L2L_2 relative to Cond. KNO by 22.9% on Burgers (0.06480.0648 vs. 0.08400.0840), 33.5% on NS at ν=10−3\nu=10^{-3} (0.01610.0161 vs. 0.02420.0242), 27.0% on NS at ν=10−4\nu=10^{-4} (0.01760.0176 vs. 0.02410.0241), and 22.0% on SWE (0.01980.0198 vs. 0.02540.0254). PKNO also outperforms Cond. FNO, PK-NN, and CAPE on these tasks.

TO CONTINUE

AuthorsOpenReview ↗

Response to Reviewer mDV2 (Part 2 of 4)

Table 2. Main benchmark results (mean ± std over three seeds).

DatasetModelStep L2L_2 ↓Full L2L_2 ↓nRMSE ↓
Fixed physical conditions
BurgersFNO0.0093±0.00060.0093±0.00060.0125±0.0008
AM-FNO0.0492±0.00310.0492±0.00310.0429±0.0022
WNO0.2862±0.01540.2862±0.01540.3172±0.0185
KNO0.0077±0.00050.0077±0.00050.0100±0.0006
iKNO0.0094±0.00050.0094±0.00050.0134±0.0008
PKNO0.0051±0.00030.0051±0.00030.0081±0.0004
NS 10−310^{-3}FNO0.0175±0.00100.0176±0.00110.0177±0.0009
AM-FNO0.0340±0.00170.0339±0.00210.0356±0.0021
WNO0.2157±0.01420.2388±0.01530.2409±0.0128
KNO0.0149±0.00070.0154±0.00080.0156±0.0007
iKNO0.0096±0.00060.0102±0.00060.0103±0.0006
PKNO0.0149±0.00080.0153±0.00090.0155±0.0008
NS 10−410^{-4}FNO0.4049±0.02390.4656±0.02070.4702±0.0255
AM-FNO0.4733±0.02910.5277±0.03010.5227±0.0285
WNO0.7262±0.03520.7906±0.04020.8252±0.0507
KNO0.4200±0.02280.4658±0.02790.4980±0.0268
iKNO0.4030±0.02460.4839±0.02800.4981±0.0255
PKNO0.4096±0.02400.4657±0.02770.4683±0.0241
SWEFNO0.0112±0.00050.0114±0.00050.0110±0.0006
AM-FNO0.0223±0.00110.0230±0.00100.0232±0.0012
WNO0.0041±0.00030.0042±0.00030.0043±0.0003
KNO0.0075±0.00040.0079±0.00040.0076±0.0005
iKNO0.0046±0.00030.0046±0.00030.0045±0.0003
PKNO0.0114±0.00070.0120±0.00070.0120±0.0005
Time-varying physical conditions
BurgersCond. FNO0.1078±0.00580.1023±0.00530.1035±0.0049
Cond. KNO0.0872±0.00450.0840±0.00450.0845±0.0051
PK-NN0.1360±0.00810.1321±0.00700.1227±0.0067
CAPE0.0752±0.00320.0720±0.00370.0720±0.0038
PKNO0.0685±0.00340.0648±0.00320.0662±0.0033
NS 10−310^{-3}Cond. FNO0.0188±0.00090.0190±0.00110.0201±0.0010
Cond. KNO0.0245±0.00150.0242±0.00130.0255±0.0012
PK-NN0.0224±0.00120.0258±0.00120.0314±0.0020
CAPE0.0176±0.00090.0168±0.00100.0175±0.0008
PKNO0.0164±0.00080.0161±0.00080.0156±0.0010
NS 10−410^{-4}Cond. FNO0.0314±0.00160.0318±0.00160.0331±0.0020
Cond. KNO0.0236±0.00130.0241±0.00130.0251±0.0013
PK-NN0.0379±0.00190.0324±0.00160.0344±0.0017
CAPE0.0353±0.00160.0385±0.00240.0368±0.0020
PKNO0.0173±0.00100.0176±0.00100.0184±0.0010
SWECond. FNO0.0330±0.00210.0327±0.00170.0036±0.0003
Cond. KNO0.0264±0.00130.0254±0.00130.0027±0.0003
PK-NN0.0413±0.00230.0496±0.00300.0054±0.0003
CAPE0.0329±0.00210.0326±0.00200.0034±0.0003
PKNO0.0198±0.00110.0198±0.00100.0020±0.0003

We will incorporate these ablations, updated benchmark results, and the discussion of fixed-condition regimes into the revised manuscript.


W2: The changing condition is also just one scalar ... & Q3: Please test more problems ...

Thanks for the valuable suggestion. We have expanded our evaluation to spatially varying forcing fields, dynamic boundary conditions, OOD amplitude extrapolation, and faster temporal variations. PKNO consistently achieves the lowest rollout error across all conditional baselines (Cond. FNO, Cond. KNO, PK-NN, and CAPE).

Our formulation handles high-dimensional, spatially structured conditions directly in the spectral domain through condition-dependent propagators. Specifically, the update rule zn+1(k)=Kθ(k;cn)zn(k)z_{n+1}(k)=K_\theta(k;c_n)z_n(k) lets each Fourier mode kk adjust independently to the spatial structure encoded in cnc_n. In our original setup, varying only the scalar amplitude q(t)q(t) kept the spatial profile f0(x,y)f_0(x,y)—and thus the relative excitation across frequencies—fixed. We have tested spatially varying forcing fields an(x)a_n(x) and time-dependent boundary conditions bn(s)b_n(s). The full generation settings are summarized in Table 7 in our response to AC. The quantitative comparisons are reported in Table 3 and Table 4 in our response to AC.

(i) PKNO generalizes across spatial forcing and dynamic boundaries. On multi-mode forcing, PKNO reduces Full L2L_2​ over the strongest baseline (Cond. FNO) by 12.2% and 19.7% on ID and OOD profiles, respectively. On dynamic boundaries, the reductions over Cond. KNO are 23.0% and 30.3%.

(ii) Spatial patterns provide information beyond condition magnitude. We design PKNO-scalar, which retains the PKNO architecture but receives only the L2L_2​ norm of the condition vector. Under matched-norm profiles, PKNO receives the full spatial condition. Compared with PKNO-scalar, it reduces Full L2L_2​ by 18.8%/30.0% on ID/OOD multi-mode forcing and 20.0%/29.3% on ID/OOD boundary conditions.

(iii) Lower error under amplitude and rate shifts. PKNO has lower Full L2L_2​ than all baselines in both OOD settings, including mild amplitude extrapolation (q∈(1.3,1.5]q\in(1.3, 1.5]) and faster temporal dynamics.

We will incorporate these extended experiments and analyses into the revised manuscript.


TO CONTINUE

AuthorsOpenReview ↗

Response to Reviewer mDV2 (Part 3 of 4)

W4: The theory does not match ...

The practical closed-loop rollout does not discard Eq. (6); rather, it equals the ordered propagator product of Eq. (6) in shared observable coordinates plus step-wise re-encoding corrections that bound latent drift over long horizons (Proposition 1). Furthermore, PKNO outperforms simpler conditional baselines because structured mode-wise Koopman propagation in shared coordinates provides a stronger inductive bias than generic nonlinear transitions—reducing 40-step Full L2L_2 by 49.5% through step-wise re-encoding, by 12.6% through history-conditioned propagation, and by 12.2% over a parameter-matched nonlinear latent update (M5).

Extension of Theorem 1. Expanding the closed-loop latent recursion shows how step-wise re-encoding acts as a bounded correction to the open-loop product in Eq. (6). Whereas the open-loop rollout in Eq. (6) evolves hn→Eηzn→Kθ(cn)z~n+1→Dωx^n+1h_n\xrightarrow{E_\eta}z_n\xrightarrow{K_\theta(c_n)}\widetilde z_{n+1}\xrightarrow{D_\omega}\widehat x_{n+1} without re-encoding, our closed-loop implementation updates hn+1=Shift⁡(hn,x^n+1)h_{n+1}=\operatorname{Shift}(h_n,\widehat x_{n+1}) from x^n+1=Dω(Kθ(cn,S(hn))zn)\widehat x_{n+1}=D_\omega(K_\theta(c_n,S(h_n))z_n) and re-encodes (zn+1,sn+1)=(Eη(hn+1),S(hn+1))(z_{n+1},s_{n+1})=(E_\eta(h_{n+1}),S(h_{n+1})). Defining the one-step re-encoding residual as δn+1enc=Eη(hn+1)−KnEη(hn)\delta_{n+1}^{\mathrm{enc}} = E_\eta(h_{n+1}) - K_n E_\eta(h_n) with Kn=Kθ(cn,S(hn))K_n = K_\theta(c_n,S(h_n)), the latent update is zn+1=Knzn+δn+1encz_{n+1} = K_n z_n + \delta_{n+1}^{\mathrm{enc}}. Unrolling this recursion over r≥1r\ge 1 steps yields Proposition 1 (Path-wise decomposition of history-adaptive rollout):

zn+r=Kn+r−1⋯Knzn+∑j=0r−1Kn+r−1⋯Kn+j+1δn+j+1enc,∥zn+r−Kn+r−1⋯Knzn∥≤ρ∑j=0r−1κj,z_{n+r} = K_{n+r-1}\cdots K_n z_n + \sum_{j=0}^{r-1} K_{n+r-1}\cdots K_{n+j+1}\delta_{n+j+1}^{\mathrm{enc}}, \qquad \left\|z_{n+r}-K_{n+r-1}\cdots K_n z_n\right\| \leq \rho\sum_{j=0}^{r-1}\kappa^j,

whenever ∥Kj∥≤κ\|K_j\|\le\kappa and ∥δjenc∥≤ρ\|\delta_j^{\mathrm{enc}}\|\le\rho. Here, the first term is the ordered composition of history-adaptive propagators along the realized trajectory, while the second term captures step-wise re-encoding as an explicit correction to Eq. (6).

Separating shared coordinates, condition-dependent evolution, and history adaptation provides a structured inductive bias with controlled approximation error. Rather than learning an unrestricted joint field-to-field mapping, PKNO factors the predictor as Pθ(h,c)=Dω(Kθ(c,S(h))Eη(h))P_\theta(h,c)=D_\omega(K_\theta(c,S(h))E_\eta(h)) into three distinct roles: (i) keeping EηE_\eta and DωD_\omega condition-independent ensures that successive propagators KnK_n act on compatible latent coordinates without requiring cross-condition coordinate transforms; (ii) parameterizing Kn(k)=Kθ(k;cn,sn)K_n(k)=K_\theta(k;c_n,s_n) mode-wise allows distinct spatial frequencies to respond differently to changing physical laws; and (iii) conditioning on sn=S(hn)s_n=S(h_n) adapts the operator to trajectory context—providing finite-step temporal context in fully observed systems (Burgers, 2D NS) and capturing unresolved memory in partially observed systems (such as unobserved momentum (hu,hv)(hu,hv) in shallow water ∂th+∇⋅(hu,hv)=0\partial_t h+\nabla\cdot(hu,hv)=0). Formally, if the target one-step dynamics admit the factorization F⋆(h,c)=D⋆(K⋆(c,S⋆(h))E⋆(h))F^\star(h,c)=D^\star(K^\star(c,S^\star(h))E^\star(h)), uniform module errors (ϵE,ϵS,ϵK,ϵD)(\epsilon_E,\epsilon_S,\epsilon_K,\epsilon_D) with target latent bound MEM_E and sensitivity LK,SL_{K,S} yield the component-wise bound

∥Pθ(h,c)−F⋆(h,c)∥≤ϵD+LD[κϵE+ME(ϵK+LK,SϵS)].\|P_\theta(h,c)-F^\star(h,c)\| \leq \epsilon_D + L_D\big[\kappa\epsilon_E + M_E(\epsilon_K + L_{K,S}\epsilon_S)\big].

Furthermore, whenever two history contexts induce distinct updates K⋆(c,s1)z≠K⋆(c,s2)zK^\star(c,s_1)z\neq K^\star(c,s_2)z at the same (c,z)(c,z), no single condition-only propagator K(c)K(c) can reproduce both simultaneously, proving that history-conditioned propagation strictly extends condition-only Koopman models.

TO CONTINUE

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.