acceptodds
Under review as a conference paper at ICLR 2027

Analytical Neural Operator (ANNO): Learning Reusable Analytical Structures for Time Evolution in PDE Families

Abstract

Neural operators have emerged as an efficient framework for learning solution operators of PDE families. For time-dependent PDEs, existing learning-based solvers typically learn temporal responses from transient trajectories, either by predicting full spatiotemporal solution fields or by learning recursive propagators. For systems governed by known PDEs, however, the temporal evolution is already constrained by the governing equations. We propose the Analytical Neural Operator (ANNO), which learns reusable analytical structures from steady-state problems and reuses them for time evolution within PDE families. ANNO explicitly represents the lifting function, spatial eigenmodes, and modal coefficients, and constrains the learned modes through the corresponding spatial eigenproblem. After steady-state training, the Learnable Mode Network and Learnable Coefficient Network are frozen and reused for time-dependent PDEs in the same family, while the dynamic modal coefficients are advanced solely according to the target PDE dynamics. This requires neither transient training trajectories nor network retraining. Across five experiments, ANNO accurately recovers steady-state solutions and spatial eigenstructure, and reuses the learned analytical structure for 1D diffusion, 2D wave, and 2D engineering thermal dynamics under nonzero initial conditions, spatiotemporally varying sources, time-varying boundaries, and varying physical parameters. On these transient tasks, ANNO achieves mean relative errors of 0.2531%, 1.8646%, and 1.0506%, respectively, while substantially outperforming fully supervised transient baselines trained on complete trajectories. Ablation studies further support the effectiveness and computational efficiency of the proposed analytical structure. These results show that reusable analytical structures learned from steady-state problems can support accurate and data-efficient time evolution within PDE families.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Scientific machine learning, Neural operators, Analytical solutions, PDE families, Time-dependent PDEs

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_WdkvOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

Neural operators learn mappings between function spaces to solve families of PDEs efficiently, but for time-dependent problems they usually learn the temporal evolution from full transient trajectories. This paper proposes the Analytical Neural Operator (ANNO), which learns the spatial eigenmodes and modal coefficients of the solution, not the solution field directly, and reuses them across a PDE family that shares the same domain, spatial operator, and homogeneous boundary conditions. ANNO is trained only on steady-state problems: it predicts the parameters of a known eigen structure, namely the frequencies and amplitudes of trigonometric modes and the steady-state modal coefficients, while an eigenstructure loss keeps the learned modes consistent with the true eigenproblem. The trained model recovers steady-state solutions and their eigenmodes accurately, and it is reused without retraining for 1D diffusion, 2D wave, and 2D engineering thermal dynamics, with mean relative L2 errors of 0.25%, 1.86%, and 1.05%, well below those of fully supervised transient baselines. An ablation study that removes the analytical lifting, replaces the trigonometric modes with an MLP, removes the frequency-band strategy, or drops the eigenstructure loss shows that each variant has higher steady-state or transient error, with the full ANNO performing best.

Critical strengths

The key strength of ANNO is that it learns the spatial eigenmodes and modal coefficients, not the solution field directly, so a model trained only on steady-state problems can be reused for the time-dependent equations of the same PDE family without transient data or retraining. Its eigenstructure loss constrains the learned modes to satisfy the underlying eigenproblem, which makes them true eigenmodes and not just a basis that fits the steady-state data, and this is what allows them to be reused correctly in time evolution.

Critical weaknesses

The main weakness of ANNO is that it learns only the parameters (frequencies and amplitudes) of a hard-coded trigonometric eigenmode form, not the form itself, so the eigenstructure must be known in advance and the architecture cannot represent eigenmodes for variable coefficients or irregular geometries. The evaluation is also limited to linear, constant-coefficient problems on intervals and rectangles, where the eigenmodes are closed-form or can be found by simple one-dimensional root-finding, so the classical modal solution is already cheap and exact and the benefit of learning them is unclear.

Questions for authors
  1. Is there evidence that predicting eigenmodes with the proposed neural operator is better than computing them with classical methods?

  2. Can the method be generalized to PDE families whose eigenmodes cannot be solved trivially or whose eigen structure is not known in advance?

  3. Has computing an eigendecomposition and reusing it for time-dependent problems been used in classical PDE solving, and how does ANNO differ from such methods?

6 comments
AuthorsOpenReview ↗

Response to Weakness 1

Thank you for your comments and questions. We would first like to clarify that our contribution is not to reintroduce eigenfunction expansions or modal time integration, but to investigate how reusable spatial analytical structures can be learned from steady-state problems within a parametric PDE family, transferred directly to different temporal dynamics, and integrated with neural networks.

This perspective was already established in the original manuscript. We define PDEs sharing the same computational domain, spatial operator, and homogeneous boundary-condition space as a PDE family, and separate the solution process into reusable spatial analytical structures and equation-specific temporal evolution. ANNO learns the lifting function, boundary-dependent spatial modes, and modal coefficients using only steady-state data, then transfers these structures to diffusion or wave dynamics without transient training data or network retraining. The relevant discussion is provided in the revised Introduction (pages 1–2, lines 74–112).

Following the reviewers' suggestions, we added two evaluations in this revision: a full-trajectory comparison with optimized classical modal methods and an extension to 3D thermal conduction.

Critical Weakness

The main weakness of ANNO is that it learns only the parameters (frequencies and amplitudes) of a hard-coded trigonometric eigenmode form, not the form itself, so the eigenstructure must be known in advance and the architecture cannot represent eigenmodes for variable coefficients or irregular geometries. The evaluation is also limited to linear, constant-coefficient problems on intervals and rectangles, where the eigenmodes are closed-form or can be found by simple one-dimensional root-finding, so the classical modal solution is already cheap and exact and the benefit of learning them is unclear.

We agree that the current ANNO uses the known trigonometric eigenstructure as an inductive bias, and its direct applicability is limited to linear, constant-coefficient problems on regular domains with separable spatial operators. We did not claim that the current architecture can handle arbitrary variable coefficients or irregular geometries. This limitation was already acknowledged in the original Conclusion and Future Work section.

The question we investigate is whether, when the analytical structure of a PDE is known, we can avoid learning temporal evolution already prescribed by the governing equations from transient trajectories, and instead learn spatial structures reusable within a PDE family.

Importantly, knowing the analytical form does not mean that obtaining the eigenstructure across a parametric PDE family is computationally free. For a fixed spatial operator and homogeneous boundary conditions, classical methods can compute the eigenbasis once and reuse it indefinitely. However, when Robin parameters or the spatial operator change, classical methods must compute a new eigenstructure. ANNO instead learns the mapping from problem parameters to spatial modes and modal responses through steady-state training, allowing modes for new configurations to be obtained through network inference.

We added a full-trajectory comparison with classical numerical modal methods. Under the same finite modal budget, ANNO achieves lower prediction errors with both MC and BP state-transfer strategies. With the same online caching strategy, ANNO is slightly faster under MC, while their computation times are comparable under BP. Detailed results are provided in Section 4.7 (pages 9–10) and Appendix K.

We understand the reviewer's concern regarding variable coefficients and irregular geometries, which the current ANNO cannot directly handle. This limitation was already identified in the original manuscript and remains an important direction for our future work.

AuthorsOpenReview ↗

Response to Q1

Question 1

Is there evidence that predicting eigenmodes with the proposed neural operator is better than computing them with classical methods?

This question should be addressed from three aspects: prediction accuracy, online computational cost, and offline cost.

1. Prediction Accuracy Under the Same Finite Modal Budget

The original manuscript already included an ablation comparing learned and numerical modes with learned and analytical coefficients. We retrained and reevaluated these configurations using the corrected frequency bands and refined transient datasets. The results are reported in Table 33 (page 49):

Modes / CoefficientsMedian Steady-State ErrorMedian Full-Trajectory ErrorSingle-Step Time
Learned / Learned0.3915%0.6433%0.513 ms
Numerical / Learned0.6528%0.9348%1.594 ms
Learned / Analytical0.3470%0.7305%0.564 ms
Numerical / Analytical0.6507%0.9219%1.685 ms

Under the same finite modal budget, the learned-mode configurations achieve lower full-trajectory errors and shorter single-step computation times.

The new Experiment 7 further compares ANNO with an optimized classical numerical modal method on an independent 288×192288\times192 test dataset with time-varying Robin boundary conditions. Both methods use K=12K=12 modes per spatial direction:

MethodMC Mean ErrorBP Mean Error
ANNO1.6681%0.8493%
CNM1.9912%1.3479%

The results are reported in Table 2 (page 10) and Table 41 (page 55). Classical modal methods approximate solutions by truncating the eigenmode expansion, whereas ANNO learns to approximate target solutions using a finite number of modes. This explains ANNO's higher prediction accuracy under the same modal budget.

2. Online Computational Cost

Experiment 7 uses an efficient one-dimensional symmetric tridiagonal eigensolver and online lazy caching, with identical state-transfer strategies for ANNO and CNM.

Basis Update Ratio ρ\rhoANNO–MCCNM–MCANNO–BPCNM–BP
00.2695 s0.3056 s0.4146 s0.4123 s
0.50.4118 s0.4520 s0.5507 s0.5569 s
1.00.5589 s0.5998 s0.7041 s0.7031 s

The results are reported in Table 2 and Table 43 (page 56).

The original ANNO implementation uses Modal Coefficient Carry (MC), which directly carries coefficients with the same modal indices when the eigenbasis changes, avoiding explicit state reprojection. We additionally evaluated Basis Projection (BP) for both methods.

With the same caching strategy, ANNO has lower full-trajectory computation times under MC, while the two methods have comparable times under BP. Classical methods can also reduce computational costs through MC and caching. Therefore, this advantage does not imply that neural networks have universally faster temporal integration.

3. Offline Cost

We also added an offline cost analysis in Section 4.7 (page 10, lines 496–498). ANNO requires approximately 35 s for steady-state dataset generation and 3709 s for network training.

At ρ=1\rho=1 under MC, ANNO saves 0.0409 s per trajectory compared with CNM, requiring approximately 91,500 trajectories to offset its offline costs.

Therefore, classical methods remain more economical for a small number of solves. ANNO's main advantages are higher prediction accuracy under the same finite modal budget and the ability to perform large numbers of repeated solves within the same PDE family after offline training.

AuthorsOpenReview ↗

Response to Q2 and 3

Question 2

Can the method be generalized to PDE families whose eigenmodes cannot be solved trivially or whose eigen structure is not known in advance?

ANNO's overall framework should be distinguished from its current trigonometric modal parameterization.

The ANNO framework consists of boundary lifting, spatial modal representation, structural constraints, and equation-specific modal time integration. It does not inherently require trigonometric eigenmodes. The trigonometric form is a structured parameterization chosen for the constant-coefficient, separable problems studied here, as explained in the original manuscript.

The original MLP Basis ablation also shows that the modal network is architecturally replaceable. However, an unstructured MLP produces larger transient errors under the current training setup, highlighting the importance of appropriate spatial inductive biases.

For higher-dimensional separable problems, the approach extends through tensor products. In this revision, we added a 3D battery-module thermal conduction experiment with a 24×54×3224\times54\times32 spatial grid and six modes per direction, forming 216 three-dimensional tensor-product modes. Trained only on steady-state data, ANNO achieves a mean full-trajectory error of 1.2012% across 144 transient test trajectories. The results are reported in Section 4.5 (page 8, lines 409–427) and Appendix H.

For problems with known approximate spatial structures but no simple root-finding solution, the trigonometric reconstructor could be replaced with other parameterized basis networks, while retaining operator residual, boundary residual, orthogonality, and mode-ordering constraints. This is a natural extension of ANNO.

For problems with entirely unknown eigenstructures, strongly heterogeneous coefficients, nonseparable operators, or irregular geometries, the current ANNO cannot be directly applied. This limitation and future research directions were acknowledged in the original Conclusion and Future Work section, rather than introduced only in this revision.

Question 3

Has computing an eigendecomposition and reusing it for time-dependent problems been used in classical PDE solving, and how does ANNO differ from such methods?

Yes. Computing spatial eigenstructures through separation of variables, eigenfunction expansions, or modal superposition, and reusing them for time evolution in diffusion, wave, and vibration problems, is standard in classical PDE solving. We have never claimed eigenfunction expansions or modal time integration as new contributions. The original Introduction attributes the separation of spatial representation and temporal evolution to classical PDE theory (page 1, lines 51–53).

The main difference between ANNO and classical fixed-eigenbasis reuse lies in the problem setting and how spatial structures are obtained:

  • Classical methods compute an eigenbasis for a fixed spatial operator and homogeneous boundary conditions, and reuse it in subsequent solves.
  • ANNO targets parametric PDE families and learns the mapping from problem parameters to spatial modes and steady-state modal responses using steady-state data.

When the spatial operator and homogeneous boundary conditions remain unchanged, classical methods can reuse cached eigenbases. When Robin parameters or the spatial operator change, a new eigenstructure is required. ANNO learns this parametric mapping through steady-state training, obtaining modes for new configurations through network inference.

Basis updates must be distinguished from state transfer. Under time-varying boundary conditions, the original ANNO uses the Modal Coefficient Carry (MC) approximation: even when the eigenbasis changes, coefficients with the same modal indices are carried directly, without explicit basis projection, trading some accuracy for efficiency. We have clarified and compared these strategies in Section 2.3 and Appendix K.

Furthermore, ANNO learns both spatial modes and their corresponding steady-state modal coefficients, reusing these structures for the temporal evolution of different equations. Its central idea is to learn reusable analytical structures from steady-state problems at the PDE-family level, rather than reuse a fixed eigenbasis for the same problem. This perspective and the absence of transient training were explained in the original Introduction (pages 1–2, lines 74–112).

The new comparison with classical numerical modal methods further demonstrates this distinction. Under the same finite modal budget and state-transfer strategy, ANNO achieves lower full-trajectory errors. With identical caching, ANNO has lower online computation times under MC, while the methods have comparable times under BP. Detailed results are provided in Section 4.7 (pages 9–10) and Appendix K.

Reviewer_SCB3OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

ANNO learns a reusable spatial representation from steady-state solutions and applies known modal evolution formulas to related diffusion and damped-wave problems without transient training. It separates nonhomogeneous boundary data through analytical lifting, predicts boundary-conditioned trigonometric modes and steady coefficients, and constrains the modes with a discrete eigenstructure loss. Two-dimensional modes are tensor products. The frozen model supplies the spatial representation and instantaneous steady coefficients used in subsequent time evolution.

The contribution is primarily an amortized boundary-to-eigenbasis construction coupled to classical modal dynamics. Validation covers one-dimensional Poisson and diffusion problems, a two-dimensional membrane, a thermal plate, and component ablations. Reported mean errors are 0.1643% for one-dimensional steady prediction and 0.2531%, 1.8646%, and 1.0506% for the three transient studies. Appendix H includes a useful numerical-mode/analytical-coefficient comparator with no learned component. The results support an interpretable proof of concept for the specified linear, constant-coefficient, separable families; they do not provide a general learned-model accuracy or stability guarantee.

Critical strengths
  1. Interpretable reuse of steady-trained structure. Learning boundary-conditioned spatial modes and then supplying the known temporal laws is a coherent way to avoid learning entire trajectories. The explicit lifting and eigenstructure penalty distinguish this from unconstrained low-rank solution fitting. This is consequential because it identifies a useful role for learning within a well-understood numerical method.

  2. Meaningful within-family validation and controls. The tests span multiple boundary combinations, sources, nonzero diffusion initial states, dynamic forcing, and both diffusion and wave evolution. Appendix E.4 examines baseline failures, and Table 28 includes a genuinely nonlearned numerical/analytical modal comparator. Full ANNO's median thermal trajectory error is 0.8384%, versus 1.1946% for that comparator. These controls provide more persuasive evidence than comparisons with neural baselines alone, although they do not resolve the concerns below.

Critical weaknesses
  1. The learned representation and its reuse are not fully consistent with the stated analytical mechanism. Appendix B.3 assumes a fixed homogeneous boundary operator, whereas G.4 changes h(t), reevaluates the Robin modes, and applies the same recurrence without documenting state transport. For v = sum_j a_j(t) phi_j(x,t), even exact instantaneous orthonormal eigenmodes produce the additional coupling sum_j <phi_i,partial_t phi_j> a_j. A consistent stepwise scheme can instead reproject the old physical state into the new basis, but that step is not specified. Separately, C.2 lists only projected forcing and log D as coefficient-network inputs, although a_k = q_k/(D lambda_k) also depends on the spectrum. Equation (134)'s frequency bands also exclude admissible low-Robin modes: for NR boundaries with alpha = 0.2, beta = 1 on [0,1], omega tan(omega) = 0.2 has its second root at 3.203935, below the second channel's lower bound 1.25 pi = 3.926991. These are material questions about whether the proposed representation supports the claimed reuse, rather than requests for unrelated extensions. They do not by themselves disprove the reported errors on the sampled data.

  2. The incremental accuracy-cost benefit remains insufficiently established. The learned/analytical variant nearly matches full ANNO's median trajectory error, 0.8561% versus 0.8384%, so learned coefficient inference is not necessary for the reported accuracy. Table 28's 0.488 versus 2.204 ms comparison measures a steady sample, not a full trajectory. For these separable problems, scalar frequency root-finding and spectrum reuse are important classical alternatives to a fresh numerical eigensolve. The large neural-baseline gaps also need cautious interpretation: ST-FNO has width 4 or 8, while POD propagation is trained over eight steps and evaluated over 128 or 256. The existing diagnostics are valuable but do not establish competitive tuning. Independent parameter samples, rather than correlated time frames alone, should inform data-efficiency comparisons. These limitations leave the practical advantage of the learned components uncertain.

Questions for authors
  1. When h(t) changes in G.4, exactly how is the current state transferred between successive boundary-conditioned bases? Please give the update, including lifting and any Gram-matrix treatment, and show time-step convergence for Cooling/Combined. Is reprojection present in the implementation but omitted from the manuscript?

  2. Does the coefficient network receive eigenvalues or boundary information through an unreported input or normalization? With D = 1, forcing equal to the first normalized eigenfunction gives q = (1,0,...) under both DD and DN conditions, but requires coefficients 1/pi^2 and 4/pi^2. How is this ambiguity resolved? Please distinguish a general coefficient map from correlations specific to the sampled source distribution.

  3. How do the implemented frequency bounds accommodate the NR example above and other weak-Robin spectra? Please reconcile the indexing/bounds with Equation (134) and report eigenvalue/eigenresidual errors and mode orthogonality across the retained modes and boundary regimes. If the implementation uses different bounds, which reported results correspond to them?

  4. What is the full-trajectory accuracy-runtime tradeoff against an efficient root-based or cached modal solver, with hardware, precision, batching, and offline costs specified? Would the large transient-baseline gaps persist with a modest capacity sweep and a propagation training horizon appropriate to evaluation? The numerical/analytical comparator and E.4 diagnostics already provided are useful starting points.

  5. What reference time integrator, tolerances, and grid/time convergence checks generated the transient targets? Please also resolve whether the thermal and ablation relative errors use absolute temperature, temperature rise, or the lifted residual, and provide an absolute-temperature error. An anonymized implementation and configurations would help verify these points.

5 comments
AuthorsOpenReview ↗

Response to Q1

We sincerely thank you for your valuable guidance, which has greatly helped improve the quality of ANNO. Before addressing your questions point by point, we would like to summarize the main revisions:

  1. Added convergence verification for the reference numerical solvers and datasets, and refined the transient reference solutions in Experiments 3 and 4.
  2. Corrected the frequency-band definition and retrained all affected models.
  3. Added full-trajectory accuracy and computational cost comparisons with classical numerical modal methods (CNM).
  4. Added ST-FNO capacity and POD-DeepPropNet training rollout length experiments.
  5. Added a 3D battery-module thermal conduction experiment.

After these revisions, the mean full-trajectory relative L2L_2 errors for 1D diffusion, 2D wave, and 2D thermal dynamics decreased from 0.2531%, 1.8646%, and 1.0506% to 0.1910%, 0.3823%, and 0.7849%, respectively.

Question 1

When h(t) changes in G.4, exactly how is the current state transferred between successive boundary-conditioned bases? Please give the update, including lifting and any Gram-matrix treatment, and show time-step convergence for Cooling/Combined. Is reprojection present in the implementation but omitted from the manuscript?

The original implementation did not perform explicit basis projection. Instead, it directly carried modal coefficients with the same indices between successive bases. This was an intentional design choice to improve computational efficiency at the cost of a small loss in accuracy. We apologize for not explaining this approximation in the original manuscript.

In the revised manuscript, we have added the state reprojection formulation for time-varying eigenbases in Appendix B.3, including lifting and Gram-matrix projection. We have also clarified in Section 2.3 (page 4, lines 169–172) that the original ANNO implementation uses approximate modal coefficient carry. Appendix K.2 distinguishes two state-transfer strategies:

  • Modal Coefficient Carry (MC): Directly carries coefficients with the same modal indices.
  • Basis Projection (BP): Reconstructs the physical residual field using the old basis and then projects it onto the new basis using the Gram matrix.

The definitions are provided in Appendix K.2 (page 54, lines 2884–2894). For BP, we first subtract the lifting function from the current physical state:

vn=Tn−ℓn.v_n=T_n-\ell_n.

We then solve the projection system in the new basis:

=\langle\phi^{\mathrm{new}},v_n\rangle, \qquad G^{\mathrm{new}}_{ij} =\langle\phi_i^{\mathrm{new}},\phi_j^{\mathrm{new}}\rangle.$$ We further compared these two strategies on an independent $288\times192$ test dataset with time-varying Robin boundary conditions: | Method | State Transfer | Mean Error | Median Error | |---|---|---:|---:| | ANNO | MC | 1.6681% | 1.3553% | | CNM | MC | 1.9912% | 1.8746% | | ANNO | BP | 0.8493% | 0.8817% | | CNM | BP | 1.3479% | 1.3666% | The results are reported in Table 41 (page 55, lines 2917–2922). Explicit BP reduces ANNO's mean error from 1.6681% to 0.8493%, while MC avoids the computational cost of explicit state projection. We discuss the computational advantages in detail in Question 4 rather than repeating them here. For the Cooling and Combined cases, we also examined the time-step convergence of the reference solutions. The internal time steps in Experiment 4 were refined from 128 in the original manuscript to 1024. Using the 2048-step solution as the reference, the mean trajectory difference decreased from 0.3336% to 0.0179%, and the maximum difference decreased to 0.0345%. We therefore regenerated the transient dataset for Experiment 4 and retrained or reevaluated all affected models. The results are provided in Appendix J.3 (pages 52–53).
AuthorsOpenReview ↗

Response to Q2 and 3

Question 2

Does the coefficient network receive eigenvalues or boundary information through an unreported input or normalization? With D = 1, forcing equal to the first normalized eigenfunction gives q = (1,0,...) under both DD and DN conditions, but requires coefficients 1/pi^2 and 4/pi^2. How is this ambiguity resolved? Please distinguish a general coefficient map from correlations specific to the sampled source distribution.

Yes, the actual implementation uses spectral normalization that was not sufficiently explained in the original manuscript. The coefficient network does not directly receive eigenvalues or Robin parameters, but it also does not use raw source projections. For the learned modes under the current boundary conditions, we first construct the modal stiffness matrix KϕK_\phi and compute

q~ϕ,k=qϕ,k(Kϕ)kk+ϵ.\widetilde{q}_{\phi,k}=\frac{q_{\phi,k}}{(K_\phi)_{kk}+\epsilon}.

The coefficient network receives q~ϕ\widetilde{\mathbf{q}}_\phi and log⁡D\log D. Since the stiffness matrix is constructed from the learned modes under the current boundary conditions, this normalization incorporates boundary-dependent spectral information. We have clarified this in Section 3 (page 4, lines 196–200) and Appendix C.2.1 (page 29, lines 1523–1539) of the revised manuscript.

For the reviewer's example with D=1D=1, under an exact normalized eigenbasis,

(Kϕ)kk=λk.(K_\phi)_{kk}=\lambda_k.

Therefore, although the first normalized source projection is identical under DD and DN boundary conditions, the normalized inputs to the network are different:

q~DD≈1π2,q~DN≈4π2.\widetilde{q}_{\mathrm{DD}}\approx\frac{1}{\pi^2},\qquad\widetilde{q}_{\mathrm{DN}}\approx\frac{4}{\pi^2}.

Thus, the two boundary conditions are distinguishable in the actual network inputs. Resolving this ambiguity does not rely on correlations between source terms and boundary conditions in the training data.

Question 3

How do the implemented frequency bounds accommodate the NR example above and other weak-Robin spectra? Please reconcile the indexing/bounds with Equation (134) and report eigenvalue/eigenresidual errors and mode orthogonality across the retained modes and boundary regimes. If the implementation uses different bounds, which reported results correspond to them?

The reviewer has identified an actual design error in the frequency bands used in the original manuscript. The original implementation adopted shifted frequency bands:

ω0∈[0,1.25π],\omega_0\in[0,1.25\pi],

\qquad k=1,\ldots,K-1.$$ Therefore, the second NR eigenfrequency identified by the reviewer cannot be represented by its corresponding channel. We have corrected the frequency bands to $$\omega_k\in[k\pi,(k+1)\pi], \qquad k=0,\ldots,K-1.$$ The corrected definition is provided in Appendix C.1.3 (page 28, lines 1500–1511). To verify the necessity of this correction, we compared three frequency-band strategies using the same dataset, network architecture, loss function, and training settings: 1. **Original shifted bands:** $$\omega_0\in[0,1.25\pi],\qquad \omega_k\in[(k+0.25)\pi,(k+1.25)\pi].$$ 2. **Corrected non-overlapping bands:** $$\omega_k\in[k\pi,(k+1)\pi].$$ 3. **Overlapping bands:** Designed to reduce gradient propagation difficulties caused by hard band boundaries, but found to be ineffective. With $\delta=0.25$, $$\omega_k\in[\max(0,k-\delta)\pi,(k+1+\delta)\pi].$$ | Frequency-Band Strategy | Mean Steady-State Error | Median Steady-State Error | Mean Transient Error | Median Transient Error | |---|---:|---:|---:|---:| | Original shifted bands | 0.4824% | 0.3732% | 0.9094% | 0.7197% | | Corrected non-overlapping bands | 0.5141% | 0.3915% | **0.7849%** | **0.6433%** | | Overlapping bands | 0.6810% | 0.5206% | 6.4357% | 3.9317% | The original shifted bands achieve slightly lower steady-state errors but cannot cover the full target Robin spectrum, resulting in higher transient errors than the corrected bands. The overlapping bands allow adjacent channels to converge to similar frequencies, causing mode crowding and substantially increased transient errors. Therefore, the final version adopts the corrected non-overlapping bands, and all affected models have been retrained using this configuration. After the correction, the median overlaps between the first four learned modes and their numerical reference modes are **0.9999999, 0.9999937, 0.9999777, and 0.9999447**, respectively. These results demonstrate a substantial improvement in the learned modal representation, making this correction particularly important. See Table 12 (page 33, lines 1754–1760).
AuthorsOpenReview ↗

Response to Q4

Question 4

What is the full-trajectory accuracy-runtime tradeoff against an efficient root-based or cached modal solver, with hardware, precision, batching, and offline costs specified? Would the large transient-baseline gaps persist with a modest capacity sweep and a propagation training horizon appropriate to evaluation? The numerical/analytical comparator and E.4 diagnostics already provided are useful starting points.

Following your suggestion, we added an independent full-trajectory experiment to compare the accuracy and computational cost of ANNO with classical numerical modal methods (CNM).

The experiment uses a 288×192288\times192 spatial grid, 129 time points, and K=12K=12 modes per spatial direction, with a symmetric tridiagonal eigensolver for one-dimensional eigenproblems and tensor-product modes in two dimensions. Both methods use the same state-transfer and online lazy caching strategies. Details are provided in Section 4.7 (pages 9–10) and Appendix K.

1. Comparison with Classical Numerical Modal Methods

Under the same finite modal budget, ANNO achieves lower full-trajectory errors than CNM with both MC and BP. The mean errors are 1.6681% for ANNO–MC, 1.9912% for CNM–MC, 0.8493% for ANNO–BP, and 1.3479% for CNM–BP. See Question 1 and Table 41.

With online lazy caching, the full-trajectory computation times are:

Basis Update Ratio ρ\rhoANNO–MCCNM–MCANNO–BPCNM–BP
00.2695 s0.3056 s0.4146 s0.4123 s
0.50.4118 s0.4520 s0.5507 s0.5569 s
1.00.5589 s0.5998 s0.7041 s0.7031 s

The complete results are reported in Table 43 (page 56, lines 2981–2987). With the same caching strategy, ANNO is faster than CNM under MC, while their computation times are comparable under BP.

Unlike classical modal methods that approximate solutions by truncating the eigenmode expansion, ANNO learns to approximate target solutions using a finite number of modes. This explains ANNO's higher prediction accuracy under the same modal budget.

We also compared continuous root-finding CNM with efficient tridiagonal CNM. Their prediction errors are nearly identical, but the tridiagonal method has substantially lower computational cost. We therefore use tridiagonal CNM as the main classical baseline in the paper. The results are reported in Table 44.

2. Further Evaluation of Transient Neural Baselines

We shared the concern that the relatively weak performance of the transient baselines might result from implementation issues. For this reason, we had already conducted several diagnostic experiments in the original manuscript. Following your suggestion, we further added model capacity and training rollout length experiments in this revision. The results indicate that the baselines require more training data and greater model capacity to approach ANNO's accuracy on the complex problems studied here. However, further optimization of these baselines is beyond the scope of this work. The details are as follows.

We had already investigated these concerns through three baseline diagnostics in the original manuscript. Increasing ST-FNO training trajectories from 384 to 6144 reduced its error from 35.7777% to 16.0419% (Fig. 11(a)). POD-DeepPropNet's decreasing eight-step training loss did not improve full-trajectory accuracy (Fig. 11(b)). On a simplified diffusion family, ST-FNO and POD-DeepPropNet achieved errors of 0.7012% and 2.9226%, respectively (Fig. 12).

Based on these analyses and your suggestion, we added the following experiments:

MethodConfigurationParametersMean Trajectory Error
ANNOFrozen steady-state model57,3580.1910%
ST-FNOwidth 458,27322.7527%
ST-FNOwidth 8231,10512.6513%
ST-FNOwidth 16921,2177.4489%
POD-DeepPropNetrollout 858,96845.5992%
POD-DeepPropNetrollout 6458,96823.5622%
POD-DeepPropNetrollout 25658,96824.4566%

The results are reported in Table 19 (page 37, lines 1945–1955).

Increasing the ST-FNO width from 4 to 16 reduced its mean error from 22.7527% to 7.4489%, but increased its parameter count to approximately 16 times that of ANNO. For POD-DeepPropNet, increasing the training rollout from 8 to 64 steps reduced the error from 45.5992% to 23.5622%. Further increasing the rollout to 256 steps did not improve accuracy.

These results show that increasing model capacity and training rollout length can improve the transient baselines, but they still do not reach ANNO's accuracy on the complex PDE families evaluated in this work. ANNO, in contrast, is trained only on steady-state data and requires no transient training trajectories.

Reviewer_qYrPOpenReview ↗
Rating
4
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper presents ANNO, which learns spatial eigenmodes and modal coefficients from steady-state problems and reuses them with analytical time evolution. Experiments on 1D diffusion and 2D wave and thermal problems demonstrate accurate transient predictions without transient training or network retraining.

Critical strengths

I believe this is a well-executed paper with a useful contribution to neural PDE solvers. The central construction separates learning a spatial representation from advancing the temporal dynamics already specified by the PDE. Learning eigenstructure from steady-state problems and reusing it for diffusion or wave evolution is a practical way to avoid generating transient training trajectories and retraining a network for each temporal task. The explicit roles of the lifting function, spatial modes, and modal coefficients make the method understandable and provide a clear explanation of what is reused within a PDE family.

The reported accuracy is encouraging, especially because the networks use no transient training data. Taken together, the analytical organization and demonstrated reuse make this a solid contribution. I support acceptance, while suggesting more direct comparison with classical spectral methods and a three-dimensional example to strengthen the practical case.

Critical weaknesses

My main suggestion is to strengthen the comparison with classical spectral or modal methods. ANNO uses a spatial eigenbasis and equation-specific modal evolution, which are also natural ingredients of established numerical approaches. Comparisons with transient neural operators show the value of incorporating this structure, but do not by themselves establish when learning the spatial modes and steady-state coefficients is preferable to constructing those components numerically.

A comparison of full transient accuracy and computational cost would make this distinction clearer. In particular, the reader should be able to see how the initial construction or training cost is amortized over repeated trajectories, and how the conclusions change when numerical modes are cached. This would better establish the operating conditions in which the neural components offer an advantage, while recognizing that both approaches can use known modal dynamics.

My other suggestion is to include at least one 3D example. The membrane and battery cooling plate demonstrate relevant applications, but both use two-dimensional spatial models. A three-dimensional thermal or wave problem consistent with the current separable formulation would help establish whether the compact modal representation remains effective at a scale closer to practical engineering use. These suggestions would strengthen a paper that I already consider suitable for acceptance.

Questions for authors

Could the authors clarify the accuracy–cost comparison with a classical spectral/modal solver when the numerical eigenstructure is reused across trajectories? In particular, which changes in the evaluated inputs require a fresh eigensolve, and how much of the reported advantage remains when the basis can be cached?

4 comments
AuthorsOpenReview ↗

Summary of Major Revisions and Key Updates

We sincerely thank you for your positive assessment of our work and your specific suggestions. Following your recommendations, we added a full-trajectory comparison with classical numerical modal methods and a 3D thermal conduction experiment. We address these additions in detail in the following responses.

We would also like to note that other reviewers identified two important issues: an error in the original frequency-band design and insufficient temporal resolution in the transient datasets. We therefore reran the affected experiments, which changed several reported results.

First, the original shifted frequency bands,

ω0∈[0,1.25π],ωk∈[(k+0.25)π,(k+1.25)π],\omega_0\in[0,1.25\pi],\qquad \omega_k\in[(k+0.25)\pi,(k+1.25)\pi],

could not cover some valid low-frequency eigenvalues under weak Robin boundary conditions. We corrected them to

ωk∈[kπ,(k+1)π],k=0,…,K−1,\omega_k\in[k\pi,(k+1)\pi],\qquad k=0,\ldots,K-1,

and retrained all affected ANNO models using the corrected bands.

Second, time-step convergence checks of the reference numerical solvers revealed insufficient temporal resolution in the original transient datasets for Experiments 3 and 4. Both datasets originally used 128 internal time steps. Compared with 2048-step reference solutions, their mean trajectory differences were 1.4936% and 0.3336%, respectively. We increased both to 1024 internal steps, reducing the mean differences to 0.0190% and 0.0179%, and the maximum differences to 0.0341% and 0.0345%, respectively. We therefore regenerated the complete transient datasets for Experiments 3 and 4 and retrained or reevaluated all models depending on these datasets, including ANNO, ST-FNO, POD-DeepPropNet, and the ablation models.

The main experimental results changed as follows:

ExperimentMetricOriginalRevised
Exp. 1: 1D steady PoissonMean relative L2L_2 error0.1643%0.0899%
Exp. 2: 1D diffusionMean full-trajectory error0.2531%0.1910%
Exp. 3: 2D waveMean full-trajectory error1.8646%0.3823%
Exp. 4: 2D thermalMean full-trajectory error1.0506%0.7849%
Full ANNO ablation configurationMedian full-trajectory error0.8384%0.6433%
Learned modes + Analytical coefficientsMedian full-trajectory error0.8561%0.7305%
Numerical modes + Analytical coefficientsMedian full-trajectory error1.1946%0.9219%
AuthorsOpenReview ↗

Response to Weakness 1

Critical Weakness 1

My main suggestion is to strengthen the comparison with classical spectral or modal methods. ANNO uses a spatial eigenbasis and equation-specific modal evolution, which are also natural ingredients of established numerical approaches. Comparisons with transient neural operators show the value of incorporating this structure, but do not by themselves establish when learning the spatial modes and steady-state coefficients is preferable to constructing those components numerically.

A comparison of full transient accuracy and computational cost would make this distinction clearer. In particular, the reader should be able to see how the initial construction or training cost is amortized over repeated trajectories, and how the conclusions change when numerical modes are cached. This would better establish the operating conditions in which the neural components offer an advantage, while recognizing that both approaches can use known modal dynamics.

Thank you for this important suggestion. In response, we added Section 4.7 (pages 9–10) to the revised manuscript, reporting the full-trajectory accuracy, online computational time, and offline costs of ANNO compared with classical numerical modal methods (CNM). Detailed experimental settings and additional comparisons are provided in Appendix K.

The experiment uses a 288×192288\times192 spatial grid, 129 time points, and K=12K=12 modes per spatial direction. CNM uses an efficient one-dimensional symmetric tridiagonal eigensolver and constructs two-dimensional tensor-product modes. Both methods use the same temporal integration, state-transfer strategies, and online lazy caching mechanism. The settings are provided in Table 40 (page 54).

We consider two state-transfer strategies: Modal Coefficient Carry (MC), which directly carries coefficients with the same modal indices, and Basis Projection (BP), which reprojects the reconstructed physical residual field using the Gram matrix.

Under the same modal budget, ANNO achieves lower mean full-trajectory errors with both strategies:

MethodMCBP
ANNO1.6681%0.8493%
CNM1.9912%1.3479%

Classical modal methods approximate target solutions by truncating the eigenmode expansion, whereas ANNO learns to approximate target solutions using the same finite number of modes. This explains ANNO's higher prediction accuracy under the same modal budget.

With online lazy caching, the full-trajectory computation times are:

Basis Update Ratio (ρ\rho)ANNO–MCCNM–MCANNO–BPCNM–BP
00.2695 s0.3056 s0.4146 s0.4123 s
0.50.4118 s0.4520 s0.5507 s0.5569 s
1.00.5589 s0.5998 s0.7041 s0.7031 s

The accuracy and online time comparisons are summarized in Table 2 (page 10), with complete results in Tables 41 (page 55) and 43 (page 56). Under the same caching strategy, ANNO has lower online computational time under MC, while the two methods have comparable times under BP.

Appendix K also reports results without caching and cold- and warm-cache comparisons between continuous root-finding CNM and tridiagonal CNM. The two classical implementations achieve nearly identical prediction accuracy, but the tridiagonal method has substantially lower computational cost. We therefore use tridiagonal CNM as the main classical baseline in the paper.

We also added an offline cost and amortization analysis in Section 4.7 (page 10, lines 496–498). ANNO requires approximately 35 s for steady-state dataset generation and 3709 s for training. At ρ=1\rho=1 under MC, ANNO saves 0.0409 s per trajectory compared with CNM, requiring approximately 91,500 trajectories to offset its offline costs.

Therefore, classical methods remain more economical for a small number of solves. ANNO's main advantages are higher prediction accuracy under the same finite modal budget and the ability to reuse learned structures across large numbers of repeated queries.

AuthorsOpenReview ↗

Response to Weakness 2

Critical Weakness 2

My other suggestion is to include at least one 3D example. The membrane and battery cooling plate demonstrate relevant applications, but both use two-dimensional spatial models. A three-dimensional thermal or wave problem consistent with the current separable formulation would help establish whether the compact modal representation remains effective at a scale closer to practical engineering use. These suggestions would strengthen a paper that I already consider suitable for acceptance.

Following your suggestion, we added Experiment 5, a 3D battery-module thermal conduction problem. The results are presented in Section 4.5 (page 8, lines 409–427) and Appendix H (pages 44–47).

The model consists of six independently heat-generating cells arranged in a 3×2×13\times2\times1 configuration. The computational domain measures

0.081×0.296×0.092m3,0.081\times0.296\times0.092\\ \mathrm{m}^3,

with a 24×54×3224\times54\times32 spatial grid containing 41,472 points. The material has anisotropic thermal conductivity, with an adiabatic bottom surface and Robin convective boundary conditions on the other five surfaces. The physical parameters are provided in Table 29 (page 45).

ANNO retains six modes per spatial direction, forming 63=2166^3=216 three-dimensional tensor-product modes. The networks are trained on only 1,024 steady-state temperature fields, without any transient training trajectories or fine-tuning.

The transient tests cover six scenarios: Uniform, Pulse, Periodic, Cell Imbalance, Cooling, and Combined. Each scenario contains 24 independent trajectories, giving 144 test trajectories in total, with 129 time points per trajectory.

ANNO achieves a mean full-trajectory relative L2L_2 error of 1.2012% across all 144 test trajectories, as reported in Section 4.5 and Figure 6.

This experiment demonstrates that ANNO's learned spatial analytical structures can be extended to three dimensions through tensor products, enabling accurate prediction of 3D thermal evolution without transient training data.

AuthorsOpenReview ↗

Response to Q1

Question for Authors

Could the authors clarify the accuracy–cost comparison with a classical spectral/modal solver when the numerical eigenstructure is reused across trajectories? In particular, which changes in the evaluated inputs require a fresh eigensolve, and how much of the reported advantage remains when the basis can be cached?

Whether a new eigensolve is required depends on changes to the spatial operator, computational domain, and homogeneous boundary operator.

Classical methods must recompute the eigenstructure when the computational domain, spatial discretization, spatial operator, or homogeneous boundary operator changes. For example, in thermal problems, changing the Robin parameter ratio h/kh/k changes the eigenbasis.

In contrast, changes to the source term, initial condition, or nonhomogeneous boundary values do not require a new eigensolve if the homogeneous spatial operator remains unchanged. Previously computed eigenbases can also be reused through caching.

It is important to distinguish basis updates from state transfer. The original ANNO implementation uses the Modal Coefficient Carry (MC) approximation: even when the eigenbasis changes, coefficients with the same modal indices are carried directly, without explicit basis projection. This design improves computational efficiency at the cost of some accuracy. We have clarified this approximation in Section 2.3 (page 4, lines 169–172) and added a comparison with explicit Basis Projection (BP) in Appendix K.

Following your suggestion, we further added online lazy caching experiments in Appendix K.3–K.4 (pages 54–56). Both ANNO and CNM use the same caching strategy: a new modal basis is constructed and cached when a configuration first appears, and reused thereafter. We also evaluate both MC and BP for each method to distinguish the effects of state-transfer approximation and caching.

The full-trajectory computation times with caching are:

Basis Update Ratio (ρ\rho)ANNO–MCCNM–MCANNO–BPCNM–BP
00.2695 s0.3056 s0.4146 s0.4123 s
1.00.5589 s0.5998 s0.7041 s0.7031 s

The results are reported in Table 2 (page 10) and Table 43 (page 56). Under the same caching strategy, ANNO has lower online computation times under MC, while the two methods have comparable times under BP. For both methods, MC reduces computational cost by avoiding explicit state reprojection.

Under the same modal budget of K=12K=12 per spatial direction, ANNO also achieves higher prediction accuracy. The mean errors are 1.6681% for ANNO versus 1.9912% for CNM under MC, and 0.8493% versus 1.3479% under BP. Thus, ANNO maintains low prediction errors while using the MC approximation to avoid the cost of state reprojection.

These results show that classical methods can also reduce computational costs through caching and MC, while ANNO retains an accuracy advantage under the same modal budget and state-transfer strategy. We have also added an offline cost analysis in Section 4.7: at ρ=1\rho=1 under MC, approximately 91,500 trajectories are required to offset ANNO's dataset generation and training costs.

Finally, we would like to express our sincere gratitude once again for your positive assessment and valuable suggestions. Your encouragement means a great deal to us and is deeply motivating.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.