acceptodds
Under review as a conference paper at ICLR 2027

The Operator Atlas: Learning the Geometry of PDE Solution Manifolds

Abstract

Existing PDE learning methods treat each equation in isolation and lack a representation that captures relations between equations. We introduce STAR, a contrastive encoder trained on the local differential features of solution fields. Three objectives jointly shape the latent space: a contrastive loss recovers the topological relations of the solution manifold, a reconstruction loss makes the coordinates physically decodable, and a coefficient loss anchors the shared representation to physical parameters. Trained on only three equation families, the model produces an Operator Atlas over twelve families in a train-free manner. The atlas organizes families by solution morphology, unfolds within-family parameters into continuous axes, and reveals a differential core shared by all twelve families () together with two cross-family relations: Poisson and Biharmonic, which differ in order, are mutual nearest neighbours under both metrics, and Helmholtz and mKdV are mutual nearest neighbours under centroid distance. Once frozen, the atlas solves four forward and inverse problems without per-query training: geodesic interpolation reconstructs unseen fields (median of ), patch voting identifies the equation family (), manifold position infers parameters (absolute error ), and departure from the manifold detects anomalies (AUROC ). The first three tasks are further paired with ground-truth-free uncertainty metrics, enabling reliable deployment without labelled validation data.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
operator learning, PDE solution manifolds, contrastive learning, representation calibration, zero-shot generalization

Reviews

Among:
Shape:
1234this paper, Bottom 9%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.50 over 4 reviews, above 9% of the others.
Reviewer_L9MnOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2(+1)
Read review
Paper summary

The paper trains a contrastive encoder on local features of 1D PDE solutions from three equation families, then applies it to twelve families to build a shared latent space ("Operator Atlas"). The authors analyze the structure of this space and use it, without retraining, for reconstruction, equation identification, parameter inference, and anomaly detection.

Critical strengths

The question of whether one representation can capture relations between different PDE families is interesting and relevant. The paper also examines the learned space from many angles (structure analysis, ablations, extensions to 2D and time, and downstream tasks), and it openly reports some negative results, such as the weaker transfer to 2D and time-dependent settings.

Critical weaknesses

W1. Key parts of the method are not clearly defined, which makes the results hard to interpret.

  • W1.1 It is unclear what a single training sample is. The model works on individual grid points, but the paper also refers to "patches" (30 per solution) without defining them, and q, k and sim(·,·) in the contrastive loss (Eq. 4) are not defined either.

  • W1.2 The design choices are not justified. Why does each point use exactly these six features (one of which, the second difference, can be computed from the other two differences)? Why is the atlas 16-dimensional, and how do the findings change with this dimension?

  • W1.3 It is unclear what each point in the t-SNE/UMAP plots represents (a single grid point or a whole solution). If each point is a single grid point, structures such as the Allen–Cahn energy regions or the Helmholtz u > 0 / u < 0 split may simply reflect the input value u rather than learned geometry.

W2. The claim that the model generalizes from three families to twelve is not supported.

  • W2.1 The paper does not explain how the three training families were chosen or how sensitive the results are to this choice.

  • W2.2 Three of the twelve evaluated families are the training families themselves, and some of the others are nearly the same equations in the 1D steady setting (e.g., steady heat vs. Poisson/Laplace).

  • W2.3 Results are always averaged over all twelve families. The paper never shows how the model performs on the nine unseen families alone.

  • W2.4 The paper does not show whether the pretrained representation actually helps when learning a new family.

Questions for authors

Q1. (W1.1, W1.3) Please define precisely what a "patch" and a contrastive sample are, and what q, k and sim(·,·) are in Eq. 4. Also, what does each point in the t-SNE/UMAP plots represent?

Q2. (W1.2) Why were these six input features and a 16-dimensional atlas chosen? How do the main findings change with a different atlas dimension (e.g., 8 or 32)?

Q3. (W2.1) How were the three training families selected, and how do the results change if the model is trained on a different set of three families?

Q4. (W2.2, W2.3) Could you report the results separately for the three training families and the nine unseen families?

Q5. (W2.4) To test whether the learned representation really transfers, could you compare validation curves on each unseen family for (a) the model trained from scratch and (b) the model fine-tuned from the three-family checkpoint, with the same data and budget? If (b) gives no clear advantage, the claim of a shared, transferable representation would be weakened.

1 comment

We thank the reviewer.

W1.1 / Q1 — Samples, patches, and loss

A sample is one interior grid point, not a field. With N=64N=64, an instance has 62 candidates. Input si∈R6\mathbf{s}_i\in\mathbb{R}^6 contains uiu_i, differences (ui−ui−1)/Δx(u_i-u_{i-1})/\Delta x, (ui+1−ui)/Δx(u_{i+1}-u_i)/\Delta x, second difference (ui+1−2ui+ui−1)/Δx2(u_{i+1}-2u_i+u_{i-1})/\Delta x^2, xix_i, and Δx\Delta x. A patch is this stencil, not a window. Drawing 30 points per instance with replacement, weighted by ∣∂xu∣|\partial_xu|, yields 3×1000×30=90,0003\times1000\times30=90,000 samples. Testing uses all 62 points (or the first QQ in few-shot).

For an anchor, q=oi=gop(zi)∈R16q=o_i=g_{\mathrm{op}}(z_i)\in\mathbb{R}^{16}; candidates are other batch embeddings (B=512B=512), cosine similarity, τ=0.1\tau=0.1. All same-family patches are positives; the loss groups embeddings by family. The single-positive shorthand is removed; retrieval budget qq becomes QQ.

W1.2 / Q2 — Features and atlas dimension

The second difference is linearly dependent; explicit curvature spares MLP reconstruction—feature engineering, not new information.

We retrained with opdim∈{4,8,16,32,64}\mathrm{op_dim}\in\{4,8,16,32,64\}, two seeds each; seed ranges (to 0.143), not mean±std, define noise. Dimension 16 gives separation 0.586 on both seeds (untrained: 0.496); 32 is weaker, while 4 and 8 lie within seed variability. Dimension 64 exceeds 16 on both seeds only for unseen families (0.706 vs 0.575); trained-family gain is absent: possible headroom. All 10 cross-dimension comparisons correlate positively (Spearman 0.40–0.61, p=0.002p=0.002).

W1.3 — Do structures merely reflect u?

We probed Allen–Cahn bands using uu, raw features, the atlas, and 2-D t-SNE. Since labels are deterministic under E=(u2−1)2/4E=(u^2-1)^2/4, 1-NN on uu reaches 0.994, above the atlas (0.927) and t-SNE (0.893). We removed the structural claim, retaining continuity.

Band direction is linearly readable from the atlas (logistic regression: 0.744), less from raw features (0.587): linearization only. Field-grouped ε\varepsilon ridge R2R^2 is 0.971 (raw), 0.911 (uu), 0.796 (atlas): no information beyond uu—a stated limitation.

W2.1 / Q3 — Training-family selection

Two seeds each: A is the original; B removes Helmholtz, C substitutes convection–diffusion, D adds Burgers. All exceed chance (0.083) and the untrained control (0.496), preserving angular organization. Only B consistently degrades separation (0.489/0.519 vs A: 0.586/0.586); C does not substitute (0.519/0.519). Thus Helmholtz is load-bearing, though the triple is not essential. D may help (0.579/0.692), but the 0.113 seed gap makes this suggestive. Cross-variant angles remain positively correlated (Spearman 0.21–0.53, p=0.002p=0.002).

W2.2 — Near-duplicate families

Steady source-free heat is Laplace-like; Poisson adds a source, and the atlas should quantify that proximity. Near-duplicates do not drive results: all three heat fields map to trained advection–diffusion, sharing random sources with heat. Eight unseen families achieve 3/3 identification; the sole error is Fisher–KPP→Burgers. Leakage would favor trained families, contrary to the split.

W2.3 / Q4 — Trained versus unseen families

MetricTrained (3)Unseen (9)
Identification0.667 (6/9)0.963 (26/27)
Median relative reconstruction error79.4%2.3%
Median parameter-inference error0.5440.035

Plausibly, independent random sources make these fields irregular under geodesic interpolation and nearly coincident. Anomaly AUROC is pooled (97+36 nominal vs 192 anomalous), precluding a family split. This argues against a lookup table.

W2.4 / Q5 — Fine-tuning versus random initialization

For each unseen family, we compare the checkpoint with two random initializations under identical data, optimizers, and 100 epochs; the two largest parameters are held out. Only readout heads (18,641 parameters) adapt; the frozen trunk complements zero-shot evaluation.

The checkpoint gives the lowest held-out reconstruction loss for 8/9 families, except Biharmonic (p=200/500p=200/500 extrapolation). After fine-tuning, mean identification is 0.886 vs 0.701/0.747 from random initialization, winning 9/9 against both seeds (median gap 0.167; range 0.056–0.278). Its zero-shot score (0.889) is unchanged.

Initialization gives no parameter-inference gain (0.084 vs 0.054/0.052), as in the untrained control. Reconstruction-only fine-tuning reduces anomaly AUROC by ~0.03, near the noise boundary. We limit transfer claims to identification and reconstruction geometry; parameter inference depends on readout and data.

Reviewer_mvk8OpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The authors propose STAR as a method to construct an operator atlas for operators associated with Partial Differential Equations (PDEs). The paper introduction (Section 1) provides a general summary of the paper, and related work is developed in Section 2, noting the known generalization failure of neural operators. Section 3 introduces the method (STAR) and objectives minimized to obtain the embedding, with sections 4 and 5 discussing results and featuring ablation studies. Section 6 describes the benefits of obtaining such a trained embedding, before a discussion and concluding remarks.

Critical strengths

In general, I find the approach to the defined problem creative, and I commend the authors on the graphical illustrations that make it possibly to easily follow their methodology in the text. The general idea of attempting to embed families of PDE operators using data-driven (i.e. deep & contrastive learning) approaches is quite interesting, and could have many likely implications in scientific machine learning. I am very interested in how these ideas mesh with and can be informed with classical PDE theory, and definitely encourage the authors to develop this angle further (especially several observations that the authors identify as unexpected).

Critical weaknesses
  • Despite the clarity added by the figures, it is very hard, at times, to follow the manuscript. I presume this is due to AI writing, based on my own experience (here I want to be clear that my issue is not what I believe is an artifact of AI writing itself, rather the fact that I actually had trouble reading). For example, there are times where variable values are represented without clear meaning, or times where a mathematical term is used (e.g. "cone") and it is unclear if that is an interpretative or rigorous statement. A better presentation of the text would significantly improve the quality of the manuscript.
  • A very important point I am confused by is that the contribution, in the sense of answering `why go through all the trouble of training such a model and obtaining the embedding space', is not yet there in the main text. Elements of it are, I believe, in section 6, but that is a short itemized summary that does not justify the presented method. The interpretation of the results of section 4 is welcome, but it is too qualitative to be a contribution (see also the comments that follow).
  • PDE solution spaces are generally infinite dimensional, depending on initial and boundary condition (IC/BC) richness; given fixed initial and boundary conditions, the finite-dimensional character of the sets of solutions is then implied by the smooth dependence on the parameter space of the PDE (family). But, the richness and behavior of the solution space is pre-determined by the IC/BC. So, the atlas would likely change when constructed using different domains, different IC/BCs, etc. Do the authors suggest sampling a range of both? What are the settings that reasonably determine these parameters before training STAR? And what is the usefulness?
  • This is somewhat minor, but t-SNE embeddings heavily depend on hyperparameters, and trusting a particular visualization would need a more robust argument.
  • The authors should consider engaging with the general geometric theory of PDE equivalence (see e.g. Lie or Olver) which considers different cases under which PDEs can be considered equivalent under different classes of changes of variables. These kinds of transformations do not fix the parabolic/hyperbolic/elliptic character of the PDEs though.
Questions for authors

I primarily want to ask the authors to clarify the second bullet point from above; If the goal is to move towards a data-driven embedding that is useful for a particular problem (that problem can be inference in the sense of parameter interpolation) then there are concrete metrics to compare against other frameworks. The idea that differential operators `share a core' can perhaps be made rigorous, but it is unclear what that means theoretically, and if it is limited by the problem instance; It it perhaps an artifact of the current setup? I think understanding what the authors consider the main compelling contribution of the current manuscript would help me re-evaluate my initial assessment.

2 comments

We thank the reviewer for the engaged reading. We answer the Questions for Authors first, then each weakness; details: Reviewer_3_Detailed_Responses_and_Supporting_Analyses.pdf ([R3-Supp]). Q: the main compelling contribution. STAR is an instrument, not a surrogate solver. (1) It is, to our knowledge, the first pipeline embedding twelve 1D PDE operator families into one shared, calibrated coordinate system from local differential features, trained on three families. (2) It yields measurements we believe are new: the families' local-solution subspaces collapse from random-embedding near-orthogonality (median pairwise angle 62.7 deg) to 0.22–11.23 deg; the families share a low-dimensional differential subspace aligned with the derivative channels (3 seeds); and the atlas recovers relations with classical meaning — the Poisson→Laplace degeneration appears as a manifold intersection (d_g < 0.05 at small source amplitude; App. B.13), and the Helmholtz manifold shows a sharp threshold at k = 2π with OOD k remaining on the smooth branches (App. B.10) — alongside cross-type relations (Helmholtz–mKdV). (3) The coordinates are immediately useful with zero further training (unseen-family identification 96.3%, anomaly AUROC 0.811). Two open questions follow: (OQ1) whether a family of solution manifolds admits a relation-preserving finite-dimensional embedding — classical theory covers single manifolds; no n-width analogue is known; (OQ2) whether any intrinsic classification principle for operators exists in solution-manifold geometry, when the measured geometry reproduces neither the classical trichotomy (R2 = 0.14) nor morphology (R2 = 0.29), even de-cored (R2 = 0.33; App. B.2.2). We will rewrite the close of Section 1 and Section 6 accordingly. Q: concrete metrics. Our uses are zero-training-budget, so comparisons must match that budget: untrained-encoder controls (identification 0.889 vs 0.69–0.72); random-cloud and permutation nulls throughout (e.g. App. B.2); a new from-scratch vs fine-tuned comparison on the nine unseen families (pretrained init wins 9/9, gap 0.167); and OP-based parameter inference against raw-feature controls under the same protocol, including where raw features win. Q: the shared core — rigorous meaning, artifact? The core is an empirical statement with explicit nulls: leading eigenvalue 0.982 vs random-subspace null median 0.560; absent in an untrained encoder (dimension 1 vs 3); stable across 3 seeds. Is it limited by the problem instance? Partly: it is observed in the 1D atlas and does not persist in 2D, and Open Question 2 shows that even de-cored geometry reveals no classical classification — so it is neither a random artifact nor the classical taxonomy in disguise. Its theoretical meaning — whether it reflects the common differential-algebraic content of these operators' stencils — we pose as a question, not a claim. It is conditioned on the declared configuration (domain, grid, BCs), so every reading ships with controls. IC/BC. This concern shaped the design: all twelve families share identical Dirichlet conditions u(0) = 0, u(1) = 1 on one domain and grid, chosen to reduce complexity and isolate the operator from boundary effects. Under this fixed-BC setting the reviewer's own premise is our foundation: finite-dimensionality follows from smooth parameter dependence (well-posedness), and the atlas is a chart of this finite-dimensional family. Different IC/BCs or domains would re-calibrate the instrument — a different, equally valid atlas; what should persist is the organizational signature, and our 2D/time extensions give first evidence (intra/inter 1.51/6.48 vs 31.18 deg random) while the coordinates change. A systematic IC/BC sweep is future work. Two training families also draw independent random sources per parameter — forcing richness is already sampled: if the atlas memorized fixed-configuration fields, trained families would be easiest, yet identification is 0.963 unseen vs 0.667 trained. Lie/Olver. This bridge enters the revised discussion. Classical equivalence theory gives Boolean verdicts under admissible transformations; the atlas gives a continuous proximity requiring no known transformation. The Poisson→Laplace intersection above is a classical limit relation made quantitative; the recurrent Poisson–Biharmonic neighborhood (3 seeds) has a candidate classical explanation via the w = Δu factorization; Helmholtz–mKdV links across the type boundary the reviewer notes. We do not claim atlas proximity coincides with equivalence classes; when the two notions agree is an open direction we pose. Presentation. Mechanical repairs (detailed in [Reviewer_L9Mn_Detailed_Responses_and_Supporting_Analyses]): a notation table; every symbol defined at first use (sample, patch, q, k, sim; q collision renamed Q); "cone" and similar interpretive terms made precise or removed. t-SNE. All quantitative claims are computed in the native 16-dimensional atlas space; t-SNE/UMAP are display only.

Reviewer_mvk8OpenReview ↗

Response to Rebuttal

I thank the authors for their detailed response, and engaging with my comments. Overall, I am still not convinced of the theoretical maturity and soundness of the proposed methodology, despite the discussion above clarifying the objectives of STAR and some connections to classical equivalence theory; I therefore maintain my initial score. However, I do encourage the authors to continue in refining this research direction.

Reviewer_Mos4OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes STAR, a contrastive learning approach for organizing PDE solution families in a shared latent space based on their solution morphology. The method is motivated by the idea that parameterized PDE solutions can form low-dimensional manifolds and that contrastive learning may capture geometric relationships among these solutions.

STAR constructs six-dimensional local differential features from PDE solution fields and maps them into a 16-dimensional latent representation, called the Operator Atlas. The model is trained on three PDE families and applied to twelve families. The authors report that the learned representation captures known relationships, including the Poisson–Laplace degeneration intersection, and identifies a shared three-dimensional differential core among the families.

The authors also evaluate the frozen representation on four downstream tasks: solution reconstruction through geodesic interpolation, PDE family identification, parameter inference, and anomaly detection.

Critical strengths
  1. The idea of using contrastive learning to identify geometric relationships among PDE solution families is interesting. The authors attempt to construct a shared representation that captures relationships between different PDEs, instead of learning a separate representation for each equation. The observed Poisson–Laplace degeneration intersection provides an example of how the learned space reflect known mathematical relationships.

  2. Useful visualization: The architecture diagram and latent-space visualizations help explain the proposed method.

Critical weaknesses
  1. Although the diagrams and visualizations help explain the overall approach, some results are difficult to interpret because several numerical values and evaluation metrics are introduced without sufficient explanation of their meaning or the rationale behind the selected criteria. For example, in Section 5, the authors define a successful Poisson–Laplace degeneration intersection using the condition (0 < d_g1 < 0.05), but the main text does not explain why this particular threshold was chosen. Similarly, metrics such as lambda_top and spread (H_d) are introduced, but their detailed definitions and interpretation are deferred to the appendix, making it difficult to assess the reported results from the main text alone. Another example is the abstract, which reports a PDE family identification score of 0.889 without explicitly identifying the evaluation metric, and a parameter inference error of 0.081 without clarifying its scale or normalization.

  2. The main result section which includes reconstruction of the unseen solutions (section 6) are presented in a short paragraph. Given that train-free reconstruction is a central contribution, a more detailed quantitative evaluation is needed to support the reported performance. Also the results are not reported separately for the three training families and the nine unseen families. This makes it difficult to assess whether the method performs consistently across different PDE families or whether the reported accuracy is influenced by the families included during training.

  3. The paper proposes a shared representation of different PDE families, but the practical benefit of constructing this representation is not clearly demonstrated. Existing scientific machine learning methods, including Subramanian et al. (2023), POSEIDON (Herde et al., 2024), CoDA-NO (Rahman et al., 2024), and PROSE-PDE (Sun et al., 2025), have already investigated transferable representations and generalization across PDE systems. Although these methods have different objectives and input requirements, they are relevant to the paper's motivation.

  4. The reconstruction procedure also requires further explanation. For an example, it is unclear how geodesic interpolation is used to reconstruct solutions, how the learned atlas contributes to this process, and what information is available at inference time.

  5. Some design choices are insufficiently justified. In Section 3.2, the authors construct a six-dimensional feature vector at each interior grid point, but it is unclear why these particular features are selected. Similarly, the choice of a 16-dimensional atlas representation is not adequately explained. Section 3.4 states that 30 local patches are sampled per solution instance, but the relationship between these patches and the grid-point features is not clearly described.

[1] Subramanian, S., Harrington, P., Keutzer, K., Bhimji, W., Morozov, D., Mahoney, M. W., & Gholami, A. (2023). Towards foundation models for scientific machine learning: Characterizing scaling and transfer behavior. Advances in Neural Information Processing Systems, 36. https://doi.org/10.52202/075280-3119

[2] Herde, M., Raonić, B., Rohner, T., Käppeli, R., Molinaro, R., de Bézenac, E., & Mishra, S. (2024). Poseidon: Efficient foundation models for PDEs. Advances in Neural Information Processing Systems, 37. https://doi.org/10.52202/079017-2311

[3] Rahman, M. A., George, R. J., Elleithy, M., Leibovici, D., Li, Z., Bonev, B., White, C., Berner, J., Yeh, R. A., Kossaifi, J., Azizzadenesheli, K., & Anandkumar, A. (2024). Pretraining codomain attention neural operators for solving multiphysics PDEs. Advances in Neural Information Processing Systems, 37. https://doi.org/10.52202/079017-3306

[4] Sun, J., Liu, Y., Zhang, Z., & Schaeffer, H. (2025). Towards a foundation model for partial differential equations: Multioperator learning and extrapolation. Physical Review E, 111, 035304. https://doi.org/10.1103/PhysRevE.111.035304

Questions for authors
  1. Could the authors provide a table comparing the reconstruction errors of STAR, PINNs, and FNOs for each PDE family, using comparable inputs and computational budgets? Reporting mean relative L2 errors with standard deviations or standard errors across multiple test instances would help assess the consistency of the results.

  2. What is the main practical advantage of the Operator Atlas compared with existing scientific machine learning foundation models and transferable PDE representations? Could the authors provide a comparison on a suitable downstream task, or explain why such a comparison is not appropriate?

  3. What information is required by STAR to reconstruct a solution at an unseen parameter value? Since the local features include solution values and spatial derivatives, how are these features obtained when the target solution is unknown? Could the authors explain the geodesic interpolation procedure and how the learned atlas contributes to the reconstruction?

2 comments

W5: design choices. The six features are exactly the contents of a three-point stencil — the value, the two one-sided differences, the centred second difference, position, and grid spacing — selected by ablation; making the second difference explicit spares the small MLP one linear recombination. The 16-dimensional atlas is a stable operating point, not a tuned optimum: a sweep over {4, 8, 16, 32, 64} × 2 seeds shows 16 dimensions to be stable (seed range 0.000) and above the untrained control, with the relational skeleton preserved (Spearman 0.40–0.61). A “patch” is the stencil centred on one interior grid point, the object from which that 6-dimensional vector is computed; 30 are drawn per solution, with probability proportional to |du/dx|, i.e. where local structure is non-trivial.

Q3 / W4: what reconstruction requires, and what the atlas contributes. Reconstruction at an unseen parameter value requires three things: the family identity, the parameter value, and solved support fields of the same family at bracketing parameters. The unknown solution is never touched and its features are never computed — this is what “train-free” means here. No learned component enters the arithmetic: the two bracketing ground-truth profiles are mixed with weights linear in the parameter value.

Q1: comparison table. A STAR-vs-PINN-vs-FNO table would conflate budgets: PINN trains on each target instance, FNO trains on many target-family instances, and STAR trains on nothing in the target family. The budget-matched comparison is against train-free alternatives on the same support set, reported per family as mean relative L2 with standard deviation and split into the three training families and the nine unseen ones: linear mixing in the parameter (our protocol), a constant-mean-profile baseline, and a negative ablation — atlas-weighted mixing (worse; see Q3). Headline: unseen families outperform the trained ones on every task (reconstruction 2.3% vs 79.4%; identification 0.963 vs 0.667; parameter inference 0.035 vs 0.544) — a direction that by itself argues against a lookup-table reading.

Q2 / W3. The four works cited learn transferable surrogates: a pretrained model is fine-tuned to solve downstream PDEs. STAR learns a calibrated coordinate system over operator families — 73k parameters, trained on the local features of three families, then used frozen. Its advantage is not solving PDEs better but answering, with no additional training, the questions a surrogate does not address: which operator generated the field (0.889; the strongest KNN baseline on the same features reaches 0.694), what its parameter is (median absolute error 0.081, with 47% of instances within 5% of the ground truth; the Gaussian-process baseline gives 8.58, SINDy recovers the parameter for 1 of 36 fields, random forest 0.209, KNN 0.375), and whether the field is anomalous (AUROC 0.811, against 0.622 for the strongest outlier detector, Isolation Forest;). Reconstruction and parameter inference additionally carry a ground-truth-free prior uncertainty, correlating with the actual error at 0.954, 0.771 and 0.289, and anomaly detection has no uncertainty metric.

Thank you for the additional clarifications. The response helped clarifying some implementation and reconstruction details. However, I am still not fully convinced of the practical benefit of the learned Operator Atlas compared with existing methods whether the main contribution is sufficiently demonstrated. Therefore, I will maintain my original rating.

Reviewer_z3uHOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

STAR encodes local PDE solution features into a shared atlas using contrastive, reconstruction, and coefficient-prediction objectives. The paper studies relationships across twelve equation families and evaluates reconstruction, identification, parameter inference, and anomaly detection, with preliminary spatial and temporal extensions.

Critical strengths
  1. Learning relationships across PDE families is a useful research direction, and the combination of geometric analysis with several downstream tasks is promising.

  2. The paper includes objective ablations and reports limitations of higher-dimensional and temporal transfer rather than only successful visualizations.

Critical weaknesses
  1. The theoretical justification overreaches the cited result. HaoChen et al. (2021) analyze a specific spectral contrastive loss, whereas Eq. (4) uses log-softmax InfoNCE. That result does not directly establish the claimed spectral equivalence or recovery of PDE solution geometry.

  2. Atlas-coordinate decodability is not established: the input contains the target field value, and the reconstruction head reads both the shared representation z and atlas coordinates o, allowing a bypass. Table 2 also shows noPred and atlasOnly passing the main geometric criterion, weakening the claimed necessity of all three objectives.

Questions for authors
  1. Clarification: What theoretical result applies to the actual loss and positive-pair construction?

  2. Assessment-changing evidence: Does an o-only decoder preserve reconstruction and parameter inference under clearly separated support/query splits? The supplied 11-page PDF lacks the cited appendices, so those protocols remain unverified.

1 comment

We thank the reviewer for stating precisely what would change the assessment. Missing appendices — our error. The cited appendices were not transmitted with the 11-page submission, for which we apologize. A maps every protocol referenced as unverified — reconstruction support/query splits, identification budget, parameter-inference protocol, anomaly battery — to its appendix location. Clarification (theory). The reviewer is correct, and the revision will say so plainly. HaoChen et al. (2021) prove spectral equivalence for the spectral contrastive loss; our Eq. (4) is the multi-positive log-softmax InfoNCE — the supervised-contrastive form of Khosla et al. (2020) — in which every same-family patch in the batch is a positive. For this loss the cited theorem does not apply, and we do not claim it does. The revision labels each link of the motivation chain: (a) Hadamard well-posedness gives smooth dependence of solutions on parameters — classical, and the reason a low-dimensional parameter geometry exists at all; (b) the graph-Laplacian / eigenmap view (Belkin & Niyogi) is exact for spectral losses and serves here as motivation only; (c) for the actual loss, formal support comes from the alignment–uniformity analysis of InfoNCE (Wang & Isola, 2020) and identifiability results for contrastive learning (Zimmermann et al., 2021) — these support the type of geometry recovered, not spectral equivalence; (d) the reported geometry is then verified empirically against null controls rather than asserted from theory. All "establishes"-type wording will be downgraded accordingly . The eigenmap check (supports the above scoping). We embedded four graph variants by normalized-Laplacian eigenmaps and compared their family-angle matrices to the atlas. The literal same-family-positive graph is disconnected (zero-eigenvalue multiplicity 12): its spectral embedding encodes only the partition, with cross-family angles of 90 deg by construction and 91% of family pairs degenerate; any label-only construction is permutation-equivariant and cannot yield label-independent proximity. kNN graphs built on the raw features or on the learned OP do produce angles, but with 0–90 deg spread and at most rho = 0.31 agreement with the atlas matrix. No graph-only baseline reproduces either the 0.22–11.23 deg contraction or the relation matrix : the atlas geometry is learned from the feature manifold through the encoder, not inherited from the positive-pair graph — consistent with (b), since spectral theory could only explain what the label graph already contains. Assessment-changing evidence (o-only decoder). We ran the requested test; the answer is informative in both directions. A pointwise o-only decoder does NOT preserve reconstruction: held-out relative L2 0.252 under the only legitimate protocol (ô fitted from support), indistinguishable from a constant mean-profile baseline (0.248), and 0.440 under held-out-parameter extrapolation. The diagnosis is sharp: within one field o varies more (0.72) than between fields (0.15; ratio 5.5), so o ↦ u is one-to-many and any pointwise decoder collapses to the conditional mean; adding the position coordinate improves only to 0.205, and a capacity control (same decoder on a random projection of the raw features: 0.075) shows the decoder is not the bottleneck. Crucially, none of our reported reconstruction numbers come from a decoder: Section-4 reconstruction blends the ground-truth profiles of the two bracketing support fields with weights from the atlas geodesic, achieving 0.005 (interpolation) and 0.021 (extrapolation) relative L2 — using only the geometry of o plus support ground truth, never z. The bypass the reviewer identifies exists only in a training auxiliary that no evaluation reads; what the coordinates carry is geometry — retrieval, interpolation, identification, parameter inference (already a purely o-geometric estimator), anomaly scores — not a pointwise code of u . On Table 2. We agree that noPred and atlasOnly passing the main geometric criterion shows the contrastive objective alone suffices for that criterion, and the revision scopes the claim accordingly: the three objectives are not jointly necessary for geometry. Each loss constrains what it reads — contrastive organizes family topology, the prediction objective anchors scalar field content, the coefficient objective fixes the parameter direction — and the necessity we claim is per property, as the downstream columns of Table 2 show. The stronger wording will be removed.

Comments on the paper

1 comment

Dear Reviewers,

The authors have posted their rebuttals. As the author-reviewer discussion period is coming to a close soon, please read their responses and engage in the discussion. At a minimum, please acknowledge that you have read the authors’ responses. Thank you for your time and effort.

Best,

AC

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.