acceptodds
Under review as a conference paper at ICLR 2027

Typed Frame Search: Reference-Free Coordinate Selection for PDEs

Abstract

Many partial differential equation (PDE) solutions become stationary profiles in moving or rescaled coordinates, yet an accurate field fit need not identify a unique coordinate representation. We introduce typed frame search, which selects an explicit frame–profile program from the PDE and prescribed initial and boundary data, without observing future solution trajectories. Candidates from a finite transformation grammar are fitted to physics and data constraints, ranked on independent validation designs, and penalized for frame complexity to favor canonical representations. We establish conditions for frame identifiability and reliable selection within the fitted bank. Across 28 in-bank configurations from 14 PDE families and ten seeds, the method recovers 96.8% of canonical frames (271/280), with median relative L2 error of 1.43 × 10^-6. Compared with a degree-four continuous-clock baseline using the same number of optimization starts, it achieves approximately 90-fold lower median error and 66.6-fold lower median search time. Tests with unseen profiles and two-dimensional affine flows recover all 50 and 60 frames, respectively. Ablations distinguish the constraints that determine the field from the complexity penalty that resolves equivalent representations. These results support interpretable coordinate recovery within a finite grammar, with limitations from initial-data noise and unsupported transformations.

Primary area:
neurosymbolic & hybrid AI systems (physics-informed, logic & formal reasoning, etc.)
Keywords:
frame search, coordinate transformations, PDEs, program search, reference-free selection, interpretable scientific machine learning

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_fDpbOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Typed Frame Search, a reference-free framework for identifying interpretable coordinate transformations for PDE solutions. The method searches over a predefined grammar of frame transformations and profile families, fits candidates using PDE residuals and initial/boundary conditions, and selects a canonical representation through validation and complexity regularization. Experiments across multiple synthetic PDE families demonstrate high frame recovery accuracy and improved efficiency over continuous-clock baselines.

Critical strengths
  1. Interpretable framework. The method provides explicit and interpretable coordinate transformations without requiring future solution trajectories. The separation between field recovery and canonical frame identification is well motivated.

  2. Comprehensive evaluation within the defined setting. The paper evaluates multiple PDE families, includes useful ablation studies, and reports strong in-bank recovery performance. The comparisons with continuous-clock baselines demonstrate the computational benefits of the structured search space.

Critical weaknesses
  1. Limited methodological novelty and generalizability. The proposed approach relies on a manually designed grammar of transformations and profile families, combined with standard PDE residual optimization and complexity-based model selection. While effective within this predefined space, the method does not demonstrate the ability to discover transformations outside the grammar. The theoretical results also primarily characterize identifiability and selection within a finite candidate set. This limits the broader novelty and applicability of the approach.

  2. Limited evidence of downstream benefits. Although the recovered frames are interpretable, their practical advantages are less convincing. The downstream numerical experiments show improvements mainly at coarse resolutions, while the selected-frame solver becomes less accurate at finer resolutions and remains slower when search costs are included. Furthermore, the benefits over simpler reduced-coordinate representations are not fully established. These observations weaken the practical significance of the proposed framework.

Questions for authors
  1. How could the proposed framework handle PDEs whose useful coordinate transformations are not included in the predefined grammar? Could the authors clarify its advantages over classical symbolic similarity reduction methods?

  2. Could the authors provide stronger evidence of downstream benefits, particularly for more challenging PDEs where the appropriate coordinate transformation is not straightforward?

  3. How sensitive is the method to the design and size of the candidate grammar, and what would be required to extend it to more general transformation families?

3 comments
AuthorsOpenReview ↗

Official Comment by Authors

Thank you for recognizing the interpretability and for your constructive comments on novelty, grammar dependence and practical value. They motivated same-bank structure-extraction controls and stronger PDE baselines.

W1: Proposed method, novelty and generalizability.

We propose Typed Frame Search to jointly infer a coordinate-law type and its profile from the PDE and prescribed IC/BC, without future trajectories. Its contribution combines joint frame/profile inference with canonical selection among field-equivalent programs. Translation is equivalent to translation–dilation with zero dilation. IC anchoring and description cost favor a concise law. Removing that cost lowers strict recovery from 96.8% to 68.2% while field equivalence stays 100% (Sec. 6.2, p. 8). The optimizer is standard. Guarantees condition on the fitted bank.

We give continuous fitting the same bank, then apply projection and joint refinement. Typed recovers 69/70 strict labels, versus 70/70 after refinement. Both recover all clocks at RMSE below 0.01. Median complete CPU costs are 3.24 seconds for Typed, 12.38 for fixed degree one plus refinement and 27.46 for degree two plus refinement. This supports lower-cost structure inference with accurate clock recovery (Table 5 (p. 9), Table 45 (p. 38)).

IC splines recover 50/50 unseen-profile frames and 2D affine tests recover 60/60, supporting profile transfer and dimensional extension (Secs. 6.3–6.4, pp. 9–10).

W2: Downstream benefit, refinement and total cost.

An amplitude- and dimension-matched physical-xx control isolates coordinate choice. At 128 samples, the paired selected/control error ratio is 0.374, with 95% family-cluster interval [0.105,0.884][0.105,0.884]. Against validation-selected continuous coordinates it is 0.999, [0.997,1.001][0.997,1.001], indicating similar downstream precision (Table 46 (p. 38)).

The PDE test fixes IC data at 2572257^2 with independently checked references. At diffusion 0.005 and perturbation 0.03, selected errors decrease through 0.0621, 0.0302, 0.0163 and 0.00856 on grids 332,652,1292,257233^2,65^2,129^2,257^2 (Table 49 (p. 41)). App. F.3 (pp. 38–39) separately tests IC resolution, remapping and reference discretization.

Costs include search, setup and solve. For pure affine nondiffusive flows at 100-IC reuse, Typed takes 1.055–1.091 times continuous-affine wall time and 2.09–2.17 times coefficient-informed characteristics, with matching errors (Table 53 (p. 42)). The comparisons distinguish structure-inference efficiency from solver performance.

Q1: Unrepresented transformations and symbolic similarity reduction.

An absent law is handled through explicit vocabulary extension. Adding damping with acceleration, damping with oscillation, and acceleration with dilation raises field success from 13/45 to 45/45 and lowers median error from 0.138 to 2.94×10−62.94\times10^{-6} (Table 33 (p. 28)).

Symbolic similarity reduction establishes exact invariance. Our numerical criterion compares coordinate programs for the given IC/BC, including sampled spline profiles. The affine application also ranks approximate frames and evolves their residual dynamics. Our SymPy pdsolve calls support 13/28 original configurations, with median error 6.95×10−86.95\times10^{-8} versus 6.87×10−76.87\times10^{-7} for Typed. We report its supported-subset accuracy advantage and all attempts (Table 3 (p. 8)).

Q2: More challenging PDEs and stronger application evidence.

The 80-case affine extension retains diffusion and nonaffine residual dynamics after coordinate selection. At ν=0.005\nu=0.005, ϵ=0.03\epsilon=0.03 and grid 1292129^2, physical, selected and continuous upwind median errors are 0.0897, 0.0163 and 0.0443, compared with 0.00193 for the semi-Lagrangian control. Across the fixed grid/time menu, a 1% error target is attained by 10/20 selected cases and 20/20 semi-Lagrangian cases. The latter has the better accuracy–cost frontier. These results support motion removal within the shared upwind backend and clarify the solver-dependent downstream benefits (Table 51 (p. 41), Table 52 (p. 42)).

Q3: Grammar sensitivity and extension requirements.

On 70 instances, removing acceleration, oscillation or moving-self-similar programs lowers strict recovery from 98.6% to 91.4%. Adding 8/16 nearby clocks raises candidates from 55 to 95/135, preserves 98.6% recovery and increases median objective calls from 5,465 to 9,037/12,888.5 (Table 32 (p. 28)).

Extension requires differentiable, domain-invertible maps, IC anchoring, bounds, equivalence rules and independent validation. At fixed confidence, score bound and separation, sufficient validation grows logarithmically with bank size. Nearby alternatives may narrow the gap (Sec. 4, pp. 5–6).

Thank you again for your constructive comments. We would appreciate your taking another look at the revised evidence.

Thanks for the additional experiments. The same-bank comparisons show a useful efficiency advantage, but my main concern about generalizability remains.

In particular, the new composition cases require manual grammar extension, while continuous fitting already handles them. I would consider raising my rating to 3 if the method could automatically discover new combinations of existing transformation primitives, without manually adding each combination, while remaining competitive with continuous fitting.

I appreciate the improvements, but for now I'll keep my rating at 2.

AuthorsOpenReview ↗

Official Comment by Authors

Thank you for the precise clarification and for considering the additional evidence. We agree that automatic composition, rather than registering each combined law, is the relevant test. We have implemented and evaluated that capability.

Automatic generation from unchanged primitives.

The extension reuses the eight clock primitives already present in the original eleven programs. A uniform rule inserts unused primitives, composes their field actions and merges commuting arrangements, producing 144 words of depth at most three. Translation–dilation order is retained. An eight-word beam expands using training scores, jointly fits the five profiles and selects using independent validation. No target-specific combination, reference label or future interior field enters fitting (Sec. 3.5, Sec. 6.5, App. G).

For example, ordered composition of dilation and oscillatory translation generates s(t)=veρdsin⁡ds(t)=v e^{\rho d}\sin d, although that combined clock was never registered. Together with quadratic damping, automatic search recovers the ordered three-primitive program, with field error 2.38×10−62.38\times10^{-6} (App. G.4).

Unseen combinations and competitive continuous controls.

We test all 28 new two-primitive structures absent from the original catalogue, plus 12 triple structures chosen by a fixed grammar-only rule. Each has three coefficient settings and five collocation seeds, yielding 420 and 180 instances. These repetitions are not counted as independent structures. Settings were frozen before compositional evaluation. All methods share public data, profiles, objective and validation. Continuous controls also receive automatic projection and joint refinement over the same generated vocabulary. Complete CPU cost includes every upstream stage.

On pairs, beam and degree-two continuous fitting plus extraction (CC2+S) both recover 420/420 active programs, with median field errors 2.33×10−62.33\times10^{-6} and 2.32×10−62.32\times10^{-6}. Median CPU times are 2.20 and 3.64 seconds. The paired CPU ratio is 0.605, with structure-cluster interval [0.579,0.633][0.579,0.633] (Table 6 (p. 10), Table 57 (p. 46)).

Triples expose a harder search tradeoff. Beam recovers 168/180 at 6.31 seconds, versus CC2+S's 175/180 at 4.05. The already-defined exhaustive control, extended unchanged to all triples, recovers 179/180 at 10.20 seconds. Continuous fitting of the same primitive basis followed by extraction recovers 180/180 at 15.74. We therefore report the beam and exhaustive policies separately. Both generate combinations automatically, with different recovery and cost. The earlier 45 composition instances and 70 original-family instances are all recovered by beam under the common active-program metric (Table 56 (p. 44)).

Checks and scope.

The appendix includes degree/warm-start controls, depth and penalty ablations, complete outcomes and same-core serial timing checks. Independent coefficient-ODE integration validates all 129 new transport references. The remaining exhaustive miss is retained and diagnosed as a numerical-fit failure on an expressible word. The expanded exhaustive study is identified as coverage extension after primary analysis, with unchanged settings.

These results provide the requested automatic discovery of new combinations within the existing primitive vocabulary and quantify competitiveness against strong continuous controls. They support bounded compositional generalization, with the depth and optimization tradeoff made explicit. Thank you again for the helpful guidance and for taking another look at the revised method and evidence.

Reviewer_1mxZOpenReview ↗
Rating
3
Confidence
1
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes a framework called "Typed Frame Search" for recovering the interpretable coordinate representation of the solution to partial differential equations (PDEs) from control operators and given initial/ boundary conditions, without requiring the observation of the reference solution trajectory.

This method enumerates 55 candidate frame-contour combinations from the predefined coordinate programs and contour representations. Each candidate solution is fitted by minimizing a target function that includes PDE residuals, initial/boundary conditions, optional conservation constraints, and dimensionality reduction coordinate residual drift. The final representation is selected through independent validation design and the penalty term for the number of frame parameters.

The core contribution lies in expressing the coordinate recovery as an explicit finite structure selection problem and emphasizing the distinction between field accuracy and standard representation recovery.

Critical strengths

This paper points out the ambiguity in the coordinate representation. It evaluates multiple families of PDEs, coefficient variations, different time transformation mechanisms, and two-dimensional affine flows. The target ablation experiments in Table 2 are particularly valuable. These results illustrate the effects of different constraint terms.

Critical weaknesses

The method compared the Continuous Clock with degree 4 and achieved a speed increase of 66 times. However, in Appendix Table 6, degree 2 yielded a better RMSE. The article did not provide sufficient evidence to explain the reason for comparing it with degree 4.

Questions for authors

Why was the degree 4 Continuous clock chosen as the baseline for comparison? Why does the Continuous clock method yield the lowest Median RMSE when the degree is set to 2?

1 comment
AuthorsOpenReview ↗

Official Comment by Authors

Thank you for carefully checking the degree comparison and for the helpful questions. Your observation is correct: degree 2 has the lowest median clock RMSE in submitted Table 6, now Table 8 (p. 15). We address the weakness and both questions explicitly below.

W1 & Q1: Why was degree 4 used as the baseline?

Degree 4 provides a flexible continuous parameterization of translation, log-amplitude and log-dilation. The full benchmark includes nonpolynomial similarity clocks such as log⁡(t/t0)\log(t/t_0) in heat and porous-medium solutions. Five profiles and eleven starts per profile give 55 continuous starts, matching Typed's candidate count. Equal starts do not imply equal objective evaluations or time, which we report separately (Sec. 3.4, p. 5).

We agree that a fair assessment should also include other degrees and tuning strategies. The revision reports degrees 1–4, validation-selected degree, warm starts and budget controls. The submitted 66.6-fold search-time ratio concerns its specified degree-four implementation. It is now presented alongside the broader degree-dependent accuracy–cost evidence (Sec. 6.1, pp. 7–8, App. E.2, pp. 24–26).

Q2: Why does degree 2 yield the lowest median clock RMSE?

The original degree table covers 40 instances from four nonconstant-clock families. With Δt=t−t0\Delta t=t-t_0, accelerating translation has s∝(Δt)2s\propto(\Delta t)^2, quadratic-time damping has g∝(Δt)2g\propto(\Delta t)^2, and exponential dilation has h∝Δth\propto\Delta t. Here s,g,hs,g,h are translation, log-amplitude and log-dilation clocks. Thus 30/40 true clocks are exactly representable by degree 2 or below. The oscillatory family requires approximation (App. B, pp. 15–16).

A degree-four polynomial contains a degree-two clock when the higher coefficients are zero. This describes approximation capacity. Actual fitting minimizes a finite-budget, nonconvex PDE/condition objective, whereas clock RMSE is a separate evaluation metric. Additional coefficients can affect optimization and initialization sensitivity, so increasing degree need not reduce measured clock RMSE. The subset's low-order structure and the optimization controls below help interpret the result without assigning it a unique cause.

The two reported comparisons also have different populations and metrics: the original degree table measures clock RMSE on 40 instances, while the main table measures field relative L2L^2 on 280. We now label both explicitly.

Additional evidence addressing W1, Q1 and Q2.

In the supplementary degree study, median clock RMSE on the 40-instance subset is 2.61×10−62.61\times10^{-6} for degree 2 and 5.51×10−55.51\times10^{-5} for degree 4. On all 280 instances, their median field errors are 1.34×10−31.34\times10^{-3} and 1.13×10−41.13\times10^{-4}. Both outcomes and every family's results are retained (Table 26 (p. 24), Table 27 (p. 24), Table 30 (pp. 25–26)).

Initializing degree 4 from degree 2 lowers full-suite median field error to 3.33×10−53.33\times10^{-5}. Selecting among degrees 1–4 by independent public validation gives 3.47×10−53.47\times10^{-5}. Warm-pipeline cost includes degree-two fitting, and validation-selection cost includes all four searches (Table 28 (p. 25)). Function caps, optimization-time allocations and actual end-to-end usage, including setup, validation and overruns, are reported in Table 29 (p. 25).

A further 70-instance comparison uses the same frame/profile bank and tests fixed-degree fitting followed by structural projection and refinement. Fixed degrees one and two each recover 70/70 strict labels at median complete CPU costs of 12.38 and 27.46 seconds. Typed recovers 69/70 at 3.24 seconds. All four fixed degrees, projection-only variants and paired intervals are reported (Table 44 (p. 37), Table 45 (p. 38)). This directly examines the cheaper alternative of fixing a lower degree before extracting structure.

App. E.1 (pp. 23–24) documents the supplementary protocol. Together, the added controls make the effects of degree, optimization strategy and complete inference cost explicit.

Thank you again for identifying this comparison issue. We would appreciate your taking another look at the clarified metrics and expanded controls.

Reviewer_m5xJOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes a typed frame search for finding the interpretable coordinate laws from PDE solutions. The paper tries to answer that given a solution of a PDE function, how does motion, rescaling changes in this field. This paper has identified different profile banks with different combinations of different characteristics of the system (?), then uses L-BFGS-B to do the frame profile fitting and validation. Under different in-bank settings and different PDE families, the proposed method did recover most of the frames and outperformed its competitors. The authors have also done ablation studies for different settings to validate the results.

Critical strengths
  1. From my understanding, this paper has conducted thorough and complete experiments, including a thorough ablation study.
  2. The paper has made a clear classification of the problem setting and the method analysis.
Critical weaknesses
  1. The results are mainly on the synthetic PDEs. Whether this method can be generalized to broader applications or real-world PDEs is not sure. I think the authors could at least provide some intuitive discussions or simple applications to analyze the generalizability of the method.
  2. I think the comparison is rather incomplete. There are some methods discussed in the related work but not compared in the quantitative results. The authors should make more thorough comparisons.
  3. The authors could make a table to summarize and list the profile family, similar to the optimization settings stated in the paper, to make it clear.
Questions for authors

Please see the detailed comments above. I am not familiar with this research field. I can only provide general comments on experiment design, motivations, etc. I think the main issue is that the arguments in the paper are not fully backed up by the paper. The performance comparisons and applications could be improved.

2 comments
AuthorsOpenReview ↗

Official Comment by Authors

Thank you for recognizing the experiments and ablations, and for your constructive suggestions on applications, comparisons and presentation. We have strengthened these aspects and respond to each concern below.

W1: Broader applicability and simple applications.

We added a complete 2D transport–diffusion application with four flows (rotation, moving rotation, shear and anisotropic flow), two diffusivities, two nonaffine perturbation levels and five initial fields, giving 80 cases. The affine-search extension fits and validates coordinates using the given PDE and initial observations, then numerically evolves the remaining advection and diffusion. Future fields are opened only for evaluation. This illustrates coordinate selection when motion removal simplifies the dynamics without making the entire solution stationary (Sec. 6.4, p. 10, App. F.4, pp. 39–42).

All methods receive the same 2572257^2 initial samples, with independently checked references. At ν=0.005\nu=0.005, ϵ=0.03\epsilon=0.03 and solver grid 1292129^2, median field errors are 0.0897 for physical upwinding, 0.0163 for selected coordinates, 0.0443 for continuous coordinates and 0.00193 for the semi-Lagrangian control (Table 49 (p. 41)). Across the fixed grid/time menu, a 1% error target is attained by 10/20 selected-frame cases and 20/20 semi-Lagrangian cases. The application demonstrates coordinate benefits within the shared upwind backend and the stronger solver's accuracy–cost advantage (Table 51 (p. 41)).

The original unseen-profile and 2D affine tests recover 50/50 and 60/60 frames. Explicitly extending the grammar also solves 45/45 new composition cases. These tests examine changes in profile shape, dimension and dynamics, with the scope of each extension stated (Secs. 6.3–6.4, pp. 9–10, Table 33 (p. 28)).

W2: More complete quantitative comparisons.

We added continuous fitting followed by projection and refinement using the same eleven frames, five profiles, public constraints and validation rule. Complete costs include all upstream searches and postprocessing. On 70 instances, Typed recovers 69/70 strict labels in median 3.24 CPU seconds. Fixed degree-one and degree-two fitting followed by refinement each recover 70/70, in 12.38 and 27.46 seconds. Both routes recover all clock functions at RMSE below 0.01. All fixed degrees 1–4, validation-selected degree, warm starts and paired comparisons are reported (Table 5 (p. 9), Table 44 (p. 37)).

We also ran SymPy 1.14 pdsolve on all 28 original configurations. Thirteen calls return the required representation, and the remaining 15 invocation outcomes are retained. On the supported subset, median field error is 6.95×10−86.95\times10^{-8} for SymPy and 6.87×10−76.87\times10^{-7} for Typed. Each PDE is submitted to SymPy once, while ten numerical seeds give 130 Typed fits. This recognizes symbolic solving's accuracy on supported templates alongside numerical selection of candidate representations (Table 3 (p. 8), Table 31 (p. 27)).

The degree, validation-selection, warm-start and budget studies appear in App. E.2 (pp. 24–26), Tables 26–30. The submitted 66.6-fold timing ratio remains tied to its degree-four implementation. Sec. 2 (pp. 2–3) distinguishes methods by their inputs and outputs. Trajectory-based discovery uses additional solution observations, while a general field solver needs structure extraction to assess canonical recovery. We therefore compare the semi-Lagrangian method directly on complete-PDE accuracy and cost. App. E.1 (pp. 23–24) specifies the supplementary protocols.

W3: A profile-family summary table.

Thank you for this helpful suggestion. Table 1 (p. 3) in Sec. 3.1 lists the Gaussian, positive-quadratic, tanh, sinusoidal and IC cubic-spline families, their formulas and fitted profile-parameter counts. Five profiles combined with eleven frame programs give 55 candidates. Spline coefficients come from the public IC samples, with no learned profile parameter, accommodating shapes without a prescribed simple analytic formula.

Thank you again for your time and helpful comments. We would be grateful if you could take another look at the revised manuscript and these additional comparisons.

AuthorsOpenReview ↗

Follow-up on our response

Thank you again for your constructive feedback and for recognizing the breadth of our experiments and ablation studies. In our earlier response, we addressed your three suggestions with a two-dimensional transport–diffusion application, additional quantitative comparisons, and a summary table of the profile families.

The central contribution of Typed Frame Search is to recover an interpretable coordinate law from the PDE and its initial and boundary conditions, without using future solution trajectories. By jointly fitting the coordinate frame and solution profile, then using independent validation and a complexity penalty to select a concise representation, the method approximates the solution field and describes how it moves and rescales over time.

If you have a chance to revisit our response, we would appreciate hearing whether the added experiments and clarifications address your main concerns. Thank you for your time and consideration.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.