acceptodds
Under review as a conference paper at ICLR 2027

Hypothesize, Evaluate, Refine: A Scientific Agent for Open-Form PDE Discovery with Unknown Spatial Coefficient Fields

Abstract

Understanding dynamics in heterogeneous media requires identifying both the governing differential operator and the spatial properties that determine its local action. Discovering these jointly is difficult: operator scope changes the law, while flexible coefficient fields can mask structural error. Existing open-form methods generally assume scalar or explicitly parameterized coefficients; nonparametric field estimators typically operate within prescribed operator libraries. We introduce Hypothesize, Evaluate, Refine for PDE Discovery (HER-PDE), coupling a scientific Agent with the Hypothesis Evaluation Interface (HEI). The Agent composes operators and field nodes into complete expression trees; HEI fits their time-invariant spatial fields and scores short-horizon prediction across two excitations within its supported evolution class. This separates structural reasoning from continuous estimation while preserving operator scope and shared-field identity. On seven two-dimensional heterogeneous PDE families with 5% state noise, including nonlinear cubic absorption, HER-PDE recovers 70/105 exact structures and achieves 94.3% Pass@3, compared with 39/105 exact recoveries for recursive typed evolution using the same HEI. Registered-structure fits attain median field correlation 0.891 across 65 fields; targeted controls test noise, solver, and morphology shifts. In a real spatial-drug case study, an Agent-proposed density-inhibitory effective model reduces continuous-prediction error by 15.7% relative to a refitted linear reference in post-discovery technical-repeat validation on measured microbial trajectories.

Primary area:
neurosymbolic & hybrid AI systems (physics-informed, logic & formal reasoning, etc.)
Keywords:
PDE discovery, scientific agents, symbolic regression, variable-coefficient PDEs, nonparametric spatial fields, heterogeneous media

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

The paper introduces HER-PDE, a scientific-agent framework for discovering open-form PDEs with unknown, nonparametric, time-invariant spatial coefficient fields. The key idea is to separate structural hypothesis generation from continuous coefficient estimation: an LLM agent constructs complete PDE expression trees, while the proposed Hypothesis Evaluation Interface (HEI) fits the spatial fields and evaluates candidates using short-horizon prediction across two excitations sharing the same underlying fields.

Critical strengths
  1. The combination of open-form structural discovery with nonparametric spatial fields is important and technically challenging. Explicitly preserving operator scope and shared-field identity is particularly valuable and distinguishes the framework from fixed-library variable-coefficient approaches.
  2. The evaluation includes 35 environments and 105 runs, multiple search baselines, field-sharing ablations, sealed prediction, field recovery, noise and morphology shifts, and a real-data case study. The improvement over matched controls is substantial: 70/105 exact recoveries versus 39/105 for typed evolution and 27/105 for the Flat Agent.
Critical weaknesses
  1. The discovery space is still substantially constrained by HEI. Although described as “open-form,” practical discovery is limited to the evaluator’s supported evolution class, including scalar autonomous dynamics, specified derivative support, time-invariant coefficient fields, and affine field dependence. Thus, the framework does not yet establish that the approach generalizes to broader PDE systems, time-varying fields, coupled states, or more general field interactions.
  2. The contribution of individual Agent-design components is not isolated. The Flat Agent ablation simultaneously removes the phase transition, minimum-hypothesis requirement, raw-data freeze, global/local search instructions, and score-cadence reminders. Consequently, the large 70/105 vs. 27/105 improvement establishes the effectiveness of the controller as a whole, but does not reveal which components are responsible for the gain. This weakens the mechanistic evidence for the proposed agent design.
Questions for authors

Could you please explain or address the weaknesses raised above?

2 comments

W1:

The discovery space is still substantially constrained by HEI. Although described as “open-form,” practical discovery is limited to the evaluator’s supported evolution class, including scalar autonomous dynamics, specified derivative support, time-invariant coefficient fields, and affine field dependence. Thus, the framework does not yet establish that the approach generalizes to broader PDE systems, time-varying fields, coupled states, or more general field interactions.

A:

Thank you for your thoughtful comments and valuable ablation suggestions. We have added the comparisons you suggested and explain below how we have clarified the scope of our work.

We appreciate you pointing out that “open-form” could be read more broadly than the class HEI currently evaluates. We apologize that the original manuscript did not make this distinction clear enough.

Open-form search composes differential operators, nonlinear state functions and unknown spatial fields without a predefined term library. HEI evaluates scalar autonomous evolution equations with time-invariant fields entering affinely, including through their derivatives.

This class includes important physical models. For example, ut=∇⋅(D(x,y)∇u)u_t=\nabla\cdot(D(x,y)\nabla u) describes source-free heat conduction in heterogeneous materials with uniform volumetric heat capacity, or groundwater flow with uniform storage. The field DD represents spatially varying thermal or hydraulic diffusivity. Replacing the flux divergence by D(x,y)ΔuD(x,y)\Delta u removes the contribution from spatial variation in DD, changing the conservation law. Searching field placement therefore addresses a physical modeling choice.

Nonlinear reaction terms also describe biological growth. In our measured bacterial growth–migration case (Section 4.4, Figure 3), the Agent proposed ut=βuxx+a(x)u−u2u_t=\beta u_{xx}+a(x)u-u^2. Here diffusion represents migration between wells, a(x)a(x) describes spatially varying growth under antibiotic exposure, and −u2-u^2 represents density-dependent inhibition. This connects equation discovery to an interpretable model of measured population dynamics.

Although we do not claim generalization to broader PDE systems, coupled states, time-varying fields or non-affine field interactions in the present study, your question highlights valuable directions for extending our framework. We thank you for bringing these directions into focus and plan to pursue them in future work. We have added Appendix K, “Future work” (p. 48), outlining extensions to coupled states, evolving fields and more general field interactions.

To address your concern, we have revised the Introduction, paragraph 3 (pp. 1–2), and Section 3.1, paragraph 2 (p. 5), to explain how open-form construction relates to HEI's evaluation scope. We have also made the supported class explicit in Section 3.2's opening paragraph (p. 5), so readers can see this boundary when the evaluator is introduced. Revisions and additions are marked in green in the revised manuscript. We are grateful for your suggestion to make this distinction clearer.

W2:

The contribution of individual Agent-design components is not isolated. The Flat Agent ablation simultaneously removes the phase transition, minimum-hypothesis requirement, raw-data freeze, global/local search instructions, and score-cadence reminders. Consequently, the large 70/105 vs. 27/105 improvement establishes the effectiveness of the controller as a whole, but does not reveal which components are responsible for the gain. This weakens the mechanistic evidence for the proposed agent design.

A:

The same HEI supports HER-PDE, Flat and typed evolution, letting us study how different controllers organize hypothesis testing.

The original Flat comparison measures the controller as a whole. We therefore ran four variants, each removing one mechanism while retaining the others: seven families, the first environment of each, and three runs per environment, giving 84 additional runs. Full HER-PDE and original Flat below are the corresponding 21-run subsets of the primary benchmark. We also added 21 matched-evaluation Flat runs, described below.

ControllerExact recoveryMean HEI evaluations
Full HER-PDE16/2131.05
Without minimum-hypothesis requirement14/2130.81
Without raw-data freeze10/2119.29
Without global/local search guidance11/2123.29
Without score reminders15/2128.57
Flat Agent, original7/2115.43
Flat Agent, matched HEI count12/2131.05

The four component ablations retain the same data, HEI, language, model request settings and budget ceilings: 36 exploration turns, two selection turns and 96 HEI evaluations. They retain the phase schedule; removing the minimum-hypothesis requirement keeps the turn-16 checkpoint. Appendix C.3, pp. 30–31, and Table 19 report this component study. The additional matched-evaluation control is reported here.

The controller affects both how extensively the Agent tests hypotheses and how it uses their feedback. Under the same 96-evaluation ceiling, HER-PDE used 31.05 evaluations per run, versus 15.43 for Flat. Sustaining this search is itself part of the controller's contribution. To examine recovery at the same evaluation count, we required each new Flat run to reach its paired HER-PDE run's actual count, also using that count as its cap. These quotas ranged from 8 to 61. Flat received the same observations, DSL, HEI and model request settings, plus a remaining-budget counter. It freely analyzed data and revised candidates, without HER-PDE's two-phase schedule or global/local guidance. We allowed up to 128 exploration turns, followed by two selection turns, to give Flat enough opportunity to reach its quota.

All 21 runs reached their exact quotas and submitted a final equation: 652 HEI evaluations for each method, averaging 31.05 per run. Matched-count Flat recovered 12/21 structures (57.1%), up from the original 7/21, while HER-PDE recovered 16/21 (76.2%). These results support two complementary contributions of the combined controller: sustaining more hypothesis testing within the available budget, and achieving higher recovery at matched evaluation counts. The latter advantage is 19.0 percentage points on this subset.

Across the full benchmark, HER-PDE recovers 70/105 structures versus Flat's 27/105 with the same HEI. The new controls strengthen the evidence for individual design choices and the combined controller at matched evaluation counts.

We sincerely appreciate the time and care you have devoted to reviewing our work. Your feedback has helped us strengthen both the paper and its supporting evidence. If our clarifications and additional results have addressed your main concerns, we would be very grateful if you would consider revisiting your assessment and score. Should any concerns remain, we would sincerely welcome the opportunity to discuss them further.

Reviewer_EY1DOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper introduces HER-PDE, an LLM-based scientific agent that discovers PDE structures and unknown, time-invariant spatial coefficient fields from noisy observations. Candidate equations are represented as recursive expression trees that preserve differential-operator scope and shared-field identities. Its hypothesis evaluator estimates coefficient fields with Gaussian kernels and differentiable short-horizon simulations, fitting shared fields across different excitation trajectories. The agent combines data analysis, global hypothesis exploration, and local refinement guided by numerical feedback. Validation covers seven PDE families with five spatial environments each, totaling 105 runs at 5% observation noise, and reports a 66.7% exact structural recovery rate, along with comparisons, held-out prediction evaluations, and ablations on the search strategy and field-sharing constraints.

Critical strengths

The paper tackles a meaningful challenge in PDE discovery, namely recovering both the governing structure and unknown spatial coefficient fields, and explicitly preserving operator scope and shared-field identities is a useful part of the formulation. The evaluation is reasonably thorough for the setting considered, with multiple PDE families, spatial environments, and targeted ablations offering useful evidence on the contributions of the search strategy and field-sharing constraints.

Critical weaknesses

Describing SEQL/HEQL as seeking compact coefficient formulas seems inaccurate, since these methods learn position-dependent coefficients without requiring compact symbolic laws. Xu et al. [1] also discover PDEs with unknown spatial fields inside differential operators, though within a prescribed flux-form template. These precedents call for a more precise qualification of the claimed novelty. The GP-IDENT-style control provides a useful fixed-library comparison, but additional comparisons with closely related methods such as SEQL/HEQL, on mutually supported cases and common metrics, would better support the claimed advantage over prior approaches.

[1] H. Xu, D. Zhang, and N. Wang. “Deep-learning based discovery of partial differential equations in integral form from sparse and noisy data.” Journal of Computational Physics, 445:110592, 2021.

Questions for authors

See weaknesses

5 comments

Describing SEQL/HEQL as seeking compact coefficient formulas seems inaccurate, since these methods learn position-dependent coefficients without requiring compact symbolic laws. Xu et al. [1] also discover PDEs with unknown spatial fields inside differential operators, though within a prescribed flux-form template. These precedents call for a more precise qualification of the claimed novelty.

A:

Thank you for pointing us to these closely related studies and suggesting direct comparisons on problems supported by both approaches. We have revisited their relationship to HER-PDE and completed the additional experiments you recommended. We first explain the classification and the methodological distinction, then describe how we made the comparisons and what we learned.

We apologize that the main-text wording blurred the distinction between the problems studied and the mechanisms used to estimate coefficients. The submitted Appendix I already states that we classify “the most permissive capability demonstrated in each paper's equation-discovery experiments.” We retain that criterion and clarify its application to coefficient regimes in the revised text. Figure 1's axes build on Lou et al. [2]. For example, Zhang et al. test advection–diffusion with f(x)=−1.5+cos⁡(2πx/5)f(x)=-1.5+\cos(2\pi x/5) and Burgers' equation with f(t)=−(1+sin⁡(t)/4)f(t)=-(1+\sin(t)/4) [1, Eqs. (11)–(12)]. SEQL/HEQL construct nonlinear expressions, while the true coefficients in these experiments follow explicit analytic laws. This places them in “Open-form library × Expressible by equations.”

There is an important distinction here: SEQL/HEQL estimate coefficient values through pointwise weights or a hypernetwork, without requiring a symbolic formula for those coefficients [1]. Our original phrase “seek compact coefficient formulas” did not describe this accurately. We have corrected it in Section 2, paragraph 4 (p. 3). Revisions and additions are marked in green in the revised manuscript.

Xu et al. study ut=∂x(D(x)ux+vu)u_t=\partial_x(D(x)u_x+vu) with a KLE-generated random diffusion field D(x)D(x), as well as a wave equation with a random stiffness field [3, Section 3.3]. Their genetic search generates derivative products within an integral-flux template. We therefore place Xu et al.'s study in “Expandable library × Time-invariant spatial fields.” To make these placements easier to follow, we have explained the existing experiment-based criterion in the Introduction (paragraph 3, pp. 1–2), Figure 1's caption (p. 2), Section 2's first paragraph (p. 3), and Appendix I (pp. 41–44). Xu et al. are now included, bringing the audit to 88 papers.

SEQL/HEQL's published PDE experiments demonstrate open-form discovery with coefficients given by analytic formulas [1]. We study open-form discovery with irregular spatial fields, including random, discontinuous and channelized patterns. HER-PDE combines a scientific Agent that proposes and revises complete equations with HEI, which fits their unknown fields and tests predictions across trajectories. It searches both the equation terms and how fields are placed and reused within differential operators.

Our framework recovers complete structures across seven PDE families and gives an interpretable model of measured bacterial growth–migration data. In the comparison you suggested, its term-recovery F1F_1 reaches 87.0%, versus 58.0% for SEQL and 53.9% for HEQL, even after choosing their term-removal thresholds using the test answers.

The GP-IDENT-style control provides a useful fixed-library comparison, but additional comparisons with closely related methods such as SEQL/HEQL, on mutually supported cases and common metrics, would better support the claimed advantage over prior approaches.

A:

Thank you for suggesting these comparisons. We followed up in three steps: (1) reproducing SEQL/HEQL on Zhang et al.'s data; (2) evaluating both methods on our spatial-field benchmark; and (3) testing HER-PDE on the authors' advection–diffusion system. Tables R1–R3 report the results, respectively.

1. Reproduction of SEQL and HEQL

We reproduced Zhang et al.'s noiseless 1D advection–diffusion experiment using their original code and data [1]. Following their protocol, we ran 80 trials per method and selected the model with the lowest training error. Both methods reached the published coefficient accuracy. Table R1 also reports exact term recovery over all 80 trials: a term is retained if its coefficient magnitude exceeds the stated cutoff anywhere in space. The 0.05 column is a separate sensitivity check; both lowest-error models recover the exact term set at that cutoff.

Table R1. Selected-model MSE averaged over three coefficient profiles, and exact term recovery over 80 trials.

MethodOur MSEPublished MSE [1, Table III]Exact terms, cutoff 0.01Exact terms, cutoff 0.05 (sensitivity)
HEQL1.24e-31.37e-30/8027/80
SEQL1.99e-32.30e-31/808/80

2. SEQL and HEQL on our spatial-field benchmark

We adapted the official implementations to our two-dimensional data, using the same noisy A/B observations and shared coordinate-dependent weights. Settings were selected on development data and frozen before the 105 runs per method. Table R2 compares which terms each method recovers, after expanding derivatives of products. Exact term recovery requires no missing or extra terms. Following Rao et al. [4], precision and recall measure the fractions of proposed terms that are correct and target terms that are found; F1F_1 is their harmonic mean. We compute all three metrics per run and then average over all 105 runs, including failures. HER-PDE recovers 78/105 exact term sets, including 70/105 complete structures, and achieves 87.0% F1F_1. SEQL/HEQL find many target terms but retain more extras.

Table R2. Common term-recovery comparison on our 35 environments, three runs each.

Method / exportExact termsComplete structurePrecisionRecallF1F_1
HEQL / full0/105N/A27.6%85.3%39.9%
HEQL / best F1F_1 cutoff*0/105N/A41.8%80.5%53.9%
SEQL / full0/105N/A8.8%95.2%15.6%
SEQL / best F1F_1 cutoff*0/105N/A44.5%88.6%58.0%
HER-PDE / final submission78/10570/10589.0%88.1%87.0%

*The cutoffs use test answers, favoring SEQL/HEQL; HER-PDE is evaluated as submitted. Optimizing exact matches instead gives 1/105 per baseline. N/A means the outputs do not encode the field-sharing relationships required for complete-structure recovery.

The zero exact-match counts in the full outputs arise despite high recall because extra terms remain: HEQL and SEQL retain an average of 9.9 and 38.5 extra terms per run. For example, one HEQL conservative-diffusion run finds all four correct terms but also includes a spatial source and three sine terms, so the equation fails the exact-match test. Pruning reduces extras but also removes true terms, leaving zero exact matches at the best-F1F_1 settings in Table R2.

3. HER-PDE on the authors' advection–diffusion data

We also tested HER-PDE on Zhang et al.'s system, using both initial preparations supplied by their generator and withholding the equation from the Agent. Native EQL training uses the first preparation. Table R3 reports all three HER-PDE runs: one recovers the exact conservative structure ut=∂x(C0(x)u)+c1uxxu_t=\partial_x(C_0(x)u)+c_1u_{xx}, preserving the field's position inside the derivative.

Table R3. Three HER-PDE runs on the authors' system; errors measure held-out state prediction.

RunFinal submitted structurePrediction error
r01Incorrect nonlinear structure3.83e-4
r02Exact conservative structure2.64e-8
r03All target terms, with extras4.18e-8

Adaptation, tuning, term assessment and the native-data protocols are documented in Appendix C.1 (Tables 14–16, pp. 27–28).

Your suggested comparisons complement the main benchmark, matched-HEI controls and measured-data study. SEQL/HEQL reach their published accuracy in our reproduction; the common-metric comparison supports HER-PDE's stronger term recovery on heterogeneous spatial-field data. Together, the results support the combination of scientific search and candidate-specific numerical testing.

We sincerely appreciate the time and care you have devoted to reviewing our work. Your feedback has helped us strengthen both the paper and its supporting evidence. If our clarifications and additional results have addressed your main concerns, we would be very grateful if you would consider revisiting your assessment and score. Should any concerns remain, we would sincerely welcome the opportunity to discuss them further.


[1] Zhang et al. Deep Learning and Symbolic Regression for Discovering Parametric Equations. TNNLS, 2024. https://arxiv.org/abs/2207.00529

[2] Lou et al. Data-driven discovery of governing differential equations across physical systems. 2026. https://arxiv.org/abs/2606.09638

[3] Xu et al. Deep-learning based discovery of partial differential equations in integral form from sparse and noisy data. JCP, 2021. https://doi.org/10.1016/j.jcp.2021.110592

[4] Rao et al. Discovering Nonlinear PDEs from Scarce Data with Physics-encoded Learning. ICLR, 2022. https://arxiv.org/abs/2201.12354

Reviewer_P7PLOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces HER-PDE, a framework consisting of a scientific agent and a numerical evaluator for PDE discovery. The paper claims to be the first framework that can jointly identify the governing PDE structure and estimate unknown time-invariant spatial coefficient fields from spatiotemporal observations.

Critical strengths
  1. While most PDE discovery methods focus on recovering equation structure under the assumption of constant coefficients, this work pushes the problem toward a more general setting with unknown, time-invariant spatial coefficient fields. This makes the task more realistic and more challenging.
Critical weaknesses

One of my major concerns is the identifiability between equation terms, spatial coefficient fields, and noise. In realistic PDE discovery settings, we do not know whether a mismatch comes from a missing term, an incorrectly estimated coefficient field, or simply noise in the observations and derivative estimates. The authors propose cross-trajectory validation to mitigate this issue, but it is not clear to me whether this is sufficient to reliably disentangle these factors. Overall, the paper presents a complete workflow, but the theoretical understanding of when the correct structure and coefficient fields are identifiable is still limited.

Questions for authors

Please see the weakness

4 comments

In realistic PDE discovery settings, we do not know whether a mismatch comes from a missing term, an incorrectly estimated coefficient field, or simply noise in the observations and derivative estimates. The authors propose cross-trajectory validation to mitigate this issue, but it is not clear to me whether this is sufficient to reliably disentangle these factors.

A:

Thank you for raising these questions about cross-trajectory validation and identifiability. We examine both below.

(To make it easier for you to see how cross-trajectory validation addresses your concern, we have added a plain-language summary at the end. If you prefer, you can read the summary first; results are provided below, with detailed settings in Appendix J (pp. 45–47). Additions to the revised manuscript are marked in green.)

To address your questions about why cross-trajectory validation helps and when the correct structure and coefficient fields can be identified, we would first like to share some intuition and examples from the literature.

Rudy et al. [1, Figure 3] give a concrete example. Their data come from the nonlinear KdV equation, ut+6uux+uxxx=0u_t+6uu_x+u_{xxx}=0. A single isolated pulse travels without changing shape: u(x,t)=f(x−vt)u(x,t)=f(x-vt). This same motion also satisfies the much simpler advection equation ut+vux=0u_t+vu_x=0, so PDE-FIND selects that simpler explanation. Changing the pulse amplitude changes its speed: one fixed advection coefficient vv can no longer explain both trajectories, whereas the same KdV equation can. The connection to our problem is that a candidate can fit one trajectory without identifying the underlying law. Requiring the same coefficients to explain both trajectories adds a cross-trajectory check that the first trajectory alone cannot provide.

For a real system with an unknown PDE, how can we distinguish missing terms, inaccurate fields and measurement noise? Prediction error alone cannot resolve this. In our view, equation discovery narrows the search to explicit, testable physical hypotheses. Song et al. [4, Box 1] describe this principle: “These hypotheses are then subject to validation by experiments, resulting in their either acceptance or rejection.” Confidence comes from reproducible predictions, physical consistency and independent observations; Reinbold et al. [2] check inferred fluid forcing against independent magnetic-field measurements.

Cross-trajectory validation tests whether one equation and the same fields explain different trajectories of the same system. A field that hides a structural error on one trajectory may fail to explain another. HER-PDE fits shared fields to A/B and tests predictions at held-out time points. We appreciate your question and address structural errors, field errors and noise in turn below.

1. Structural errors, field errors and noise

We use the five existing reaction–diffusion environments, ut=∇⋅(D∇u)−kuu_t=\nabla\cdot(D\nabla u)-ku, where D(x,y)D(x,y) controls diffusion and k(x,y)k(x,y) controls decay. A and B are different trajectories with the same true fields; A/A' denotes two independently noisy measurements of A. Field errors are relative L2L_2 errors against the known fields; lower is better. Tables R1–R4 are numbered within this response.

1.1 Structural errors: missing and extra terms

We fit the complete equation, remove its decay term, or add an unnecessary quadratic term, allowing the remaining fields to adjust. Table R1 reports the clean-data results for all five environments and the extra-term results for the four environments where both fitting groups completed all three repeats. Cross-trajectory fitting gives lower average diffusion-field error for all three candidates, although the missing-term gain is small. With the extra term, cross-trajectory fitting also recovers the decay field more accurately and reduces the extra term's size relative to the true equation's right-hand side from 87.6% to 5.0%. Single-trajectory fitting assigns a much larger contribution to that unnecessary term.

Table R1. Field recovery for complete, missing-term and extra-term equations.

Candidate and dataQuantitySingle-trajectory dataCross-trajectory: A/B
Complete, cleanDiffusion field errorA: 0.0505; B: 0.05200.0440
Missing −ku-ku, cleanDiffusion field errorA: 0.8750; B: 0.88560.8748
Extra bu2bu^2, 5% noiseDiffusion field errorA/A': 0.2860.105
Extra bu2bu^2, 5% noiseDecay field errorA/A': 0.6890.170
Extra bu2bu^2, 5% noiseExtra-term contribution / true dynamicsA/A': 87.6%5.0%

1.2 Field errors: recovery from inaccurate starting fields

We give every fitting group the same incorrect starting field, with a 20% perturbation, while keeping the equation correct and observations clean. Table R2 shows how much error remains after fitting. Cross-trajectory fitting recovers both fields more accurately on average than either single trajectory. Decay-field recovery improves in all five environments and diffusion-field recovery in four; the fifth diffusion fit stalls for all groups.

Table R2. Recovery from the same starting field error; averages over five environments.

Perturbed fieldInitial errorAfter A-only fitAfter B-only fitAfter A/B fit
Diffusion DD0.2030.0810.0830.076
Decay kk0.2100.0710.0780.066

1.3 Measurement noise: field recovery at equal observation counts

We compare two noisy measurements of the same trajectory with cross-trajectory observations, keeping the observation count equal. With the correct equation and 5% noise, cross-trajectory fitting reduces both field errors in every environment (Table R3). The comparison uses three noise realizations per environment. Thus, cross-trajectory information improves field recovery beyond the benefit of repeated measurements alone. These results also appear in Appendix C.2, Table 18 (p. 30).

Table R3. Field recovery from noisy observations; five environments, three noise realizations each.

Fitting dataDiffusion field errorDecay field error
A only0.1760.363
B only0.2050.463
A/A': two noisy measurements of A0.1490.285
A/B: cross-trajectory0.0900.129

the paper presents a complete workflow, but the theoretical understanding of when the correct structure and coefficient fields are identifiable is still limited.

A:

Thank you for recognizing our complete workflow and for highlighting this important theoretical question. We address your concern below with a local identifiability analysis and numerical experiments that test how cross-trajectory observations help distinguish coefficient fields.

2. Local coefficient identifiability

An incorrect field may predict A almost as well as the correct field, yet predict B poorly. Requiring agreement with both makes that field error easier to detect. He, Zhao and Zhong [3, Theorem 3.4] give a sufficient condition for coefficient recovery in a prescribed family of linear differential operators: solution and derivative observations must supply sufficiently independent equations. This concerns the information in the observations, not a guarantee that two trajectories identify every PDE.

For HEI, a field change hidden from joint fitting must be hidden on both trajectories: ker⁡JAB=ker⁡JA∩ker⁡JB\ker J_{AB}=\ker J_A\cap\ker J_B, where each Jacobian measures how field changes alter predictions. If the Jacobian of the continuously differentiable prediction map has full column rank at the fitted point, the field parameters are locally identifiable for the fixed equation; Appendix J provides the derivation.

We test this explanation by comparing duplicated clean single-trajectory observations with cross-trajectory observations, at equal observation counts and the same fitted fields. Table R4 measures how strongly predictions respond to the hardest-to-detect small field change. Cross-trajectory fitting increases that response in every environment. Consequently, locally, less field error can remain hidden at the same prediction accuracy. This concerns detecting field errors, not amplifying measurement noise.

Table R4. Increase in minimum sensitivity to field errors under cross-trajectory observations.

EnvironmentCross-trajectory / duplicated single-trajectory sensitivity
e011.62
e022.21
e032.33
e045.66
e052.01

3. End-to-end equation discovery

We additionally ran the full Agent workflow 105 times using only A, retaining the controller and search-budget ceilings. Exact recovery is 48/105 with A only, compared with 62/105 with separate A/B fields and 70/105 with shared-field HER-PDE in the existing benchmark. Every run is counted by its final submission.

4. Summary of cross-trajectory benefits

  • More accurate fields: with the correct equation and the same inaccurate starting fields, cross-trajectory fitting recovers both fields more accurately on average than either single trajectory.
  • Less reliance on extra terms: cross-trajectory fitting reduces the unnecessary term's contribution compared with single-trajectory fitting.
  • Better recovery under noise: at 5% noise and the same observation count, cross-trajectory fitting recovers both fields more accurately in all five environments than repeated measurements of one trajectory. The gain goes beyond collecting more measurements of the same trajectory.
  • Field errors are harder to hide: in our fixed-equation analysis, cross-trajectory observations reveal the hardest-to-detect small field changes more clearly than repeated observations of one trajectory. Thus, less local field error can remain hidden at the same prediction accuracy.
  • Better equation recovery: the full cross-trajectory workflow recovers 70/105 equations, compared with 48/105 using one trajectory.

Thank you again for your positive assessment and thoughtful suggestions. If our clarifications and additional evidence have further strengthened your confidence in the work, we would be very grateful if you would consider reflecting this in your final score. We would be very happy to discuss any remaining questions.

[1] Rudy et al. Data-driven discovery of partial differential equations. Science Advances, 2017, Figure 3.

[2] Reinbold et al. Robust learning from noisy, incomplete, high-dimensional experimental data via physically constrained symbolic regression. Nature Communications, 2021.

[3] He, Zhao and Zhong. How Much Can One Learn a Partial Differential Equation from Its Solution?. Foundations of Computational Mathematics, 2024. Theorem numbering follows arXiv v2.

[4] Song et al. Towards data-driven discovery of governing equations in geosciences. Communications Earth & Environment 5, 589, 2024. Box 1, p. 2.

Thank you for the detailed reply, I will keep my positive evaluation.

Comments on the paper

1 comment
AuthorsOpenReview ↗

What HER-PDE contributes and why it matters

Dear Area Chair,

Thank you for handling our paper. As the discussion draws to a close, we would like to briefly summarize what HER-PDE contributes and why we think it matters for discovering physical laws from data.

When a material varies from place to place, can observations tell us both how its diffusivity varies and where it belongs in the equation? The placement matters. Putting diffusivity inside the divergence, for example, introduces a term involving its spatial gradient. Moving it outside drops that term. HER-PDE searches over both possibilities and can infer when several terms share the same unknown field. In our survey, we found no earlier study that demonstrated this combination for open-form equations with irregular, time-invariant fields.

To choose between competing equations, the Agent uses a simple test. For each candidate, it fits the same fields across different trajectories, then uses the fitted fields and the resulting predictions to refine its next proposal. A check we added during review shows why this helps. An unnecessary reaction term made a large contribution when we fit two noisy measurements of one trajectory. With just as many observations from two distinct trajectories, its contribution fell sharply (Appendix J.1, Table 29).

On our seven-family benchmark, we could check the recovered equations and fields against the known answers and test predictions on held-out observations. What we found especially interesting was what we could check in the bacterial growth experiment. The Agent proposed an equation in which per-capita growth slows as density rises. We had given it the measurements from each well without telling it which wells contained the antibiotic. Yet every time we refit the model, the estimated growth field peaked in the drug-free wells (Figure 3).

Best regards,
The authors of Paper 13615

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.