acceptodds
Under review as a conference paper at ICLR 2027

MINT: Shared PDE Identification with Neural Fields and Genetic Programming

Abstract

Identifying a partial differential equation (PDE) from sparse measurements is difficult because different equations can fit the same observed solution. We propose MINT (Multi-trajectory Identification with Neural fields and Trees), which fits independent neural fields and uses genetic programming (GP) to construct a single equation shared across multiple trajectories. MINT searches for terms and coefficients simultaneously and retains the selected structure while refining the fields and coefficients. We also develop MINT-M, which uses additional family-specific motifs to guide initialization. Across six benchmarks with 30 runs each, MINT recovers the terms of 2D and 3D advection in 86.7% and 43.3% of noise-free runs, respectively, whereas two GP-based baselines recover neither. In the same setting, MINT-M recovers Burgers and Fisher–KPP terms in 56.7% and 63.3% of runs, respectively, compared with no recoveries for MINT. Both methods also recover 2D advection more often than the GP baselines at 1% and 5% noise. With 250 observations per trajectory, both improve field accuracy over the common data-only reconstruction on 2D advection, Burgers, and Fisher at both noise levels. Missing-library and noise tests further show that accurate fields and low residuals alone do not establish recovery of the governing law.

Primary area:
neurosymbolic & hybrid AI systems (physics-informed, logic & formal reasoning, etc.)
Keywords:
Partial differential equation discovery, symbolic regression, genetic programming, neural fields, multi-trajectory identification, physics-informed learning

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_2NsrOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes MINT to discover shared PDEs from multiple sparse observation trajectories generated by the same partial differential equation. The method fits an independent neural field for each trajectory and employs genetic programming to search for a shared expression tree. The pipeline consists of a structure-search stage with frozen neural fields and a continuous parameter optimization stage after the tree structure is fixed. The authors further propose MINT-M, which incorporates PDE-family motifs into the initialization of genetic programming. Experiments are conducted on six analytical benchmarks, reporting equation term recovery rates, field reconstruction errors, and performance under noise, varying observation counts, and missing candidate library settings.

Critical strengths
  1. Addresses the ambiguity of single-trajectory PDE discovery by leveraging multiple trajectories governed by the same underlying equation.
  2. MINT decouples trajectory-specific field reconstruction from shared equation structure identification, providing a clear methodological rationale.
  3. Evaluates six PDEs under varying noise levels, observation quantities, and candidate-term omissions, using both term recovery rates and field reconstruction errors.
  4. MINT-M provides preliminary evidence that family-level structural motifs can guide genetic search toward identifying nonlinear terms, offering a promising direction for improving symbolic PDE discovery.
Critical weaknesses
  1. Insufficient validation of the core contribution. The paper lacks controlled experiments varying trajectory counts under fixed observation and computational budgets, as well as ablations to isolate the gains of multi-trajectory consistency from other components.

  2. Limited generalizability across PDEs. MINT performs unevenly across PDE benchmarks, with failures on several equations. Nonlinear evaluations are limited to traveling-wave trajectories, leaving its effectiveness on general initial-value problems unverified.

  3. Unverified robustness of prior-guided MINT-M. MINT-M’s improvements rely on PDE-family motifs, but experiments do not examine incorrect or incomplete priors, leaving its robustness to prior mismatch unclear.

  4. Insufficient analysis of equation recovery failures. Low field errors do not guarantee correct equation recovery. The paper lacks systematic analysis of spurious and missing terms, as well as validation on unseen trajectories.

Questions for authors
  1. Under fixed total observations and identical search budget, how does MINT’s recovery rate change with the number of trajectories MM? Compared with single-trajectory GP, does the benefit brought by multiple trajectories still hold?
  2. Could the authors ablate shared terminal screening, local residual, CVaR and Stage II refinement separately, and report equation term recovery rates rather than only field errors, to clarify the contribution of each mechanism to the core results?
  3. How are the motifs of MINT-M constructed, and what specific prior information do they contain? How does the recovery rate change when the motif is incorrect, partially correct, or mismatched with the target equation?
  4. Could the authors provide representative expressions from successful and failed runs, distinguishing missing terms, spurious terms, redundant terms and coefficient errors? Can the residual or prediction performance of the discovered equations be evaluated on unseen trajectories?
Reviewer_oCtLOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
3
Soundness
1
Read review
Paper summary

This paper proposes MINT, a framework that jointly uses multiple trajectories to recover the terms and coefficients of a PDE. The main idea is to fit an independent neural field for each trajectory, use these neural fields to estimate unobserved quantities such as (u_x) and (u_{xx}), and then apply genetic programming to search for a single governing equation that is shared across all trajectories.

Critical strengths

PDE discovery is essentially an inverse problem, where different equations may fit the same observed trajectory, especially when the observations are sparse or noisy. To reduce this ambiguity, the authors propose a straightforward idea: instead of searching for an equation that explains a single trajectory, they search for one governing equation that is shared across multiple trajectories generated by the same PDE.

Critical weaknesses

Although the motivation is straightforward, I do not see how the current method design and experiments cleanly demonstrate the benefit of the proposed multi-trajectory idea. My main concerns are as follows.

First, MINT contains several additional components beyond the shared multi-trajectory constraint, as shown in Figure 1. I do not fully understand the rationale for introducing all of these components together. It becomes difficult to tell which component is actually responsible for the observed gains.

Second, the baseline setup is somewhat mixed. The most direct comparison should be against a matched single-trajectory GP baseline using the same grammar, derivative estimation, search budget, screening, and refinement procedure, differing only in whether the equation is shared across multiple trajectories. I do not see such a comparison. In addition, the results are quite mixed: compared with Plain GP, MINT performs substantially worse on H2, and compared with library-based methods, it underperforms on several PDEs. Maybe a detailed analysis is needed to identify the failing senario of the proposed method.

Third, I think the noisy setting should play a more central role in demonstrating the method. The ambiguity of PDE recovery is arguably more severe when the observations are noisy. However, Table 3 mainly reports field-accuracy comparisons, while the core claim of the paper is about improved equation recovery.

Questions for authors

Please refer to the weakness

Reviewer_7WSCOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

MINT combines independently fitted neural fields with genetic programming to identify one PDE shared across several trajectories. It first searches over expressions using frozen field derivatives, then fixes the selected structure and jointly refines fields and coefficients. MINT-M additionally seeds the search with family-specific motifs. Six analytic benchmarks examine term recovery and field reconstruction under noise, sparse observations, and incomplete candidate libraries.

Critical strengths
  1. The advection results demonstrate a concrete improvement within the tested symbolic-search setting. MINT recovers the correct terms in 26/30 two-dimensional and 13/30 three-dimensional noise-free runs, versus zero for both GP baselines. The two-dimensional advantage also persists in the reported noisy settings.
  2. The evaluation clearly distinguishes equation recovery from field reconstruction and residual minimization. Reporting 30 runs, unsuccessful cases, and the additional prior information supplied to MINT-M makes the empirical findings substantially easier to interpret.
Critical weaknesses
  1. The central contribution is not isolated experimentally. Comparisons with the shared frozen reconstruction assess refinement's benefit for field accuracy, but there is no controlled single-versus-multiple-trajectory comparison or ablation separating terminal screening, local residual moments, and upper-tail loss. Consequently, the observed gains do not establish which proposed mechanism improves term recovery or whether additional trajectories resolve the relevant ambiguities.
  2. Evidence for flexible nonlinear equation discovery is weak. Base MINT recovers neither Burgers nor Fisher in any noise-free run, while MINT-M may start from the correct structure. Moreover, Appendix E replaces DLGA-PDE's constructive genetic search with fixed-library mask selection, removing precisely the capability relevant to the missing-library comparison. Together with evaluation restricted to analytic solution families and nonlinear traveling waves, this leaves the advantage over existing constructive discovery methods insufficiently established.
Questions for authors
  1. How does recovery change from one to several trajectories when the total observation budget and search budget are controlled? Which term-recovery gains remain when screening and the local and upper-tail fitness terms are ablated?
  2. Can the principal comparisons include a DLGA-PDE implementation retaining derivative-gene and product construction? For MINT-M, please list the exact seeded motifs and report whether successful structures were already present at initialization.
  3. For runs counted as exact support recovery, what are the coefficient errors? Can the recovered equations predict a held-out trajectory with a different initial profile, without refitting the equation? The present coordinate-level tests on the fitted solution families do not establish this form of governing-law generalization.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.