acceptodds
Under review as a conference paper at ICLR 2027

Symbolic by Design: Zero-Prior Neural ODE Discovery via Decoupled Affine KANs

Abstract

Neural scientific discovery aims to uncover governing laws from observational data, requiring models that capture underlying physical mechanisms rather than merely interpolate trajectories. For ordinary differential equations, this objective exposes a fundamental trade-off. Flexible neural function approximators accurately fit noisy dynamics but rarely recover compact symbolic equations, whereas sparse symbolic regression methods produce interpretable formulas at the cost of strong prior assumptions. We propose DA-KAN, a decoupled affine Kolmogorov-Arnold network for zero-prior differential equation discovery. DA-KAN makes symbolifiability an architectural property by representing each ODE coefficient with an explicit global affine path, which carries the recoverable symbolic law, together with an annealed local RBF residual path that absorbs optimization mismatch and noise during training. DA-KAN combines this operator with Wronskian multi-trajectory excitation, cascade order validation, and a frozen noise-agnostic adaptive extractor based on residual-checked snapping and affine projection. Extensive experiments on representative affine-coefficient ODE benchmarks demonstrate that DA-KAN achieves robust exact symbolic recovery under noisy observations without requiring the true differential order, consistently outperforming representative neural and symbolic baselines. These results indicate that embedding symbolic structure into the network architecture substantially improves exact equation recovery for zero-prior affine-coefficient ODE discovery.

Primary area:
neurosymbolic & hybrid AI systems (physics-informed, logic & formal reasoning, etc.)
Keywords:
Neural scientific discovery, Differential equation discovery, nterpretable machine learning, Physics-inspired neural networks, Zero-prior discovery

Reviews

Among:
Shape:
1234this paper, Bottom 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.00 over 3 reviews, above 0% of the others.
Reviewer_CDtFOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
3
Soundness
1
Read review
Paper summary

DA-KAN combines affine-plus-RBF coefficient models, order selection, and symbolic extraction to recover scalar linear ODEs with affine time-dependent coefficients.

Critical strengths

The approach appears to recover the equations for affine models using multiple trajectories with exact recovery on 2% noise.

Critical weaknesses

-- The models are affine so very constrained in applicability

-- TLS appears to outperform their approach and is significantly faster. It is not clear why the KAN approach is needed.

-- Authors appear to be unaware of other equation recovery methods. I would not compare against Sindy etc.

Questions for authors

-- Why is the KAN component needed if TLS is performing well?

Reviewer_dqK2OpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper focuses on the problem of identifying dynamical equations from measured trajectory data. The problem is restricted to the case of y(n)+∑i=0n−1(ait+bi)y(i)y^{(n)}+\sum_{i=0}^{n-1} (a_it+b_i)y^{(i)} - i.e., a scalar linear nonautonomous homogeneous differential equation with affine-in-time coefficients. The proposed method, DA-KAN, heavily capitalizes on this knowledge by fitting models of the form y(n)+∑i=0n−1Φi(t)y(i)y^{(n)}+\sum_{i=0}^{n-1} \Phi_i(t)y^{(i)}, where each Φi(t)\Phi_i(t) is a sum of learnable affine function with an RBF-based KAN to absorb model errors and noise effects. The discovered model accuracy is evaluated by knocking out the KAN component, and unknown derivative order is resolved by repeating model training with increasing derivative order until model accuracy satisfies some prespecified threshold criteria. Experiments for systems with n=2,3n=2,3 shows that the proposed method can accurately recover equation coefficients even in the presence of noise.

Critical strengths

The general problem of dynamics discovery from noisy trajectory data is highly nontrivial, and developing dedicated discovery algorithms for useful classes of equations is worth investigating.

Critical weaknesses

I believe the problem class considered in the paper is too simple, the data requirements and the algorithm complexity are too strong compared to the problem difficulty, and that the experimental evaluations are insufficient even for this restricted case. I layout my concerns below.

Proposed equation family is too simple & proposed algorithm is too prior-information dependent.

The family of ODEs considered in this paper is too simple. Nonetheless, if the equation form is not known to the algorithm, this paper would have had more significance. However, the proposed DA-KAN method literally fits the exact ODE family to data, with the slight addition of a learnable KAN modulation on the time dependent weights. This is relying too much on the specifics of the problem and the authors do not provide any clue as to how this approach could be extended to more difficult problems.

Another problem with this known structure of the ground truth ODE family is that it is a linear regression problem for the affine function coefficients. For y(2)+(a0t+b0)y+(a1t+b1)y(1)=0y^{(2)}+(a_0t+b_0)y+(a_1t+b_1)y^{(1)}=0, y(2)y^{(2)} is a linear function of a0,b0,a1,b1a_0,b_0,a_1,b_1 - meaning that with multiple trajectories, time samples, and derivative trajectories, this is just an error-in-variables linear regression problem, with no need to train anything.

The data requirements are too strong

The proposed method requires both multiple trajectories and higher order derivative trajectories. The latter is especially problematic because in dynamics identification, the derivative estimation from noisy trajectories is the highly nontrivial part and is something algorithms like weak SINDy try to tackle. This study just provides this information as a slightly noised version of the ground truth (as opposed to derivative noise scaling as 1/Δt1/\Delta t per derivative order), making the problem trivial.

Experimental evaluation is insufficient

The only part of the model that is not specified and must be inferred from data is the derivative order nn, which is found through iterative fitting. However, experiments are run on n=2,3n=2,3 only, meaning that this iterative fitting is done only once. Higher derivative orders must be experimented with the properly vet this procedure.

Overall, the paper contains fundamental issues with its problem setup that does not seem fixable with a single revision round.

Questions for authors

Q1

Can the authors provide a way to run this algorithm, supplied noisy trajectories only (and no derivative trajectories; i.e., is the method robust to derivative estimation noise)?

Q2

Does KAN play a role in this procedure? One could just fit the model with out the KAN component, assess the fit quality, and raise the derivative order if the fit is not good. Why use KANs? Can the authors provide ablations experiments on this front?

Q3

Considering the authors' problem setup amounts to linear regression, I am confused as to the poor performance of the SINDy methods. Do the baseline methods have the same amount of prior information as the authors' algorithm - i.e., knowledge of the form of the ground truth ODE?

Reviewer_QiW4OpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes DA-KAN for scalar ODE discovery with affine coefficients. It combines an affine-plus-RBF model, multi-trajectory data, order selection, and residual-checked coefficient snapping. The final extractor compares learned coefficients with an independently computed total least squares estimate.

Critical strengths
  1. The method has a clear output form and explicitly states its restricted scope and derivative-noise setting.
  2. The evaluation shares noisy data across methods and reports five-seed results, failure cases, and a useful comparison.
Critical weaknesses
  1. The results do not clearly show what the neural component adds beyond the simpler TLS-based approach. This makes the main contribution difficult to assess.
  2. Some experimental results appear to differ across the figures and tables. Clarifying how these results relate would help readers judge the reported performance.
Questions for authors

What benefit does the neural component provide over TLS-only? Do all baselines receive the same coefficient-snapping procedure?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.