betaFeedback

WorksComputer Science

SPIKE: Sparse Koopman Regularization for Physics-Informed Neural Networks

Abstract

Physics-Informed Neural Networks (PINNs) provide a mesh-free approach for solving differential equations by embedding physical constraints into neural network training. However, PINNs tend to overfit within the training domain, leading to poor generalization when extrapolating beyond trained spatiotemporal regions. This work presents SPIKE (Sparse Physics-Informed Koopman-Enhanced), a framework that regularizes PINNs with continuous-time Koopman operators to learn parsimonious dynamics representations. By enforcing linear dynamics dz/dt=Azdz/dt = Az in a learned observable space, both PIKE (without explicit sparsity) and SPIKE (with L1 regularization on AA) learn sparse generator matrices, embodying the parsimony principle that complex dynamics admit low-dimensional structure. Experiments across parabolic, hyperbolic, dispersive, and stiff PDEs, including fluid dynamics (Navier-Stokes) and chaotic ODEs (Lorenz), demonstrate consistent improvements in temporal extrapolation, spatial generalization, and long-term prediction accuracy. The continuous-time formulation with matrix exponential integration provides unconditional stability for stiff systems while avoiding diagonal dominance issues inherent in discrete-time Koopman operators.

Introduction

Solving partial differential equations with reliable out-of-distribution (OOD) generalization remains a fundamental challenge in scientific computing [1]. Physics-Informed Neural Networks (PINNs), introduced by Raissi et al. [2], provide a mesh-free approach by embedding differential equation residuals directly into neural network training. Given a partial differential equation N[u]=f\mathcal{N}[u] = f where N\mathcal{N} is a nonlinear differential operator, PINNs minimize a physics loss:

Lphysics=1Nr∑i=1Nr∣N[uθ](xi,ti)−f(xi,ti)∣2(1)\mathcal{L}_{\mathrm{physics}} = \frac{1}{N_r}\sum_{i=1}^{N_r}\left|\mathcal{N}[u_\theta](x_i,t_i)-f(x_i,t_i)\right|^2 \tag*{(1)}

alongside initial and boundary condition losses. This formulation enables mesh-free solutions with minimal data requirements [2, 3]. However, PINNs face two key limitations: (1) the neural network representation uθu_\theta remains opaque: the learned weights do not directly reveal the underlying dynamics, limiting interpretability; and (2) without structural regularization, PINNs tend to overfit within the training domain, leading to rapid error growth when extrapolating beyond trained time horizons [1].

Koopman operator theory [4] provides a complementary framework where nonlinear dynamical systems admit linear representations in infinite-dimensional observable spaces. For an autonomous system x˙=f(x)\dot{x}=f(x), the Koopman operator family Kt\mathcal{K}^t acts on observables g:X→Cg:\mathcal{X}\to\mathbb{C} via Ktg(x)=g(Ft(x))\mathcal{K}^t g(x)=g(F^t(x)), where FtF^t denotes the flow map. The infinitesimal generator L\mathcal{L}, also known as the Lie operator, satisfies the fundamental relation [5]:

Lg=∇g⋅f(2)\mathcal{L}g = \nabla g \cdot f \tag*{(2)}

In finite-dimensional approximations, this yields linear dynamics dz/dt=Azd z/dt = A z where z=g(x)z = g(x) represents lifted observables and A∈RM×MA \in\mathbb{R}^{M \times M} approximates the generator.

The present work proposes a framework where the PINN remains the base model for solving PDEs via automatic differentiation, with the Koopman component serving as an auxiliary regularizer that

Preprint. Work in progress. promotes sparse, interpretable structure in the learned dynamics. This inverts the typical paradigm: rather than augmenting Koopman methods with physics constraints, the approach enhances PINNs with Koopman regularization.

Nomenclature. PIKE (Physics-Informed Koopman-Enhanced) denotes the base framework with Koopman regularization. SPIKE (Sparse PIKE) adds explicit L1 sparsity on the generator matrix AA. Both learn sparse structure; SPIKE enforces it explicitly.

Contributions. This work makes the following contributions:

  1. Koopman regularization for generalization: The sparse Koopman constraint acts as an implicit regularizer, improving temporal extrapolation by 2–184× and preventing catastrophic failure outside training regions. For open-domain systems, this enables physically meaningful spatial generalization.

  1. Parsimonious dynamics representation: L1 sparsity reduces non-zero generator entries by up to 5.7×, yielding parsimonious representations embodying the parsimony principle [6]. For inverse problems, sparse structure enables hypothesis generation.

  1. Library-latent decomposition: A dual-component observable embedding combining explicit polynomial terms with learned MLP features (up to 0.99 correlation with uxxu_{xx}), enabling both coefficient recovery and interpretability.

  1. Continuous-time Koopman formulation: Direct learning of the generator AA via dz/dt=Azd z/dt = Az, avoiding diagonal dominance issues in discrete-time formulations.

Related Work

Physics-Informed Neural Networks. The PINN framework [2] approximates solutions to differential equations using neural networks trained with physics-based loss functions. For a PDE ut+N[u]=0u_t + \mathcal{N}[u] = 0, a neural network uθ(x,t)u_\theta(x,t) minimizes L=λrLphysics+λicLIC+λbcLBC\mathcal{L} = \lambda_r\mathcal{L}_{\mathrm{physics}} + \lambda_{ic}\mathcal{L}_{IC} + \lambda_{bc}\mathcal{L}_{BC} where automatic differentiation computes required derivatives. PINNs have been applied to Navier-Stokes equations [2], turbulence modeling [3], and inverse problems [7]. However, PINNs can fail to train due to spectral bias and gradient pathologies [1], and tend to overfit within the training domain. Extensions address training difficulties through gradient balancing [3], loss modifications, and architectural innovations; Hanna et al. [8] proposed variance-based regularization, while AC-PKAN [9] combines Chebyshev-based Kolmogorov-Arnold Networks with attention mechanisms. Such approaches improve training stability or expressiveness but do not address OOD generalization—PIKE/SPIKE is orthogonal and could be combined with these architectures. Neural operators (FNO [10], DeepONet [11]) learn function-space mappings but also exhibit OOD limitations [12, 13].

Koopman Operator Theory. The Koopman operator provides a linear representation of nonlinear dynamics in an infinite-dimensional function space [5]. Dynamic Mode Decomposition (DMD) [14] and Extended DMD [15] approximate this operator from data. Deep Koopman approaches include autoencoders for invariant subspaces [16, 17] and Physics-Informed Koopman Networks (PIKN) [18], which incorporate the Lie operator constraint Lg=∇g⋅f\mathcal{L}_g = \nabla g \cdot f. Unlike PIKN, which requires explicit knowledge of the dynamics f(x)f(x), PIKE operates directly through the PDE residual. The Koopman Regularization framework of Cohen [19] proves that for an NN-dimensional system, exactly NN functionally independent Koopman eigenfunctions suffice, justifying sparse AA matrices. However, while Koopman Regularization targets system identification, PIKE/SPIKE targets OOD generalization for PDE solving.

Sparse Dynamics Discovery. SINDy [20] pioneered sparse regression using finite-difference derivatives. Extensions include PDE-FIND [21], neural hybrids [22], and mesh-free variants [23]. SPIKE addresses their noise sensitivity via autograd derivatives from the trained PINN.

Key Distinctions. Table 4 summarizes methodological differences. PIKE/SPIKE: (1) operates on PDE residuals, not explicit f(x)f(x); (2) targets OOD generalization, not training stability [8] or system identification; (3) employs continuous-time dz/dt=Azd z/dt = Az with multiple integrators (Euler/RK4/expm); (4) enables post-hoc sparse discovery without retraining.

MethodRequires f(x)f(x)OOD TargetPDE SolvingKoopman
PINN [2]NoNoYesNo
Variance Loss [8]NoNoYesNo
SINDy [20]Yes (samples)NoNoNo
Deep Koopman [16]Yes (trajectories)NoNoYes
PIKN [18]Yes (explicit)NoNoYes
Koopman Reg. [19]Yes (samples)YesNoYes
PIKE/SPIKENoYesYesYes

Table 4. Comparison of PIKE/SPIKE with Related Methods

Theoretical Foundation

This section establishes the theoretical basis connecting the Koopman infinitesimal generator to physics-informed learning. The key insight, following Liu et al. [18], is that the Lie operator relation provides a principled mechanism for enforcing dynamical consistency.

Koopman Generator and Physics Constraints

Definition 1 (Koopman Operator Family). For an autonomous dynamical system dx(t)/dt=f(x(t))d x(t)/dt = f(x(t)) with x∈X⊆Rnx \in\mathcal{X} \subseteq\mathbb{R}^{n}, let Ft:X→XF^{t} : \mathcal{X} \to\mathcal{X} denote the time-tt flow map satisfying x(t0+t)=Ft(x(t0))x(t_{0} + t) = F^{t}(x(t_{0})). The Koopman operator family Kt:G(X)→G(X)\mathcal{K}^{t} : \mathcal{G}(\mathcal{X}) \to\mathcal{G}(\mathcal{X}) acts on observables g:X→Cg : \mathcal{X} \to\mathbb{C} as:

Ktg(x)=g(Ft(x))(3)\mathcal{K}^{t}g(x) = g(F^{t}(x)) \tag*{(3)}

Proposition 1 (Lie Operator Consistency). For an autonomous system x˙=f(x)\dot{x} = f(x) and differentiable observable g:X→Cg : \mathcal{X} \to\mathbb{C}, the infinitesimal generator L\mathcal{L} (Lie operator) satisfies:

Lg=∇g⋅f(4)\mathcal{L}g = \nabla g \cdot f \tag*{(4)}

Proof in Appendix I.

Proposition 2 (Finite-Dimensional Approximation). Let g=[g1,…,gM]Tg = [g_{1}, \ldots, g_{M}]^{T} be a vector of observable functions spanning a finite-dimensional subspace, and let A∈RM×MA \in\mathbb{R}^{M \times M} approximate the generator restricted to this subspace. The physics-informed Koopman loss:

LLie=∥Ag(x)−∇g(x)⋅f(x)∥2(5)\mathcal{L}_{\mathrm{Lie}} = \lVert A g(x) - \nabla g(x) \cdot f(x) \rVert^{2} \tag*{(5)}

enforces consistency between the linear dynamics z˙=Az\dot{z} = Az (where z=g(x)z = g(x)) and the true nonlinear system.

Proof in Appendix I.

Remark 1 (Approximation Error Bounds). Proposition 2 establishes existence of finite-dimensional approximations; general error bounds for neural network observables remain open. This work focuses on empirical validation following standard practice in physics-informed learning [2].

Proposition 3 (Sparsity and Polynomial Representation). Under L1 regularization λs∥A∥1\lambda_s \lVert A \rVert_1, the minimizer of the combined loss tends toward sparse solutions. For polynomial observables g=[1,u,u2,uv,…]Tg = [1, u, u^2, uv, \ldots]^{T}, non-zero entries AijA_{ij} correspond to active polynomial terms, providing an interpretable representation of the dynamics.

Proof in Appendix I.

Remark 2 (Lie Loss as Regularizer). Equation 5 corresponds to the Lie operator consistency constraint used in physics-informed Koopman methods [18]. The key distinction is that SPIKE enforces this constraint as a regularizer within a PINN framework, rather than as the primary training objective.

Proposition 4 (Out-of-Distribution Generalization Bound). Let uθu_{\theta} be a PINN solution trained on domain Ω×[0,T]\Omega\times[0,T] with Koopman consistency error ϵK=E[∥Az−z˙∥12]1/2\epsilon_{\mathrm{K}} = \mathbb{E}[\lVert Az - \dot{z} \rVert_{1}^{2}]^{1/2}. Let z=g(uθ)z = g(u_{\theta}) be the observable embedding. Under the following assumptions:

*(i) Spectral bound: all eigenvalues of AA satisfy Re⁡(λi)≤ρ0\operatorname{Re}(\lambda_i) \le\rho_0

*(ii) Lipschitz decoder: ∥g−1(z1)−g−1(z2)∥≤Lg∥z1−z2∥\lVert g^{-1}(z_1) - g^{-1}(z_2) \rVert\le L_g \lVert z_1 - z_2 \rVert

the extrapolation error for t∈[T,T+δ]t \in[T, T + \delta] satisfies:

∥uθ(x,t)−u∗(x,t)∥L2(Ω)≤Lg⋅ϵKρ0(eρ0δ−1)+O(ϵtrain)(6)\lVert u_{\theta}(x,t) - u^{*}(x,t) \rVert_{L^{2}(\Omega)} \le L_g \cdot\frac{\epsilon_{\mathrm{K}}}{\rho_0}\left(e^{\rho_0\delta} - 1\right) + O(\epsilon_{\mathrm{train}}) \tag*{(6)}

where ϵtrain\epsilon_{\mathrm{train}} is the training domain error and u∗u^{*} is the true solution. (Proof in Appendix I; follows from Gronwall’s inequality.)

Remark 3 (Interpretation of Bound). Equation 6 shows: (1) small Koopman loss ϵK\epsilon_K directly reduces OOD error; (2) bounded spectral radius ρ0\rho_0 controls exponential growth; (3) linear error growth initially since (eρ0δ−1)/ρ0≈δ(e^{\rho_0\delta}-1)/\rho_0 \approx\delta for small δ\delta. This bound extends to spatial extrapolation under Lipschitz continuity assumptions (Corollary 1, Appendix § 17). The Lipschitz decoder assumption (ii) holds for neural networks with bounded weights and Lipschitz activations; for the 4-layer MLP with tanh activation used here, Lg≤∏l∥Wl∥L_g \le\prod_l \lVert W_l\rVert. Empirical estimation of LgL_g remains an open direction.

Physics-Informed Koopman-Enhanced Neural Networks

Architecture Overview

The proposed architecture augments a standard PINN with a Koopman regularization branch. The neural network encoder maps inputs (x,t)(x,t) to the solution field u(x,t)u(x,t), while a parallel embedding layer lifts uu to an observable space where linear dynamics are enforced. Figure 1 illustrates the overall framework.

PIKE/SPIKE architecture

Figure 1. PIKE/SPIKE architecture. A neural network maps coordinates (x,y,t)(x,y,t) to solution uu, which is lifted to observables g(u)=ψpoly(u)⊕ψlatent(u)g(u)=\psi_{\mathrm{poly}}(u)\oplus\psi_{\mathrm{latent}}(u). The Koopman operator AA governs linear dynamics dz/dt=Az\mathrm{d}z/\mathrm{d}t=Az. Three losses are minimized: physics residual Lphysics\mathcal{L}_{\mathrm{physics}} via automatic differentiation, Koopman consistency Lkoopman\mathcal{L}_{\mathrm{koopman}}, and L1 sparsity Lsparse=∥A∥1\mathcal{L}_{\mathrm{sparse}}=\lVert A\rVert_1 (SPIKE only).

Definition 2 (Augmented Embedding). The observable embedding g:Rn→RMg:\mathbb{R}^n\to\mathbb{R}^M decomposes as g(u)=[glib(u),gmlp(u)]Tg(u)=[g_{\mathrm{lib}}(u),g_{\mathrm{mlp}}(u)]^{T} where:

glib(u)=Wlib⋅ψd(u),gmlp(u)=σ(WL⋯σ(W1u))(7)g_{\mathrm{lib}}(u)=W_{\mathrm{lib}}\cdot\psi_d(u),\qquad g_{\mathrm{mlp}}(u)=\sigma(W_L\cdots\sigma(W_1u)) \tag*{(7)}

with ψd(u)=[1,u1,…,un,u12,u1u2,…]T\psi_d(u)=[1,u_1,\ldots,u_n,u_1^2,u_1u_2,\ldots]^T containing all monomials up to degree dd, Wlib∈RMlib×(n+dd)W_{\mathrm{lib}}\in\mathbb{R}^{M_{\mathrm{lib}}\times\binom{n+d}{d}} a learnable projection, and {Wi}\{W_i\} parameterizing an LL-layer MLP with activation σ\sigma.

The embedding comprises: (1) a polynomial library [1,u,u2,…][1,u,u^2,\ldots] mapping to symbolic expressions, and (2) learned MLP features providing additional capacity. Polynomial observables follow from Koopman theory: polynomial systems admit finite-dimensional invariant subspaces [20]. Degree 2 captures quadratic nonlinearities common in physical systems. The library can be extended to include derivative terms [ux,uxx,u⋅ux,…][u_x,u_{xx},u\cdot u_x,\ldots] for explicit representation of convective and diffusive dynamics. The current experiments use polynomial-only libraries to evaluate the latent MLP’s capacity for implicit derivative learning; results show high correlations (up to 0.99 with uxxu_{xx} for Heat; see Appendix § 13, validating that derivative structure emerges in the latent space even without explicit library terms.

Continuous-Time Koopman Formulation

A critical design choice distinguishes continuous-time from discrete-time Koopman operators. The discrete formulation zt+Δt=Kztz_{t+\Delta t}=Kz_t suffers from diagonal dominance as Δt→0\Delta t \to0:

K=eAΔt≈I+AΔt+O(Δt2)(8)K=e^{A\Delta t}\approx I+A\Delta t+O(\Delta t^2) \tag*{(8)}

For small timesteps, KK approaches the identity matrix, with off-diagonal elements becoming negligible.

Lemma 1 (Continuous Generator Advantage). Let K=eAΔtK=e^{A\Delta t} be the discrete Koopman operator and AA the infinitesimal generator. For Δt→0\Delta t \to0:

(i) The discrete operator satisfies ∥K−I∥F≤∥A∥FΔt+O(Δt2)\lVert K-I\rVert_F\leq\lVert A\rVert_F\Delta t+O(\Delta t^2), causing off-diagonal entries to vanish.

(ii) The generator AA remains Δt\Delta t-independent, with AijA_{ij} (i≠ji\ne j) directly encoding the rate at which observable gjg_j influences g˙i\dot{g}_i.

(Proof in Appendix § 17.)

The continuous-time formulation directly learns the generator:

dzdt=Az(9)\frac{d z}{d t}=Az \tag*{(9)}

Here, AA is independent of Δt\Delta t, and off-diagonal entries directly represent interaction strengths.

Matrix Exponential Integration for Stiff Systems

For stiff PDEs (Cahn-Hilliard, Kuramoto-Sivashinsky), finite-difference Koopman losses exhibit training instability. Following work on exponential integrators for neural ODEs [24, 25], the framework adopts a matrix exponential formulation providing the exact solution for linear Koopman dynamics:

z(t+Δt)=eAΔtz(t)(10)z(t+\Delta t)=e^{A\Delta t}z(t) \tag*{(10)}

The corresponding loss directly compares propagated states:

Lkoopmanexpm=1N∑i=1N∥eAΔtzi−zi+1∥2(11)\mathcal{L}_{\mathrm{koopman}}^{\mathrm{expm}}=\frac{1}{N}\sum_{i=1}^{N}\left\lVert e^{A\Delta t}z_i-z_{i+1}\right\rVert^2 \tag*{(11)}

Proposition 5 (Unconditional Stability of Matrix Exponential). The matrix exponential propagation (Equation (10)) is unconditionally A-stable: for any Δt>0\Delta t>0 and any matrix AA with eigenvalues having non-positive real parts, ∥eAΔtz∥≤∥z∥\lVert e^{A\Delta t}z\rVert\leq\lVert z\rVert.

The integrating factor Euler method zn+1=eAΔtznz_{n+1}=e^{A\Delta t}z_n has been shown to be the most reliable explicit exponential method for stiff neural ODEs [26].

The matrix exponential is computed via Padé approximation with scaling and squaring, ensuring both accuracy and differentiability for gradient-based training. For stiff PDEs where ∣λmax⁡∣\lvert\lambda_{\max}\rvert can exceed 10410^4, this formulation enables SPIKE to handle fourth-order PDEs and chaotic systems that would otherwise exhibit training collapse (detailed stability analysis in Appendix § 12).

Sparsity Regularization

L1 regularization promotes sparse structure in the AA matrix:

Lsparse=λs∥A∥1=λs∑i,j∣Aij∣(12)\mathcal{L}_{\mathrm{sparse}}=\lambda_s\lVert A\rVert_1=\lambda_s\sum_{i,j}\lvert A_{ij}\rvert \tag*{(12)}

Lemma 2 (Structured Representation via Block Sparsity). Under Definition 2, the generator matrix partitions as:

A=(Alib−libAlib−mlpAmlp−libAmlp−mlp)(13)A = \begin{pmatrix} A_{\mathrm{lib-lib}} & A_{\mathrm{lib-mlp}} \\ A_{\mathrm{mlp-lib}} & A_{\mathrm{mlp-mlp}} \end{pmatrix} \tag*{(13)}

If Wlib=IW_{\mathrm{lib}} = I (identity projection) and Alib−mlp=0A_{\mathrm{lib-mlp}} = 0, then the dynamics of the ii-th monomial observable satisfy:

ddtψi(u)=∑j[Alib−lib]ijψj(u)(14)\frac{d}{dt}\psi_i(u) = \sum_j [A_{\mathrm{lib-lib}}]_{ij}\psi_j(u) \tag*{(14)}

where each non-zero [Alib−lib]ij[A_{\mathrm{lib-lib}}]_{ij} indicates that monomial ψj\psi_j contributes to the time evolution of ψi\psi_i. (Proof in Appendix I.)

Non-zero entries in the library portion of AA directly indicate active polynomial terms, embodying the parsimony principle [6].

Remark 4 (Connection to Koopman Eigenfunction Theory). The observed sparsity in learned AA matrices aligns with the minimal set theorem of Cohen [19]: an NN-dimensional dynamical system admits exactly NN functionally independent Koopman eigenfunctions. The empirical observation that active dimensions ≪\ll latent dimension (Table 20) suggests the learned representation captures a near-minimal generating set, explaining why L1 regularization succeeds.

SystemPIKESPIKEImprovementNon-zero Reduction
Heat76.4%81.5%+5.1%–
Advection99.4%99.5%+0.1%–
Burgers92.4%99.0%+6.6%–
Allen-Cahn97.6%99.8%+2.2%–
KdV100.0%97.9%-2.1%–
Reaction-Diffusion17.3%19.0%+1.7%–
Kuramoto-Sivashinsky99.2%99.8%+0.6%32→832 \to8 (4×)
Schrödinger99.6%99.9%+0.3%17→317 \to3 (5.7×)

Table 20. Sparsity Comparison (Higher = Sparser AA Matrix)

Training Objective

The full training objective combines PDE residual, Koopman consistency, and sparsity:

Ltotal=E[∣N[u]−f∣2]+λkLkoopman+λs∥A∥1+λicLIC+λbcLBC(15)\mathcal{L}_{\mathrm{total}} = \mathbb{E}\left[\left\lvert\mathcal{N}[u] - f \right\rvert^2\right] + \lambda_k\mathcal{L}_{\mathrm{koopman}} + \lambda_s\lVert A\rVert_1 + \lambda_{ic}\mathcal{L}_{IC} + \lambda_{bc}\mathcal{L}_{BC} \tag*{(15)}

The Koopman loss Lkoopman\mathcal{L}_{\mathrm{koopman}} admits multiple formulations corresponding to different integrators:

  • Euler: LkoopmanEuler=E[∥zn+1−(zn+AznΔt)∥2]\mathcal{L}_{\mathrm{koopman}}^{\mathrm{Euler}} = \mathbb{E}\left[\lVert z_{n+1} - (z_n + Az_n\Delta t)\rVert^2\right]

  • RK4: LkoopmanRK4=E[∥zn+1−RK4⁡(zn,A,Δt)∥2]\mathcal{L}_{\mathrm{koopman}}^{\mathrm{RK4}} = \mathbb{E}\left[\lVert z_{n+1} - \operatorname{RK4}(z_n, A, \Delta t)\rVert^2\right]

  • EXPM: LkoopmanEXPM=E[∥zn+1−eAΔtzn∥2]\mathcal{L}_{\mathrm{koopman}}^{\mathrm{EXPM}} = \mathbb{E}\left[\lVert z_{n+1} - e^{A\Delta t}z_n\rVert^2\right] (Equation (10))

All formulations enforce z˙=Az\dot{z} = Az but differ in numerical stability for stiff systems. Training pairs (zn,zn+1)(z_n, z_{n+1}) are sampled at consecutive timesteps tn,tn+1=tn+Δtt_n, t_{n+1} = t_n + \Delta t from collocation points, where zn=g(uθ(x,tn))z_n = g(u_\theta(x, t_n)) is computed by evaluating the PINN at each timestep. This discrete formulation avoids explicit computation of z˙\dot{z} via autograd, instead comparing propagated states directly. The timestep Δt=0.01\Delta t = 0.01 is fixed across all systems (see Appendix C).

Loss weights follow standard multi-task physics-informed learning practice [3]; details in Appendix C.

The proposed framework is referred to as PIKE (Physics-Informed Koopman-Enhanced) and SPIKE (Sparse PIKE). Both yield sparse generator matrices AA: PIKE learns sparse structure implicitly through the Koopman constraint, while SPIKE explicitly enforces sparsity via L1 regularization on AA. Collectively, PIKE/SPIKE denotes the Koopman-regularized PINN approach.

Experiments

Experimental Setup

Experiments span 14 dynamical systems: 9 one-dimensional PDEs (Heat, Advection, Burgers, Allen-Cahn, KdV, Reaction-Diffusion, Cahn-Hilliard, Kuramoto-Sivashinsky, Schrödinger), 3 two-dimensional PDEs (2D Wave, 2D Burgers, Navier-Stokes), and 2 ODEs (Lorenz, SEIR). This selection covers parabolic, hyperbolic, dispersive, chaotic, and stiff dynamics. Full system descriptions are provided in Appendix A.

All models use identical architectures (4-layer MLP, 128 units, tanh activation) with embedding dimension 64. Training proceeds for 5000 steps; loss weights λkoopman=0.1\lambda_{\mathrm{koopman}} = 0.1, λsparse=0.01\lambda_{\mathrm{sparse}} = 0.01 are fixed across all systems (Appendix § 11). To ensure fair comparison, the PINN baseline uses the same architecture, optimizer (Adam, lr=10−310^{-3}), and training budget as PIKE/SPIKE; the only difference is the addition of Koopman regularization. Extended training (10,000+ steps) was tested for PINN on stiff systems (Cahn-Hilliard, 2D Wave) but did not resolve the failure modes, confirming these are fundamental limitations rather than undertrained baselines.

Main Results

Table 1 summarizes key performance improvements across all systems. PIKE/SPIKE achieves consistent gains over standard PINNs, with the most substantial improvements on stiff and chaotic systems.

SystemMetricImprovementBest Method
In-Domain MSE
2D WaveIn-domain8×107×8 \times10^{7}\timesPIKE-Euler
Cahn-HilliardIn-domain106×10^{6}\timesPIKE-Euler
OOD-Space MSE{sup}*`
2D Burgers{sup}†`xy∈[1,2]xy \in[1, 2]38×PIKE-Euler
Advectionx∈[3,5]x \in[3, 5]29×SPIKE-EXPM
Allen-Cahn{sup}†`x∈[3,5]x \in[3, 5]7.5×PIKE-Euler
Navier-Stokesxy∈[1,2]xy \in[1, 2]32×PIKE-Euler
Kuramoto-Sivashinskyx∈[3,5]x \in[3, 5]2.1×SPIKE-EXPM
OOD-Time MSE
Schrödingert∈[3,5]t \in[3, 5]24×SPIKE-EXPM
KdVt∈[3,5]t \in[3, 5]6.3×PIKE-EXPM
Burgerst∈[3,5]t \in[3, 5]2.4×PIKE-Euler
Kuramoto-Sivashinskyt∈[3,5]t \in[3, 5]2.8×PIKE-RK4
Chaotic Systems (Valid Prediction Time)
LorenzValid time184×PIKE-Euler

Table 1. Summary of Key Improvements over PINN Baseline (grouped by metric type)

Figure 2 compares solution quality for representative PDEs, showing that PIKE/SPIKE maintains accuracy in OOD regions where PINN exhibits oscillations and phase drift.

Challenging dynamics. For chaotic systems, PIKE-Euler achieves 184× improvement in valid prediction time for Lorenz (12.91s vs 0.07s, corresponding to 11.7 vs 0.06 Lyapunov times). For 2D Navier-Stokes channel flow (an open-domain system), PIKE-Euler yields 32× lower OOD-Space MSE, representing physically meaningful downstream prediction. The improvement magnitude correlates with system stiffness and nonlinearity: fourth-order PDEs (Cahn-Hilliard: 106×10^{6}\times) and chaotic systems benefit most from Koopman regularization, while well-posed parabolic systems (Heat) show moderate but consistent gains. For bounded-domain systems (2D Burgers, Allen-Cahn), OOD-Space improvements reflect extrapolation regularity—Koopman regularization prevents catastrophic failure and learns smoother representations, even though spatial extrapolation is not physically meaningful for these systems.

Integrator comparison. PIKE-Euler achieves lowest error on non-stiff systems; SPIKE-EXPM excels on stiff PDEs where EXPM provides unconditional stability (Euler: max⁡Re⁡(λ)=0.21\max\operatorname{Re}(\lambda) = 0.21; EXPM: 0.00). Computational overhead: <5% for Euler/RK4, ~25% for EXPM (Appendix § 15).

Solution comparison for Burgers and Advection equations

Figure 2. Solution comparison for Burgers and Advection equations. Columns show temporal evolution from in-domain (t≤1t \le1) to OOD extrapolation (t=1.5t = 1.5). PINN exhibits oscillations near the Burgers shock and phase drift in Advection; PIKE/SPIKE variants maintain accuracy through Koopman regularization. Quantitative MSE values in Appendix § 12 (Table 7–Table 9).

SystemPINNPIKE-EulerPIKE-RK4PIKE-EXPMSPIKE-EXPM
Heat1.01e-056.95e-061.20e-056.02e-068.67e-06
Advection5.04e-078.89e-054.45e-074.88e-073.90e-07
Burgers9.68e-054.63e-053.44e-053.23e-053.23e-05
Allen-Cahn1.19e-051.56e-051.44e-052.34e-057.08e-05
KdV5.23e-024.75e-025.24e-025.06e-025.14e-02
Reaction-Diffusion5.02e-052.67e-044.80e-054.63e-051.51e-04
Cahn-Hilliard3.53e-012.61e-073.50e-013.46e-011.22e-01
Schrödinger2.50e+012.53e+012.49e+012.47e+011.20e+01
Kuramoto-Sivashinsky1.83e+018.77e+016.75e+022.93e+021.91e+01

Table 7. Integrator Ablation: In-Domain Physics Residual MSE

SystemPINNPIKE-EulerPIKE-RK4PIKE-EXPMSPIKE-EXPM
Heat{sup}‡`1.52e-057.57e-061.53e-051.50e-051.46e-05
Advection1.51e-065.95e-059.21e-078.66e-075.12e-08
Burgers{sup}‡`6.56e-031.67e-027.70e-036.97e-036.97e-03
Allen-Cahn{sup}‡`1.35e-011.81e-021.55e-016.36e-025.18e-02
Cahn-Hilliard{sup}‡`1.31e+013.36e-061.28e+011.05e+015.24e-01

Table 9. Integrator Ablation: OOD-Space MSE (x∈[3,5]x \in[3, 5])

Sparsity and interpretability. L1 regularization yields up to 5.7× reduction in non-zero entries (Schrödinger: 17 → 3). Latent MLP features show high correlations with derivatives (0.99 with uxxu_{xx} for Heat), enabling post-hoc coefficient recovery with < 1% error (Appendix § 12, E).

Discussion

Key insights. Three main findings emerge: (1) Koopman regularization acts as a dynamics-aware constraint that prevents PINN overfitting, yielding 2–184× improvements in temporal extrapolation; (2) the continuous-time formulation with matrix exponential integration provides unconditional stability for stiff PDEs where standard PINNs fail; (3) sparse generator matrices (≤99.9% zeros) enable interpretable dynamics recovery while maintaining accuracy. Improvements are most pronounced for stiff PDEs, temporal extrapolation, and chaotic systems (Lorenz: 184× longer prediction). These results align with Proposition 4: smaller ϵK\epsilon_K (tight Koopman fit) and bounded ρ0\rho_0 (stable eigenvalues) directly reduce extrapolation error.

Temporal vs. spatial extrapolation. Temporal OOD is physically meaningful for all systems. Spatial OOD tests downstream prediction for open-domain problems (channel flow, Advection) and extrapolation regularity for bounded-domain problems—demonstrating that Koopman regularization prevents catastrophic failure outside training regions (details in Appendix § 12).

Unlike prior physics-informed Koopman methods [18] that require explicit dynamics f(x)f(x), SPIKE operates directly from PDE residuals. The library-latent decomposition enables both structured interpretation and correlation-based analysis. The current experiments use polynomial-only libraries to evaluate implicit derivative learning; the library can be extended to include derivative terms (uxu_x, uxxu_{xx}, u⋅uxu \cdot u_x) for explicit representation of convective and diffusive dynamics.

On improvement magnitudes. The extreme improvements for certain systems (Cahn-Hilliard, 2D Wave) reflect genuine PINN failure modes: fourth-order stiffness and phase coherence requirements that unconstrained networks cannot maintain. Koopman regularization addresses these fundamental limitations through structural constraints on learned dynamics.

Integrator selection. The choice between Euler, RK4, and EXPM involves accuracy-stability trade-offs. Euler is fastest but conditionally stable; RK4 provides fourth-order accuracy; EXPM is unconditionally stable, critical for stiff PDEs (∣λmax⁡∣>104|\lambda_{\max}| > 10^4) at ∼25% overhead. The OOD bound (Proposition 4) depends on both ϵK\epsilon_K and ρ0\rho_0: for moderately stiff systems, Euler’s tighter Koopman fit can outweigh mild instability over short extrapolation windows, while EXPM becomes essential for highly stiff systems.

Navier-Stokes case study. The 2D Navier-Stokes results illustrate integrator selection: PIKE-Euler achieves 82×82\times lower in-domain error for channel flow and 2.3×2.3\times for lid-driven cavity, with corresponding OOD improvements (32×32\times spatial, 15% temporal). Where PIKE does not outperform PINN, both achieve comparable performance within the same order of magnitude.

Inverse problem capability. Beyond forward simulation, PIKE/SPIKE enables post-hoc coefficient recovery: after training, PDE coefficients can be estimated via least-squares regression on PINN derivatives without retraining. Across 9 PDEs/ODEs, coefficient recovery achieves <1%< 1\% relative error (Heat uxxu_{xx}: 0.33%, Burgers u⋅uxu \cdot u_x: 0.01%, KdV uxxxu_{xxx}: 0.21%, Lorenz σ\sigma, ρ\rho, β\beta: <0.3%< 0.3\%). This leverages smooth autograd derivatives from uθu_\theta, avoiding noise amplification from finite differences. Details in Appendix § 13.

Limitations. The polynomial library component of g(u)g(u) assumes that relevant dynamics can be captured by low-degree polynomials in uu. For PDEs with strongly nonlinear or non-polynomial terms (e.g., sin⁡(u)\sin(u), eue^u), the library may require extension. However, the latent MLP component partially compensates by learning implicit correlations with derivative structure (up to 0.99 with uxxu_{xx}), though these remain non-symbolic. The current framework also assumes smooth solutions; discontinuities (shocks in hyperbolic PDEs) may require specialized treatment. Finally, spatial OOD improvements for bounded-domain problems should be interpreted as extrapolation regularity rather than physically meaningful prediction.

Conclusion

This work presented PIKE/SPIKE, combining PINNs with continuous-time Koopman operators for parsimonious dynamics representations with improved generalization. The augmented embedding decomposes observables into interpretable polynomial and flexible MLP components, avoiding identity collapse while enabling L1L1 sparsity (up to 5.7×5.7\times reduction). Experiments on 14 systems demonstrate consistent improvement in temporal extrapolation, with Koopman regularization preventing catastrophic failure outside training regions. For open-domain systems, this yields physically meaningful spatial generalization; for bounded domains, it ensures smooth extrapolation behavior. The latent MLP’s correlation with derivative structure (up to 0.99 with uxxu_{xx}) suggests derivative-augmented libraries as a promising extension.

System Descriptions

The benchmark suite spans 14 dynamical systems selected to cover diverse mathematical structures and physical phenomena. The selection includes: (1) parabolic PDEs (Heat, Reaction-Diffusion) testing diffusion-dominated dynamics; (2) hyperbolic PDEs (Advection, Wave) testing transport and propagation; (3) dispersive PDEs (KdV, Schrödinger) testing wave dispersion and soliton dynamics; (4) stiff PDEs (Cahn-Hilliard, Kuramoto-Sivashinsky) testing fourth-order operators with extreme eigenvalue spreads; (5) nonlinear PDEs (Burgers, Allen-Cahn, Navier-Stokes) testing shock formation and bistable dynamics; and (6) chaotic systems (Lorenz) testing sensitive dependence on initial conditions. This comprehensive coverage ensures that observed improvements are not artifacts of specific equation types.

The following dynamical systems are evaluated:

1D PDEs:

  • Heat: ut=αuxxu_t = \alpha u_{xx} (parabolic diffusion)

  • Advection: ut+cux=0u_t + c u_x = 0 (hyperbolic transport)

  • Burgers: ut+uux=νuxxu_t + u u_x = \nu u_{xx} (nonlinear convection-diffusion)

  • Allen-Cahn: ut=ϵuxx+u−u3u_t = \epsilon u_{xx} + u - u^3 (bistable reaction-diffusion)

  • KdV: ut+uux+uxxx=0u_t + u u_x + u_{xxx} = 0 (dispersive solitons)

  • Reaction-Diffusion: ut=Duxx+R(u)u_t = D u_{xx} + R(u) (pattern formation)

  • Cahn-Hilliard: ut=−ϵ2uxxxx+(u3−u)xxu_t = -\epsilon^2 u_{xxxx} + (u^3 - u)_{xx} (fourth-order phase separation)

  • Kuramoto-Sivashinsky: ut+uux+uxx+uxxxx=0u_t + u u_x + u_{xx} + u_{xxxx} = 0 (spatiotemporal chaos)

  • Schrödinger: iut+uxx+∣u∣2u=0i u_t + u_{xx} + |u|^2 u = 0 (complex dispersive)

2D PDEs:

  • 2D Wave: utt=c2(uxx+uyy)u_{tt} = c^2(u_{xx} + u_{yy}) (hyperbolic propagation)

  • 2D Burgers: ut+uux+vuy=ν(uxx+uyy)u_t + u u_x + v u_y = \nu(u_{xx} + u_{yy})

  • Navier-Stokes: ut+(u⋅∇)u=−∇p+ν∇2u\mathbf{u}_t + (\mathbf{u} \cdot\nabla)\mathbf{u} = -\nabla p + \nu\nabla^2 \mathbf{u} (incompressible flow)

ODEs:

  • Lorenz: x˙=σ(y−x),y˙=x(ρ−z)−y,z˙=xy−βz\dot{x} = \sigma(y - x), \dot{y} = x(\rho- z) - y, \dot{z} = xy - \beta z (chaotic attractor, σ=10\sigma= 10, ρ=28\rho= 28, β=8/3\beta= 8/3)

  • SEIR: S˙=−βSI/N,E˙=βSI/N−σE,I˙=σE−γI,R˙=γI\dot{S} = -\beta SI/N, \dot{E} = \beta SI/N - \sigma E, \dot{I} = \sigma E - \gamma I, \dot{R} = \gamma I

Reproduction Details

We provide complete specifications for experimental reproducibility. Let Ωtrain⊂Rd×R+\Omega_{\mathrm{train}} \subset\mathbb{R}^{d} \times\mathbb{R}^{+} denote the training domain and ΩOOD\Omega_{\mathrm{OOD}} the out-of-distribution evaluation domain.

Domain Specifications and Boundary Conditions

1D PDEs. For all one-dimensional systems, the training domain is Ωtrain=[0,1]×[0,1]\Omega_{\mathrm{train}} = [0, 1] \times[0, 1] (spatial ×\times temporal). Table 2 specifies initial conditions u(x,0)=u0(x)u(x, 0) = u_0(x), boundary conditions, and OOD evaluation domains.

SystemInitial Condition u0(x)u_0(x)Boundary ConditionsOOD Domain
Heatsin⁡(πx)\sin(\pi x)u(0,t)=u(1,t)=0u(0,t) = u(1,t) = 0t∈[1,3]t \in[1, 3]
Advectionsin⁡(2πx)\sin(2\pi x)Periodic: u(0,t)=u(1,t)u(0,t) = u(1,t)x∈[1,3]x \in[1, 3]
Burgers−sin⁡(πx)-\sin(\pi x)u(0,t)=u(1,t)=0u(0,t) = u(1,t) = 0t∈[1,3]t \in[1, 3]
Allen-Cahnx2cos⁡(πx)x^2\cos(\pi x)∂xu(0,t)=∂xu(1,t)=0\partial_x u(0,t) = \partial_x u(1,t) = 0t∈[1,3]t \in[1, 3]
KdV2sech⁡2(x−0.5)2\operatorname{sech}^{2}(x - 0.5)Periodic: u(0,t)=u(1,t)u(0,t) = u(1,t)t∈[1,3]t \in[1, 3]
Reaction-Diffexp⁡(−50(x−0.5)2)\exp(-50(x - 0.5)^2)∂xu(0,t)=∂xu(1,t)=0\partial_x u(0,t) = \partial_x u(1,t) = 0t∈[1,3]t \in[1, 3]
Cahn-Hilliard0.1N(0,1)0.1N(0,1) perturbationPeriodic: u,∂xuu, \partial_x u periodict∈[1,3]t \in[1, 3]
KScos⁡(x)(1+sin⁡(x))\cos(x)(1 + \sin(x))Periodic: u(0,t)=u(2π,t)u(0,t) = u(2\pi,t)t∈[1,3]t \in[1, 3]
Schrödingersech⁡(x−0.5)e2ix\operatorname{sech}(x - 0.5)e^{2ix}Periodic: u(0,t)=u(1,t)u(0,t) = u(1,t)t∈[1,3]t \in[1, 3]

Table 2. 1D PDE Specifications: Initial/Boundary Conditions and Evaluation Domains

2D PDEs. The training domain is Ωtrain=[0,1]2×[0,1]\Omega_{\mathrm{train}} = [0, 1]^2 \times[0, 1]. For Navier-Stokes, we evaluate two configurations: (i) channel flow with parabolic inlet u(0,y,t)=4y(1−y)u(0, y, t) = 4y(1-y), v=0v = 0, and outflow ∂xu=0\partial_x u = 0; (ii) lid-driven cavity with u(x,1,t)=(1,0)u(x, 1, t) = (1, 0) and no-slip elsewhere. Both use Re⁡=100\operatorname{Re} = 100.

SystemInitial ConditionBoundary ConditionsOOD Domain
2D Waveexp⁡(−50((x−0.5)2+(y−0.5)2))\exp(-50((x - 0.5)^2 + (y - 0.5)^2))u∣∂Ω=0u|_{\partial\Omega} = 0 (Dirichlet)x,y∈[1,2]x, y \in[1, 2]
2D Burgersexp⁡(−20((x−0.5)2+(y−0.5)2))\exp(-20((x - 0.5)^2 + (y - 0.5)^2))u∣∂Ω=0u|_{\partial\Omega} = 0 (Dirichlet)x,y∈[1,2]x, y \in[1, 2]
Navier-StokesQuiescent: u(x,0)=0\mathbf{u}(x,0) = 0No-slip / Inlet-Outletx,y∈[1,2]x, y \in[1, 2]

Table 3. 2D PDE Specifications

ODEs. For Lorenz, x0=(1,1,1)T\mathbf{x}_0 = (1, 1, 1)^T with parameters (σ,ρ,β)=(10,28,8/3)(\sigma, \rho, \beta) = (10, 28, 8/3). For SEIR, (S0,E0,I0,R0)=(0.99,0.01,0,0)(S_0, E_0, I_0, R_0) = (0.99, 0.01, 0, 0) with (β,σ,γ)=(0.4,0.2,0.1)(\beta, \sigma, \gamma) = (0.4, 0.2, 0.1) and N=1N = 1. Training domain: t∈[0,1]t \in[0, 1]; OOD: t∈[1,15]t \in[1, 15] (Lorenz), t∈[1,3]t \in[1, 3] (SEIR).

Collocation Point Sampling

Training points are sampled via Latin Hypercube Sampling (LHS) to ensure uniform coverage. Let NcolN_{\mathrm{col}}, NbcN_{\mathrm{bc}}, NicN_{\mathrm{ic}} denote the number of interior collocation, boundary, and initial condition points, respectively:

  • 1D PDEs: Ncol=10,000N_{\mathrm{col}} = 10{,}000, Nbc=200N_{\mathrm{bc}} = 200 (100 per boundary), Nic=100N_{\mathrm{ic}} = 100

  • 2D PDEs: Ncol=20,000N_{\mathrm{col}} = 20{,}000, Nbc=500N_{\mathrm{bc}} = 500 (125 per face), Nic=500N_{\mathrm{ic}} = 500

  • ODEs: Ncol=5,000N_{\mathrm{col}} = 5{,}000, Nic=1N_{\mathrm{ic}} = 1 (single initial condition)

Evaluation Protocol

Physics residual MSE is computed on a held-out uniform grid Gtest\mathcal{G}_{\mathrm{test}} disjoint from training points:

MSEphysics=1∣Gtest∣∑(x,t)∈Gtest∥N[uθ](x,t)∥2(16)\mathrm{MSE}_{\mathrm{physics}} = \frac{1}{\lvert\mathcal{G}_{\mathrm{test}}\rvert} \sum_{(x,t)\in\mathcal{G}_{\mathrm{test}}} \lVert\mathcal{N}[u_{\theta}](x,t)\rVert^{2} \tag*{(16)}

where N[⋅]\mathcal{N}[\cdot] is the PDE residual operator. Grid resolutions: 100×100100 \times100 (1D PDEs), 50×50×5050 \times50 \times50 (2D PDEs), 1,000 temporal points (ODEs).

Ground Truth Generation

Reference solutions are computed as follows:

  • Analytical: Heat (separation of variables), Advection (method of characteristics), KdV (inverse scattering), 2D Wave (d’Alembert)

  • Spectral methods: Allen-Cahn, Cahn-Hilliard (Fourier-Galerkin with N=256N = 256 modes, implicit-explicit time stepping, Δt=10−4\Delta t = 10^{-4})

  • ETDRK4: Kuramoto-Sivashinsky (exponential time differencing fourth-order Runge-Kutta, N=512N = 512 modes)

  • Split-step Fourier: Schrödinger (N=256N = 256 modes, Δt=10−5\Delta t = 10^{-5})

  • Finite differences: Burgers (Cole-Hopf transformation for analytical verification), 2D Burgers (central differences, RK4, Δx=Δy=0.01\Delta x = \Delta y = 0.01)

  • Finite volume: Navier-Stokes (staggered grid, SIMPLE algorithm, Δx=Δy=0.02\Delta x = \Delta y = 0.02, Re=100\mathrm{Re} = 100)

  • Adaptive integration: Lorenz, SEIR (SciPy solve_ivp with RK45, absolute and relative tolerance 10−1010^{-10})

All numerical reference solutions are validated against analytical solutions (where available) or refined meshes to ensure discretization error is negligible (<10−6< 10^{-6}) relative to reported MSE values.

Method Comparison

Table 4 compares SPIKE with related methods across key dimensions.

MethodDerivativesPhysics LossSparsityInverse
PINN [2]AutogradPDE residualNoTraining
SINDy [20]Finite diffNoneLASSOPost-hoc
Deep Koopman [16]NoneNoneNoNo
PIKN [18]AutogradLie operatorNoNo
GN-SINDy [27]NN-basedNoneSINDyYes
SPIKEAutogradPDE residualL1 on AAPost-hoc

Table 5. Detailed Methodological Comparison. Inverse column indicates coefficient recovery capability: “Training” requires unknown coefficients as trainable parameters during optimization; “Post-hoc” enables coefficient recovery after training on forward problems without retraining.

The comparison tables highlight three key distinctions. First, PIKE/SPIKE is the only method that combines Koopman structure with direct PDE solving capability; existing Koopman methods operate on trajectory data or require explicit dynamics f(x)f(x), while PIKE/SPIKE works directly from PDE residuals via autograd. Second, PIKE/SPIKE explicitly targets OOD generalization, whereas most methods focus on training accuracy or system identification. Third, the continuous-time formulation dz/dt=Azdz/dt = Az avoids the identity collapse problem inherent in discrete Koopman operators, enabling meaningful sparsity analysis regardless of timestep size.

Hyperparameter Settings

Table 6 summarizes the hyperparameter settings used across all PDE systems. A unified configuration was employed to ensure fair comparison: all models use identical network architectures (4 hidden layers, 128 units each) with the same embedding dimension and polynomial degree. The loss weights λkoopman=0.1\lambda_{\mathrm{koopman} } = 0.1 and λsparse=0.01\lambda_{\mathrm{sparse} } = 0.01 were selected to balance gradient magnitudes across loss terms, following standard practice in physics-informed learning [2, 3]. During training, the observed magnitudes were Lphysics∼10−4\mathcal{L}_{\mathrm{physics}} \sim10^{-4} and Lkoopman∼10−3\mathcal{L}_{\mathrm{koopman}} \sim10^{-3}, so λkoopman=0.1\lambda_{\mathrm{koopman} } = 0.1 yields comparable gradient contributions. These weights were held fixed across all systems to avoid per-system tuning and demonstrate robustness of the approach.

ArchitectureValueTrainingValue
Hidden layers4λkoopman\lambda_{\mathrm{koopman}}0.1
Hidden units128λsparse\lambda_{\mathrm{sparse}}0.01
Embedding dim64Koopman Δt\Delta t0.01
Poly degree2Training steps5000
ActivationtanhOptimizerAdam

Table 6. Hyperparameter Settings Across PDE Systems

Note: Identical hyperparameters are used across all 14 systems (9 1D PDEs, 3 2D PDEs, 2 ODEs) to ensure fair comparison and demonstrate robustness. No per-system tuning was performed.

The polynomial degree d=2d = 2 provides a library of (n+22)\binom{n+2}{2} terms (6 terms for scalar PDEs: [1,u,u2][1, u, u^2] plus MLP features). Higher degrees were tested but yielded diminishing returns while increasing computational cost. The embedding dimension of 64 balances expressivity with the sparsity objective; larger dimensions provide more capacity but require stronger regularization to achieve comparable sparsity levels.

Detailed Quantitative Results

This section provides the full quantitative analysis supporting the main experimental claims.

Integrator Ablation

Table 7 compares the in-domain physics residual MSE across different Koopman integrators. PIKE-Euler achieves best performance on non-stiff systems, while EXPM variants excel on stiff PDEs (Cahn-Hilliard, Kuramoto-Sivashinsky).

SystemPINNPIKE-EulerPIKE-RK4PIKE-EXPMSPIKE-EXPM
Heat7.52e-042.17e-037.00e-047.97e-048.03e-04
Advection4.05e-102.96e-096.02e-104.48e-101.06e-09
Burgers2.77e-021.15e-022.97e-022.91e-022.91e-02
KdV8.89e-032.49e-029.99e-031.42e-031.81e-03
Schrödinger3.09e+012.86e+013.69e+014.19e+011.29e+00

Table 8. Integrator Ablation: OOD-Time MSE (t∈[3,5]t \in[3, 5])

{sup}‡`Bounded-domain problems where spatial extrapolation demonstrates model capability rather than physical prediction.

SystemPINNPIKE-EulerPIKE-RK4PIKE-EXPMSPIKE-EXPM
Wave 2D1.72e+042.06e-042.17e+057.69e+045.11e-02
Burgers 2D9.66e-037.12e-034.92e-023.23e-029.34e-03
Navier-Stokes1.09e+011.33e-014.31e+002.25e+012.47e+01
NS Lid-Driven{sup}†`3.74e-011.62e-013.33e-011.66e-015.25e-01

Table 10. Integrator Ablation: 2D PDEs In-Domain MSE

{sup}†`Lid-driven cavity: PIKE-Euler achieves 2.3× lower in-domain MSE than PINN.

‡OOD-Space is not physically meaningful for bounded-domain problems (fixed boundaries); results demonstrate extrapolation capability rather than physical prediction.

SystemPINNPIKE-EulerPIKE-RK4PIKE-EXPMSPIKE-EXPM
Wave 2D3.19e+002.06e-046.23e-041.81e-021.16e-02
Burgers 2D7.87e-029.29e-026.80e-027.88e-026.12e-02
Navier-Stokes2.24e-022.78e-022.94e-022.67e-018.93e-01
NS Lid-Driven6.32e-021.17e-015.40e-021.18e-011.19e+00

Table 12. Integrator Ablation: 2D PDEs OOD-Time MSE (t∈[1,3]t \in[1, 3])

SystemPINNPIKE-EulerPIKE-RK4PIKE-EXPMSPIKE-EXPM
Lorenz5.48e+044.79e+045.03e+044.35e+044.35e+04
SEIR4.70e-024.26e-025.26e-025.73e-025.73e-02

Table 13. Integrator Ablation: ODE Systems In-Domain MSE

System (OOD range)PINNPIKE-EulerPIKE-RK4PIKE-EXPMSPIKE-EXPM
Lorenz (t∈[1,3]t \in[1, 3])2.96e+011.19e+011.95e+011.29e+011.29e+01
Lorenz (t∈[3,5]t \in[3, 5])9.42e-051.22e-051.87e-041.22e-051.22e-05
SEIR (t∈[1,3]t \in[1, 3])1.84e-031.42e-031.67e-032.10e-032.10e-03
SEIR (t∈[3,5]t \in[3, 5])1.40e-042.14e-045.22e-042.95e-042.95e-04

Table 14. Integrator Ablation: ODE OOD-Time MSE (t∈[1,3]t \in[1, 3] and t∈[3,5]t \in[3, 5])

Hyperparameter Sensitivity Analysis

This ablation study examines the Lorenz system with λkoopman∈{0.01,0.05,0.1,0.5,1.0}\lambda_{\mathrm{koopman}} \in\{0.01, 0.05, 0.1, 0.5, 1.0\} and λsparse∈{0.0,0.001,0.01,0.1,1.0}\lambda_{\mathrm{sparse}} \in\{0.0, 0.001, 0.01, 0.1, 1.0\}. Table 15 shows physics residual MSE, and Table 16 shows Koopman latent R2R^2.

λkoopman\lambda_{\mathrm{koopman}}λs=0.0\lambda_s=0.0λs=0.001\lambda_s=0.001λs=0.01\lambda_s=0.01λs=0.1\lambda_s=0.1λs=1.0\lambda_s=1.0
0.012.36e-022.14e-022.14e-022.14e-022.14e-02
0.052.48e-022.48e-022.48e-022.48e-022.60e-02
0.102.70e-022.70e-022.70e-022.70e-022.55e-02
0.502.79e-022.79e-022.79e-022.79e-022.79e-02
1.002.96e-022.96e-022.96e-022.96e-022.96e-02

Table 15. Lambda Ablation: Physics Residual MSE (Lorenz)

λkoopman\lambda_{\mathrm{koopman}}λs=0.0\lambda_{\mathrm{s}}=0.0λs=0.001\lambda_{\mathrm{s}}=0.001λs=0.01\lambda_{\mathrm{s}}=0.01λs=0.1\lambda_{\mathrm{s}}=0.1λs=1.0\lambda_{\mathrm{s}}=1.0
0.010.8110.8000.8000.8000.800
0.050.8410.8410.8410.8410.848
0.100.7380.7380.7380.7380.761
0.500.0930.0930.0930.0930.093
1.000.0780.0780.0780.0780.078

Table 16. Lambda Ablation: Koopman Latent R2R^2 (Lorenz)

Table 17 shows Lyapunov analysis metrics. Valid prediction time remains constant across all configurations (≈0.07\approx0.07 s), indicating that λ\lambda selection primarily affects local physics fitting and Koopman representation quality rather than long-term trajectory prediction.

λkoopman\lambda_{\mathrm{koopman}}In-Domain MSEOOD MSEValid Time (s)τ\tau ratio
0.01312.0291.80.070.06×0.06\times
0.05312.0291.70.070.06×0.06\times
0.10311.9291.60.070.06×0.06\times
0.50311.9291.50.070.06×0.06\times
1.00311.8291.40.070.06×0.06\times

Table 17. Lambda Ablation: Lyapunov Metrics (Lorenz, τλ=1.1\tau_{\lambda}=1.1 s)

Key findings: (1) Physics residual improves with lower λkoopman\lambda_{\mathrm{koopman}}, with best performance at λk=0.01\lambda_k = 0.01. (2) Koopman R2R^2 peaks at moderate λkoopman\lambda_{\mathrm{koopman}} (0.05), suggesting a trade-off between physics and Koopman fidelity. (3) High λkoopman\lambda_{\mathrm{koopman}} (≥0.5\ge0.5) severely degrades Koopman representation quality (R2<0.1R^2 < 0.1). (4) Sparsity regularization (λsparse\lambda_{\mathrm{sparse}}) has minimal effect within the tested range. (5) Trajectory prediction (valid time) is insensitive to λ\lambda choice, remaining at ≈0.07\approx0.07 s across all configurations. All configurations maintain Koopman stability except λk=0.1\lambda_k = 0.1, λs=1.0\lambda_s = 1.0.

Lyapunov Analysis (Chaotic Systems)

For chaotic systems like Lorenz, the Lyapunov time τλ≈1.1\tau_{\lambda} \approx1.1 s represents the characteristic timescale over which nearby trajectories diverge. Valid prediction time measures how long the model tracks the true trajectory before relative error exceeds 50%.

ModelIn-Domain MSEOOD MSEShort-term MSEValid Time (s)τ\tau ratio
PINN61.4065.8623.600.070.06×0.06\times
PIKE-Euler60.3669.240.1612.9111.73×11.73\times
PIKE-RK461.3466.0119.610.070.06×0.06\times
PIKE-EXPM60.2869.450.6112.9111.73×11.73\times
SPIKE-EXPM60.2869.450.6112.9111.73×11.73\times

Table 18. Lorenz Lyapunov Analysis (τλ=1.1\tau_{\lambda}=1.1 s)

The Lyapunov analysis reveals that PIKE/SPIKE achieves 184×184\times longer valid prediction time than PINN (12.91s vs 0.07s), corresponding to approximately 12 Lyapunov times versus less than 0.1 Lyapunov times. This dramatic improvement stems from the Koopman constraint enforcing globally consistent dynamics: while PINN can achieve low training loss through local fitting, it fails to capture the attractor geometry that governs long-term behavior. The short-term MSE column shows that PIKE-Euler maintains trajectory accuracy (0.16) within one Lyapunov time, whereas PINN already exhibits significant deviation (23.60).

Kuramoto-Sivashinsky also exhibits spatiotemporal chaos, though Lyapunov time is not directly applicable to PDEs. Table 19 shows SPIKE-EXPM achieves 2.1× improvement in OOD-Space extrapolation for this chaotic PDE.

MetricPINNPIKE-EulerPIKE-RK4PIKE-EXPMSPIKE-EXPM
In-Domain MSE1.83e+018.77e+016.75e+022.93e+021.91e+01
OOD-Space (x∈[3,5]x \in[3, 5])5.64e-024.09e-024.13e-023.70e-022.65e-02
OOD-Time (t∈[3,5]t \in[3, 5])3.44e-049.97e-041.21e-046.14e-041.01e-03

Table 19. Kuramoto-Sivashinsky (Spatiotemporal Chaos) Performance

Several patterns emerge from the integrator ablation. First, PIKE-Euler consistently achieves strong performance on non-stiff systems (Heat, Allen-Cahn, 2D PDEs), suggesting that the simplest integrator suffices when eigenvalue magnitudes remain moderate. Second, the dramatic improvement on Cahn-Hilliard (106×10^{6}\times for PIKE-Euler) and 2D Wave (8×107×8 \times10^{7}\times) indicates that Koopman regularization fundamentally changes the learned dynamics structure. Third, SPIKE-EXPM excels on dispersive systems (Schrödinger: 24× OOD-Time improvement) where unconditional stability prevents error accumulation during extrapolation. The choice of integrator should match the stiffness characteristics of the target PDE.

Sparsity Analysis

Table 20 compares the sparsity of the learned Koopman matrix AA between PIKE (without L1 regularization) and SPIKE (with L1 regularization). Sparsity is measured as the percentage of entries with ∣Aij∣<10−4\lvert A_{ij} \rvert< 10^{-4}.

The results demonstrate consistent sparsity improvement across most systems. Notable exceptions include KdV, where the already-high PIKE sparsity leaves little room for improvement. The most significant gains occur for Burgers (+6.6%) and the complex systems (Kuramoto-Sivashinsky, Schrödinger), where the non-zero element count is reduced by 4–5.7×. These reductions directly translate to simpler candidate term sets (Definition 3) for inverse problem coefficient estimation.

Koopman Stability Analysis

Table 21 analyzes the stability of learned Koopman generators by examining the maximum real part of eigenvalues of AA. For continuous-time dynamics z˙=Az\dot{z} = Az, stability requires max⁡iRe⁡(λi)≤0\max_i \operatorname{Re}(\lambda_i) \le0; a threshold of 0.01 is used to account for numerical precision.

SystemPIKE-EulerPIKE-RK4PIKE-EXPMSPIKE-EXPM
Heat0.00810.00000.00000.0000
Advection0.01220.00000.00000.0000
Burgers0.21120.00030.00000.0000
Reaction-Diffusion0.06140.00000.00000.0000
Cahn-Hilliard0.00830.00000.00000.0002
Kuramoto-Sivashinsky0.00230.00000.00000.0000
NS Lid-Driven0.01660.00140.00010.0000
Lorenz0.19500.00000.00000.0000

Table 21. Koopman Generator Stability: Max Real Eigenvalue (stable if ≤0.01\le0.01)

A clear pattern emerges: implicit integrators (PIKE-RK4, PIKE-EXPM, SPIKE-EXPM) consistently learn stable dynamics across all systems. Explicit Euler can yield mildly positive eigenvalues for some systems, which motivates the matrix exponential formulation. This demonstrates that integrator choice directly controls learned stability, with EXPM providing guaranteed stabilization.

Stability-performance tradeoff. The OOD bound (Proposition 4) depends on both Koopman consistency ϵK\epsilon_K and spectral radius ρ0\rho_0. For moderately stiff systems like Burgers, PIKE-Euler may learn a mildly unstable generator (ρ0=0.21\rho_0 = 0.21) yet achieve better OOD performance than PIKE-EXPM (ρ0=0\rho_0 = 0) because the simpler Euler loss enables tighter Koopman fit (smaller ϵK\epsilon_K). Over short extrapolation windows, the ϵK\epsilon_K reduction dominates the mild instability penalty. Concretely, for Burgers with δ=2\delta= 2 (OOD window t∈[3,5]t \in[3, 5]), the instability growth factor is (e0.21×2−1)/0.21≈2.5(e^{0.21 \times2} - 1)/0.21 \approx2.5, only marginally worse than the stable limit of δ=2\delta= 2, insufficient to offset PIKE-Euler’s superior Koopman fit. For highly stiff systems (Cahn-Hilliard, Kuramoto-Sivashinsky) where ∣λmax⁡∣>104|\lambda_{\max}| > 10^4, EXPM becomes essential as even small positive ρ0\rho_0 causes rapid divergence.

On PINN overfitting. Standard PINNs minimize physics residuals without structural constraints on the learned solution manifold. While the network can achieve near-zero training loss, nothing prevents it from learning spurious high-frequency components that satisfy the PDE locally but extrapolate poorly. The Koopman constraint acts as a dynamics-aware regularizer: by requiring that observables evolve linearly, the network is implicitly encouraged to learn smooth, physically meaningful representations rather than memorizing training data.

On improvement magnitudes. The extreme improvements for certain systems (Cahn-Hilliard: 106×10^6\times, 2D Wave: 8×107×8 \times10^7\times) reflect genuine PINN failure modes rather than weak baselines. Fourth-order PDEs like Cahn-Hilliard exhibit severe stiffness (∣λmax⁡∣>104|\lambda_{\max}| > 10^4), causing standard PINNs to learn unstable dynamics that diverge rapidly outside the training domain. Similarly, 2D Wave solutions require precise phase coherence that unconstrained networks fail to maintain. These results demonstrate that PINNs have fundamental limitations for certain PDE classes—limitations that Koopman regularization directly addresses by imposing structural constraints on the learned dynamics. The practical implication is that PIKE/SPIKE enables reliable solutions for stiff and wave-dominated systems where vanilla PINNs are unsuitable. Extended training (10,000+ steps) was tested for PINN on these stiff systems but did not resolve the failure modes, confirming these are fundamental limitations rather than undertrained baselines.

Invariance and Conservation Analysis

Table 22 evaluates physical conservation properties by measuring the relative standard deviation of mass (∫u dx\int u\,dx) and energy (∫u2 dx\int u^2\,dx) over time. Lower values indicate better conservation. We focus on systems with known conservation laws or physical invariants:

  • KdV: Integrable system conserving infinite quantities including mass ∫u dx\int u\,dx and energy ∫u2 dx\int u^2\,dx

  • Schrödinger: Conserves probability ∫∣u∣2 dx\int|u|^2\,dx (unitarity)

  • Wave 2D: Conserves total energy (kinetic + potential)

  • SEIR: Conserves total population S+E+I+R=NS + E + I + R = N

  • Advection: Conserves mass under periodic boundaries

Systems without conservation laws are excluded: Heat (dissipative), Burgers (shock dissipation), Allen-Cahn (reaction), Lorenz (strange attractor with volume contraction), Kuramoto-Sivashinsky (chaotic dissipation), Navier-Stokes (viscous dissipation).

SystemMass Conservation PINNMass Conservation SPIKE-EXPMEnergy Conservation PINNEnergy Conservation SPIKE-EXPM
Advection1.83e-031.21e-033.67e-032.43e-03
KdV4.95e-024.98e-029.18e-029.24e-02
Schrödinger8.31e-024.27e-021.64e-018.53e-02
Wave 2D1.03e+006.17e-011.59e-063.13e-01
SEIR5.42e-055.97e-052.73e-053.17e-05

Table 22. Conservation Properties: Relative Standard Deviation of Conserved Quantities (lower = better)

For integrable systems with exact conservation laws (KdV, Schrödinger), SPIKE achieves comparable or improved conservation over PINN. The most significant improvement occurs for Schrödinger (48% reduction in mass deviation), where the Koopman constraint helps maintain unitarity. For SEIR, both methods achieve excellent population conservation (<10−4< 10^{-4} relative std), as expected from the closed-system formulation. Wave 2D shows improved mass conservation with SPIKE but slightly worse energy conservation, likely due to the energy being distributed between kinetic and potential components not captured by ∫u2 dx\int u^2\,dx alone.

Interpretability Analysis Details

Library-Latent Decomposition Framework

The SPIKE framework provides interpretability through the augmented embedding (Definition 2) and block-sparse structure (Lemma 2). The library component glib(u)=Wlib⋅ψd(u)g_{\mathrm{lib}}(u) = W_{\mathrm{lib}} \cdot\psi_d(u) provides a structured representation: non-zero entries in Alib−libA_{\mathrm{lib-lib}} correspond to active polynomial terms via Equation (14). The latent component gmlp(u)g_{\mathrm{mlp}}(u) captures implicit correlations with physical quantities (uxu_x, uxxu_{xx}, etc.). For forward problems, this provides validation that the learned structure matches the physics specified in Lphysics\mathcal{L}_{\mathrm{physics}}; for inverse problems, it enables coefficient recovery.

PDETrue EquationLibrary dg1/dtdg_1/dtLatent Correlations
Heatut=αuxxu_t = \alpha u_{xx}+0.07g2+0.07g_2uxxu_{xx} (0.99)
Advectionut=−cuxu_t = -cu_x00uxu_x (0.96), utu_t (0.96)
Burgersut=−uux+νuxxu_t = -uu_x + \nu u_{xx}−0.06g0+0.04g1-0.06g_0 + 0.04g_1u−u3u-u^3 (0.97), uxxu_{xx} (0.27)
Allen-Cahnut=ϵuxx+u−u3u_t = \epsilon u_{xx} + u - u^3+0.03g0−0.03g1+0.03g_0 - 0.03g_1u−u3u-u^3 (0.85), uxxu_{xx} (0.57)
KdVut=−uux−uxxxu_t = -uu_x - u_{xxx}00u3u^3 (1.00), uxxu_{xx} (0.94)
Reaction-Diffusionut=Duxx+R(u)u_t = Du_{xx} + R(u)−0.03g0+0.01g1-0.03g_0 + 0.01g_1uxxu_{xx} (0.76), u3u^3 (0.91)
Kuramoto-Sivashinskyut=−uux−uxx−uxxxxu_t = -uu_x - u_{xx} - u_{xxxx}≈0\approx0 (4×\times sparser)u3u^3 (0.96), uxxu_{xx} (0.38)
Schrödingeriut=uxx+∣u∣2uiu_t = u_{xx} + |u|^2u−0.01g0-0.01g_0uRu_R (0.55), ∣u∣2|u|^2 (0.36)

Table 23. Learned Dynamics Structure: Library dg1/dtdg_1/dt and Latent Correlations

Representation Completeness Analysis

The accuracy of the Koopman representation depends on whether the true PDE terms lie within the span of the observable library.

Proposition 6 (Representation Completeness Condition). Let the true dynamics satisfy ut=F(u,ux,uxx,…)u_t = \mathcal{F}(u,u_x,u_{xx},\ldots). The library ψd(u)\psi_d(u) can exactly represent the dynamics if and only if F\mathcal{F} is a polynomial in uu of degree at most dd with no explicit dependence on spatial derivatives.

Proof. By Lemma 2, the library dynamics satisfy ψ˙i=∑j[Alib−lib]ijψj\dot{\psi}_i = \sum_j [A_{\mathrm{lib-lib}}]_{ij}\psi_j, expressing ψ˙i\dot{\psi}_i as a linear combination of monomials in uu. If F\mathcal{F} contains terms like uxu_x or u⋅uxu \cdot u_x, these cannot be expressed in the monomial basis ψd(u)=[1,u,u2,…]\psi_d(u) = [1,u,u^2,\ldots], as spatial derivatives are not functions of uu alone. The converse holds by construction: any polynomial F(u)=∑kckuk\mathcal{F}(u) = \sum_k c_k u^k with k≤dk \le d admits exact representation via appropriate Alib−libA_{\mathrm{lib-lib}} entries.

Table 24 summarizes which physical terms are captured by the library representation and which require the latent MLP.

PDECaptured TermsMissed TermsPrimary Limitation
Reaction-Diffusionuxxu_{xx} (corr. 0.76), R(u)R(u)Exact R(u)R(u) formPolynomial approx.
Advectionuxu_x, utu_t (corr. 0.96)–Library trivial
Heatuxxu_{xx} (corr. 0.99)–Library form incorrect
Allen-Cahnu−u3u-u^3 (corr. 0.85)Correct signsCoefficient accuracy
KdVuxxu_{xx} (corr. 0.94), u⋅uxu \cdot u_x (0.54)uxxxu_{xxx}Higher derivatives
Burgersuxxu_{xx} (corr. 0.27)u⋅uxu \cdot u_xConvection term
Kuramoto-Sivashinskyu3u^3 (corr. 0.96)u⋅uxu \cdot u_x, uxxxxu_{xxxx}Chaotic dynamics
Schrödingerureu_{\mathrm{re}} (corr. 0.55)uxxu_{xx}, ∣u∣2u\lvert u\rvert^2uComplex nonlinearity

Table 24. Representation Coverage: Library-Captured vs. Latent-Correlated Terms

By Proposition 6, PDEs with convective (u⋅uxu \cdot u_x) or higher-derivative terms violate the completeness condition. The latent MLP partially compensates: it receives uu values and learns implicit correlations with uxxu_{xx} through solution structure, but cannot express these symbolically.

Inverse Problem Applications

For inverse problems where PDE coefficients are unknown, the sparse AA matrix structure provides a reduced search space for coefficient estimation.

Definition 3 (Candidate Term Set). Given a sparse generator AA with sparsity pattern S={(i,j):Aij≠0}S = \{(i,j): A_{ij} \ne0\}, the candidate term set for observable ψi\psi_i is:

Ci={ψj:(i,j)∈S}(17)C_i = \{\psi_j : (i,j) \in S\} \tag*{(17)}

representing monomials that potentially contribute to ψ˙i\dot{\psi}_i.

The candidate set CiC_i provides a reduced search space: instead of considering all (n+dd)\binom{n+d}{d} possible polynomial terms, only ∣Ci∣≪(n+dd)\lvert C_i\rvert\ll\binom{n+d}{d} candidates require evaluation. For Schrödinger, L1 regularization reduces non-zero entries from 17 to 3 (5.7× reduction). For forward problems with known PDEs, this validates that the learned representation matches expected physics. For inverse problems, the sparse AA matrix serves as a structured prior for coefficient estimation, which can be refined using derivative-augmented libraries [u,ux,uxx,u⋅ux,…][u,u_x,u_{xx},u \cdot u_x,\ldots].

Post-Hoc Coefficient Recovery

A key capability of SPIKE is post-hoc inverse problem solving: after training on forward solutions, PDE coefficients can be recovered via least-squares regression on the learned PINN derivatives without retraining. This leverages the smooth autograd derivatives from uθ(x,t)u_\theta(x,t), avoiding noise amplification from finite differences.

Important distinction: Coefficient recovery uses a separate derivative-based regression on PINN outputs, not the Koopman AA matrix directly. Given the trained PINN uθ(x,t)u_\theta(x,t), autograd computes utu_t, uxu_x, uxxu_{xx}, etc. For Burgers (ut+uux=νuxxu_t + uu_x = \nu u_{xx}), the regression fits:

ut=c1(u⋅ux)+c2uxx(18)u_t = c_1(u \cdot u_x) + c_2u_{xx} \tag*{(18)}

The derivative library [u⋅ux,uxx][u \cdot u_x, u_{xx}] used for regression is distinct from the Koopman observable library [1,u,u2][1,u,u^2]. The Koopman regularization improves PINN accuracy, which in turn yields cleaner derivatives for coefficient recovery, but the AA matrix structure only informs which polynomial terms are active, not derivative terms directly.

SystemCoefficientTrueRecoveredError (%)
Heatuxxu_{xx}0.010.01000.33%
Advectionuxu_x−1.0-1.0−1.0007-1.00070.07%
Burgersu⋅uxu \cdot u_x−1.0-1.0−0.9999-0.99990.01%
Burgersuxxu_{xx}0.010.01000.19%
Allen-Cahnuxxu_{xx}0.010.01010.52%
Allen-Cahnu−u3u-u^31.00.99870.13%
KdVu⋅uxu \cdot u_x−1.0-1.0−0.9994-0.99940.06%
KdVuxxxu_{xxx}−1.0-1.0−1.0021-1.00210.21%
Reaction-Diff.uxxu_{xx}0.010.01000.41%
Lorenzσ\sigma10.09.98470.15%
Lorenzρ\rho28.027.96310.13%
Lorenzβ\beta2.6672.65890.30%

Table 25. PDE/ODE Coefficient Recovery Accuracy (PIKE-EXPM)

Table 25 demonstrates near-exact coefficient recovery with < 1% relative error across all tested systems, including both PDEs and the chaotic Lorenz ODE. The R2^2 values exceed 0.99 in all cases. This capability enables: (1) validation of known physics by checking recovered coefficients against theoretical values; (2) estimation of unknown parameters in systems with known functional form (inverse problems); (3) uncertainty quantification through bootstrap regression on PINN outputs.

Latent Correlation Analysis

A key design choice in SPIKE restricts the polynomial library to functions of uu alone, deliberately excluding spatial derivatives. This separation enables investigation of whether the latent MLP component gmlp(u)g_{\mathrm{mlp}}(u) can implicitly learn derivative-dependent structure from the solution manifold. Table 26 presents correlation coefficients between latent features and finite-difference approximations of spatial derivatives, quantifying this implicit learning.

The correlation analysis validates the design hypothesis:

  • Derivative-correlated representations: High uxxu_{xx} correlations (Heat: 0.99, KdV: 0.94) demonstrate that the MLP learns representations correlated with diffusive structure without explicit derivative features. This correlation indicates that latent features encode derivative-related information, but does not constitute symbolic extraction of derivative terms.

  • Convection structure: Lower u⋅uxu \cdot u_x correlations (Burgers: 0.01) indicate that mixed derivative-polynomial terms are more challenging to capture implicitly, suggesting these require explicit library inclusion.

  • Polynomial terms: High u3u^3 correlations (KdV: 1.00, Kuramoto-Sivashinsky: 0.96) confirm successful polynomial feature learning in the latent space.

PDEuuuxu_xuxxu_{xx}utu_tu⋅uxu \cdot u_xu3u^3
Heat1.000.020.990.370.000.95
Advection0.980.960.910.960.910.97
Burgers1.000.000.270.570.010.88
Allen-Cahn0.950.000.570.850.000.76
KdV0.990.600.940.330.541.00
Reaction-Diffusion1.000.010.760.330.000.91
Kuramoto-Sivashinsky1.000.000.380.410.010.96
Schrödinger0.55–0.09–––

Table 26. Latent Feature Correlation Matrix (Maximum Correlation per Derivative)

The correlation patterns reveal which physical terms the latent MLP can implicitly capture. Diffusion-dominated systems (Heat, Reaction-Diffusion) show strong uxxu_{xx} correlations, indicating successful implicit derivative learning. Convection terms (u⋅uxu \cdot u_x) prove more challenging, with uniformly low correlations across all systems; these require explicit library inclusion for symbolic recovery. The high polynomial correlations (u3u^3: KdV 1.00, Kuramoto-Sivashinsky 0.96) confirm that nonlinear terms within the library basis are accurately represented. Schrödinger presents unique challenges due to complex-valued solutions, resulting in lower correlations overall.

Computational Cost Analysis

Table 27 reports training time normalized per 1000 optimization steps. All experiments use identical architectures (4-layer MLP, 128 units) and were conducted on NVIDIA T4 GPUs. PIKE-Euler adds negligible overhead since the Koopman loss involves only matrix-vector multiplication. PIKE-RK4 requires four substeps but remains within 5% of PINN cost. PIKE-EXPM and SPIKE-EXPM incur ∼25%\sim25\% overhead due to Padé approximation for matrix exponential computation; this cost is justified for stiff PDEs where EXPM provides unconditional stability. The L1 sparsity penalty in SPIKE adds no measurable overhead beyond PIKE-EXPM.

SystemPINNPIKE-EulerPIKE-RK4PIKE-EXPMSPIKE-EXPM
Heat2,3122,1782,2473,0122,984
Burgers2,2092,1722,2392,8772,925
Advection2,2162,1572,2332,6552,660
Allen-Cahn2,2552,2502,2642,9492,966
Korteweg-de Vries2,6502,6082,6893,2153,216
Schrödinger1,5941,5381,5771,8971,993
Reaction-Diffusion2,2881,3722,3202,6372,652
Lorenz5459229541,0271,037
SEIR5931,0201,0681,1211,141
Navier-Stokes 2D2,8603,1142,7303,1172,834
Wave 2D2,8202,7802,8643,2942,010
Burgers 2D2,4543,4083,5814,4392,888

Table 27. Training Time (seconds per 1000 steps)

The computational overhead of Koopman regularization is modest relative to the accuracy gains. For most 1D PDEs, PIKE-Euler matches or slightly improves upon PINN training time due to the regularization effect reducing optimization difficulty. The 2D systems show more variability, with Burgers 2D exhibiting higher PIKE overhead due to the larger observable dimension. ODEs (Lorenz, SEIR) show increased relative overhead because the base PINN cost is lower, making the fixed Koopman computation more prominent. Overall, the <5%<5\% overhead for Euler/RK4 and ∼25%\sim25\% for EXPM represents a favorable cost-accuracy tradeoff given the order-of-magnitude improvements in OOD generalization.

Additional Figures

This section presents supplementary visualizations for the Lorenz chaotic system and 2D Navier-Stokes experiments, demonstrating PIKE/SPIKE’s improved long-term prediction and spatial generalization capabilities.

The Lorenz results (Figure 3) demonstrate PIKE/SPIKE’s ability to maintain trajectory coherence in chaotic systems where small errors grow exponentially. Standard PINNs diverge after ∼0.07s (less than one Lyapunov time), while PIKE-Euler tracks the true attractor for 12.9s (≈12 Lyapunov times). This 184× improvement reflects the Koopman constraint’s role in learning globally consistent dynamics rather than locally accurate but globally unstable solutions.

Lorenz system: chaotic dynamics analysis

Figure 3. Lorenz system: PIKE/SPIKE achieve 184×\times longer valid prediction time than PINN (τPIKE/τPINN=11.02/0.06\tau_{\mathrm{PIKE}}/\tau_{\mathrm{PINN}} = 11.02/0.06, corresponding to ≈11\approx11 Lyapunov times vs ≈0.06\approx0.06).

Navier-Stokes (2D). The incompressible Navier-Stokes equations present a challenging benchmark due to the pressure-velocity coupling and nonlinear convective term (u⋅∇)u\mathbf{u} \cdot\nabla)\mathbf{u}. Figure 5 shows channel flow with two complementary metrics: PIKE-EXPM achieves 5.5× lower solution error vs analytical Poiseuille flow, while PIKE-Euler achieves 23× lower physics residual. This divergence highlights that satisfying the PDE (low residual) does not guarantee recovering the correct physical solution in OOD regions—Koopman regularization helps on both fronts but with different optimal integrators. Figure 4 shows the lid-driven cavity benchmark. Standard PINNs exhibit spurious vortices and flow inconsistencies where training data is sparse, while PIKE/SPIKE maintains physical consistency by enforcing linear dynamics in observable space.

Navier-Stokes 2D lid-driven cavity flow

Figure 4. Navier-Stokes 2D lid-driven cavity flow (Re=100Re = 100). PIKE-Euler achieves 2.3×\times lower MSE than PINN (1.62e-1 vs 3.74e-1). The primary vortex structure and boundary layer near the moving lid are well-captured by PIKE, while PINN shows larger errors in the vortex core region.

Navier-Stokes 2D channel flow: spatial OOD generalization

Figure 5. Navier-Stokes 2D channel flow (Re=100Re = 100): spatial OOD generalization. Training domain: x∈[0,1]x \in[0, 1]; OOD region: x∈[1,2]x \in[1, 2] (downstream prediction within channel, y∈[0,1]y \in[0, 1]). Solution error (shown): PIKE-EXPM achieves 5.5× lower error than PINN (7.14e-2 vs 3.94e-1); SPIKE-EXPM also outperforms PINN (2.4×, 1.62e-1). Physics residual (Table 11): PIKE-Euler achieves 23× lower residual (1.18e-1 vs 2.76e+0). The discrepancy reflects that minimizing PDE residual does not guarantee matching the true solution in extrapolation regions.†

SystemPINNPIKE-EulerPIKE-RK4PIKE-EXPMSPIKE-EXPM
Wave 2D‡1.14e+001.32e-041.35e+001.38e-015.20e-03
Burgers 2D‡6.19e-031.62e-042.42e-042.79e-042.99e-03
Navier-Stokes3.28e+001.04e-012.42e+001.18e+017.76e+00
NS Lid-Driven‡6.36e+007.88e+006.43e+001.24e+011.67e+01

Table 11. Integrator Ablation: 2D PDEs OOD-Space MSE (xy∈[1,2]xy \in[1, 2])

Proofs

Proofs of Main Propositions

Proof of Proposition 1 (Lie Operator Consistency). By Definition 1, the Koopman operator family satisfies Ktg(x)=g(Ft(x))K^{t}g(x) = g(F^{t}(x)). The infinitesimal generator is defined as the limit [5]: Lg:=lim⁡t→0Ktg−gt=lim⁡t→0g∘Ft−gt\mathcal{L}g := \lim_{t \to0} \frac{K^{t}g-g}{t} = \lim_{t \to0} \frac{g \circ F^{t}-g}{t}. Evaluation at a point x(t)x(t) along a trajectory governed by x˙=f(x)\dot{x} = f(x) yields: Lg(x(t))=lim⁡τ→0g(x(t+τ))−g(x(t))τ=ddtg(x(t))\mathcal{L}g(x(t)) = \lim_{\tau\to0} \frac{g(x(t+\tau))-g(x(t))}{\tau} = \frac{d}{dt}g(x(t)). Since gg is differentiable and x(t)x(t) satisfies x˙=f(x)\dot{x}=f(x), application of the chain rule gives: ddtg(x(t))=∇g(x)⋅dxdt=∇g(x)⋅f(x)\frac{d}{dt}g(x(t)) = \nabla g(x) \cdot\frac{dx}{dt} = \nabla g(x) \cdot f(x), establishing Equation (4). □

Proof of Proposition 2 (Finite-Dimensional Approximation). Suppose gg spans a Koopman-invariant subspace. Then there exists A∈RM×MA \in\mathbb{R}^{M \times M} such that Lgi=∑jAijgj\mathcal{L}g_i = \sum_j A_{ij}g_j for all i∈[M]i \in[M]. By Proposition 1, this condition is equivalent to ∇gi⋅f=∑jAijgj\nabla g_i \cdot f = \sum_j A_{ij}g_j. In vector form, ∇g⋅f=Ag\nabla g \cdot f = Ag. Minimizing Equation (5) yields the optimal finite-dimensional approximation when the invariant subspace condition holds only approximately. □

Proof of Proposition 3 (Sparsity and Polynomial Representation). The proximal gradient step for the L1-regularized objective is: A(k+1)=prox⁡λs∥⋅∥1(A(k)−η∇ALLie)A^{(k+1)} = \operatorname{prox}_{\lambda_s\|\cdot\|_1}(A^{(k)} - \eta\nabla_A\mathcal{L}_{\mathrm{Lie}}), where the proximal operator applies element-wise soft-thresholding: [prox⁡λs(B)]ij=sign⁡(Bij)max⁡(∣Bij∣−λ,0)[\operatorname{prox}_{\lambda_s}(B)]_{ij} = \operatorname{sign}(B_{ij})\max(|B_{ij}|-\lambda,0). This sets entries with magnitude below λ\lambda to zero [20], inducing sparsity in AA. For polynomial observables g=[g1,…,gM]Tg = [g_1,\ldots,g_M]^T, the Koopman generator constraint requires g˙i=∑j=1MAijgj\dot{g}_i = \sum_{j=1}^{M} A_{ij}g_j. Each non-zero AijA_{ij} indicates that polynomial gjg_j appears in the time evolution of gig_i. L1 regularization eliminates spurious terms, retaining only those polynomials essential for representing the dynamics. □

Proofs of Lemmas

Proof of Lemma 1 (Continuous Generator Advantage). (i) The matrix exponential expansion gives K=I+AΔt+12A2Δt2+⋯K = I + A\Delta t + \frac{1}{2}A^2\Delta t^2 + \cdots, so Kij=δij+AijΔt+O(Δt2)K_{ij} = \delta_{ij} + A_{ij}\Delta t + O(\Delta t^2). As Δt→0\Delta t \to0, off-diagonal entries Kij→0K_{ij} \to0 for i≠ji \ne j, while diagonal entries Kii→1K_{ii} \to1. This “identity collapse” obscures the interaction structure. (ii) The continuous dynamics z˙=Az\dot{z} = Az imply g˙i=∑jAijgj\dot{g}_i = \sum_j A_{ij}g_j. The coefficient AijA_{ij} represents the instantaneous contribution of gjg_j to g˙i\dot{g}_i, independent of the discretization timescale. L1 regularization on AA thus directly promotes sparse observable interactions. □

Proof of Lemma 2 (Structured Representation via Block Sparsity). From the Koopman generator constraint (Proposition 2), g˙=Ag\dot{g} = Ag expands to block form with library and MLP components. With Wlib=IW_{\mathrm{lib}} = I, the library embedding reduces to glib=ψd(u)g_{\mathrm{lib}} = \psi_d(u), and the constraint Alib−mlp=0A_{\mathrm{lib-mlp}} = 0 decouples the library dynamics: ψ˙i=∑j[Alib−lib]ijψj\dot{\psi}_i = \sum_j [A_{\mathrm{lib-lib}}]_{ij}\psi_j. For polynomial systems where u˙=f(u)\dot{u} = f(u) with ff polynomial of degree pp, the time derivative ψ˙i=∇ψi⋅f(u)\dot{\psi}_i = \nabla\psi_i \cdot f(u) is itself a polynomial in uu. The matrix Alib−libA_{\mathrm{lib-lib}} encodes these coefficients. L1 regularization promotes sparsity, recovering the minimal polynomial representation. □

OOD Generalization Bound

Lemma 3 (Spectral Stability Criterion). The learned dynamics z˙=Az\dot{z} = Az are asymptotically stable if and only if all eigenvalues of AA have strictly negative real parts: Re⁡(λi(A))<0\operatorname{Re}(\lambda_i(A)) < 0 for all ii. For marginal stability (bounded but non-decaying solutions), Re⁡(λi(A))≤0\operatorname{Re}(\lambda_i(A)) \le0 with no repeated eigenvalues on the imaginary axis.

Proof. Standard result from linear systems theory. The solution z(t)=eAtz(0)z(t) = e^{At}z(0) can be expressed via eigendecomposition. For eigenvalue λ=α+iβ\lambda= \alpha+ i\beta, the corresponding mode evolves as eλt=eαt(cos⁡βt+isin⁡βt)e^{\lambda t} = e^{\alpha t}(\cos\beta t + i\sin\beta t). The magnitude ∣eλt∣=eαt|e^{\lambda t}| = e^{\alpha t} decays iff α<0\alpha< 0. For the full system, ∥eAt∥≤Ceρ0t\|e^{At}\| \le Ce^{\rho_0t} where ρ0=max⁡iRe⁡(λi)\rho_0 = \max_i \operatorname{Re}(\lambda_i). □

:::

Lemma 4 (Error Propagation under Koopman Dynamics). Let z(t)z(t) satisfy the perturbed dynamics z˙=Az+r(t)\dot{z} = Az + r(t) where ∥r(t)∥≤ϵK\|r(t)\| \le\epsilon_K is the Koopman consistency residual. Let z~(t)=eA(t−T)z(T)\tilde{z}(t) = e^{A(t-T)}z(T) be the unperturbed Koopman prediction. Then for t≥Tt \ge T:

∥z(t)−z~(t)∥≤ϵKρ0(eρ0(t−T)−1)(19)\|z(t)-\tilde{z}(t)\| \le\frac{\epsilon_K}{\rho_0}\left(e^{\rho_0(t-T)}-1\right) \tag*{(19)}

where ρ0=max⁡iRe⁡(λi(A))\rho_0 = \max_i \operatorname{Re}(\lambda_i(A)) is the spectral abscissa.

Proof. Define the error e(t)=z(t)−z~(t)e(t) = z(t)-\tilde{z}(t). Since z˙=Az+r\dot{z} = Az+r and z~˙=Az~\dot{\tilde{z}} = A\tilde{z}, it follows that e˙=Ae+r(t)\dot{e} = Ae+r(t) with e(T)=0e(T)=0. By variation of constants:*

e(t)=∫TteA(t−s)r(s) ds(20)e(t) = \int_T^t e^{A(t-s)}r(s)\,ds \tag*{(20)}

Taking norms and using ∥eAτ∥≤eρ0τ\|e^{A\tau}\| \le e^{\rho_0\tau}:

∥e(t)∥≤∫Tteρ0(t−s)∥r(s)∥ ds≤ϵK∫Tteρ0(t−s) ds=ϵKρ0(eρ0(t−T)−1)(21)\|e(t)\| \le\int_T^t e^{\rho_0(t-s)}\|r(s)\|\,ds \le\epsilon_K\int_T^t e^{\rho_0(t-s)}\,ds = \frac{\epsilon_K}{\rho_0}\left(e^{\rho_0(t-T)}-1\right) \tag*{(21)}

For ρ0≤0\rho_0 \le0 (stable systems), the bound simplifies to ∥e(t)∥≤ϵK(t−T)\|e(t)\| \le\epsilon_K(t-T). □

Proof of Proposition 4 (Out-of-Distribution Generalization Bound). Let z(t)=g(uθ(x,t))z(t)=g(u_\theta(x,t)) denote the observable trajectory from the PINN solution. On the training domain [0,T][0,T], the Koopman loss ensures ∥z˙−Az∥L2≤ϵK\|\dot{z}-Az\|_{L^2}\leq\epsilon_K, so the dynamics can be written as z˙=Az+r(t)\dot{z}=Az+r(t) where r(t)r(t) is the residual.

For extrapolation t∈[T,T+δ]t\in[T,T+\delta], the Koopman-predicted trajectory is z~(t)=eA(t−T)z(T)\tilde{z}(t)=e^{A(t-T)}z(T). By Lemma 4: ∥z(t)−z~(t)∥≤ϵKρ0(eρ0δ−1)\|z(t)-\tilde{z}(t)\|\leq\frac{\epsilon_K}{\rho_0}(e^{\rho_0\delta}-1). The Lipschitz decoder assumption gives ∥uθ−u~∥≤Lg∥z−z~∥\|u_\theta-\tilde{u}\|\leq L_g\|z-\tilde{z}\|, where u~\tilde{u} is the Koopman-extrapolated solution.

The total error decomposes as ∥uθ−u∗∥≤∥uθ−u~∥+∥u~−u∗∥\|u_\theta-u^*\|\leq\|u_\theta-\tilde{u}\|+\|\tilde{u}-u^*\|. The first term is bounded by Lg⋅ϵKρ0(eρ0δ−1)L_g\cdot\frac{\epsilon_K}{\rho_0}(e^{\rho_0\delta}-1); the second depends on training accuracy ϵtrain\epsilon_{\mathrm{train}}. Combining yields Equation (6). □

Corollary 1 (Spatial Generalization). Under the conditions of Proposition 4, for spatial extrapolation to x∈Ωext\Ωtrainx\in\Omega_{\mathrm{ext}\backslash\Omega_{\mathrm{train}}} with distance dxd_x from the training boundary, the OOD error satisfies:

∥uθ(x,t)−u∗(x,t)∥≤Lg⋅ϵKρ0(eρ0δ−1)+Lx⋅dx⋅CPDE(22)\|u_\theta(x,t)-u^*(x,t)\|\leq L_g\cdot\frac{\epsilon_K}{\rho_0}(e^{\rho_0\delta}-1)+L_x\cdot d_x\cdot C_{\mathrm{PDE}} \tag*{(22)}

where LxL_x is the spatial Lipschitz constant of the observable and CPDEC_{\mathrm{PDE}} captures PDE-specific regularity.

Proof. For spatial extrapolation, the error decomposes into temporal and spatial components. At a point x∈Ωext\Ωtrainx\in\Omega_{\mathrm{ext}\backslash\Omega_{\mathrm{train}}}, let x0∈∂Ωtrainx_0\in\partial\Omega_{\mathrm{train}} be the nearest training boundary point.

The temporal error bound from Proposition 4 applies at x0x_0 since the training domain covers this location. By spatial invariance of AA, the same Koopman dynamics govern evolution at xx, so the temporal contribution remains Lg⋅ϵKρ0(eρ0δ−1)L_g\cdot\frac{\epsilon_K}{\rho_0}(e^{\rho_0\delta}-1).

For spatial deviation, the observable Lipschitz condition gives ∥g(u(x))−g(u(x0))∥≤Lx∥x−x0∥=Lx⋅dx\|g(u(x))-g(u(x_0))\|\leq L_x\|x-x_0\|=L_x\cdot d_x. The PDE regularity constant CPDEC_{\mathrm{PDE}} accounts for how spatial errors propagate through the differential operator: for elliptic PDEs this follows from standard regularity theory; for parabolic PDEs the heat kernel provides decay estimates. Combining yields the stated bound. □

References

  1. [1]Sifan Wang, Xinling Yu, and Paris Perdikaris. When and why PINNs fail to train: A neutral tangent kernel perspective. Journal of Computational Physics, 449:110768, 2022. doi: 10.1016/j.jcp.2021.110768. URL https://www.sciencedirect.com/science/article/pii/S002199912100663X.DOI
  2. [2]Maziar Raissi, Paris Perdikaris, and George E Karniadakis. Physics-informed neural networks: A deep learning framework for solving forward and inverse problems involving nonlinear partial differential equations. Journal of Computational Physics, 378:686–707, 2019.DOI
  3. [3]Sifan Wang, Yujun Teng, and Paris Perdikaris. Understanding and mitigating gradient flow pathologies in physics-informed neural networks. SIAM Journal on Scientific Computing, 43(5): A3055–A3081, 2021.
  4. [4]Bernard O Koopman. Hamiltonian systems and transformation in Hilbert space. Proceedings of the National Academy of Sciences, 17(5):315–318, 1931.pubmed.ncbi.nlm.nih.gov/16577368
  5. [5]Steven L. Brunton, Marko Budišić, Eurika Kaiser, and J. Nathan Kutz. Modern Koopman Theory for Dynamical Systems. SIAM Review, 64(2):229–340, 2022. doi: 10.1137/21M1401243. URL https://epubs.siam.org/doi/10.1137/21M1401243.DOI
  6. [6]Yi Ma, Doris Tsao, and Heung-Yeung Shum. On the principles of parsimony and self-consistency for the emergence of intelligence. Frontiers of Information Technology & Electronic Engineering, 23(9):1298–1323, 2022.arxiv.org/abs/2207.04630
  7. [7]Maziar Raissi, Alireza Yazdani, and George Em Karniadakis. Hidden fluid mechanics: Learning velocity and pressure fields from flow visualizations. Science, 367(6481):1026–1030, 2020.DOI
  8. [8]John M. Hanna, Hugues Talbot, and Irene E. Vignon-Clementel. Improved Physics-informed neural networks loss function regularization with a variance-based term. arXiv preprint arXiv:2412.13993, 2025. URL https://arxiv.org/abs/2412.13993. Proposes ℒ = α · MSE + (1 − α) · std(e) to reduce localized error spikes.
  9. [9]Hangwei Zhang, Zhimu Huang, and Yan Wang. AC-PKAN: Attention-Enhanced and Chebyshev Polynomial-Based Physics-Informed Kolmogorov-Arnold Networks. arXiv preprint arXiv:2505.08687, 2025. URL https://arxiv.org/abs/2505.08687. Combines Chebyshev KAN with Residual Gradient Attention for PINN training stability.
  10. [10]Zongyi Li, Nikola Kovachki, Kamyar Azizzadenesheli, Burigede Liu, Kaushik Bhattacharya, Andrew Stuart, and Anima Anandkumar. Fourier Neural Operator for Parametric Partial Differential Equations. arXiv preprint arXiv:2010.08895, 2021. URL https://arxiv.org/abs/2010.08895. Published at ICLR 2021.
  11. [11]Lu Lu, Pengzhan Jin, Guofei Pang, Zhongqiang Zhang, and George Em Karniadakis. Learning nonlinear operators via DeepONet based on the universal approximation theorem of operators. Nature Machine Intelligence, 3(3):218–229, 2021.
  12. [12]Jiahao Zhang, Chenyang Zhang, Yuxiang Li, et al. Dual-branch neural operator for enhanced out-of-distribution generalization. Engineering Analysis with Boundary Elements, 169:105986, 2024. doi: 10.1016/j.enganabound.2024.105986.
  13. [13]Hao Wu et al. Temporal neural operator for modeling time-dependent physical phenomena. Scientific Reports, 15:16922, 2025. doi: 10.1038/s41598-025-16922-5.
  14. [14]Clarence W Rowley, Igor Mezić, Shervin Bagheri, Philipp Schlatter, and Dan S Henningson. Spectral analysis of nonlinear flows. Journal of Fluid Mechanics, 641:115–127, 2009.DOI
  15. [15]Matthew O Williams, Ioannis G Kevrekidis, and Clarence W Rowley. A data-driven approximation of the Koopman operator: Extending dynamic mode decomposition. Journal of Nonlinear Science, 25(6):1307–1346, 2015.
  16. [16]Bethany Lusch, J Nathan Kutz, and Steven L Brunton. Deep learning for universal linear embeddings of nonlinear dynamics. Nature Communications, 9(1):4950, 2018.DOI
  17. [17]Omri Azencot, N Benjamin Erichson, Vanessa Lin, and Michael W Mahoney. Forecasting sequential data using consistent Koopman autoencoders. In International Conference on Machine Learning, pages 475–485. PMLR, 2020.arxiv.org/abs/2003.02236
  18. [18]Yuying Liu, Aleksei Sholokhov, Hassan Mansour, and Saleh Nabi. Physics-Informed Koopman Network. arXiv preprint arXiv:2211.09419, 2022. URL https://arxiv.org/abs/2211.09419.
  19. [19]Ido Cohen. Koopman Regularization. arXiv preprint arXiv:2403.11302, 2025. URL https://arxiv.org/abs/2403.11302. Proves minimal set theorem: N-dim system has exactly N functionally independent Koopman eigenfunctions.
  20. [20]Steven L Brunton, Joshua L Proctor, and J Nathan Kutz. Discovering governing equations from data by sparse identification of nonlinear dynamics. Proceedings of the National Academy of Sciences, 113(15):3932–3937, 2016.
  21. [21]Samuel H Rudy, Steven L Brunton, Joshua L Proctor, and J Nathan Kutz. Data-driven discovery of partial differential equations. Science Advances, 3(4):e1602614, 2017.
  22. [22]Kathleen Champion, Bethany Lusch, J Nathan Kutz, and Steven L Brunton. Data-driven discovery of coordinates and governing equations. Proceedings of the National Academy of Sciences, 116(45):22445–22451, 2019.DOI
  23. [23]Sheng Li et al. Mesh-free sparse identification of nonlinear dynamics. arXiv preprint arXiv:2505.16058, 2025. URL https://arxiv.org/abs/2505.16058.
  24. [24]Alexia Jolicoeur-Martineau, Kilian Fatras, and Tal Kachman. Training Stiff Neural Ordinary Differential Equations with Explicit Exponential Integration Methods. arXiv preprint arXiv:2412.01181, 2024. URL https://arxiv.org/abs/2412.01181.
  25. [25]Hayden Schaeffer and Linan Zhang. Structure-Preserving Neural Ordinary Differential Equations for Stiff Systems. arXiv preprint arXiv:2503.01775, 2025. URL https://arxiv.org/abs/2503.01775.
  26. [26]Jesús Garrido, Eduardo Ros, and Richard R. Carrillo. Performance Evaluation of Single-step Explicit Exponential Integration Methods on Stiff Ordinary Differential Equations. arXiv preprint arXiv:2411.19374, 2024. URL https://arxiv.org/abs/2411.19374.
  27. [27]Huan Tang, Liang Li, and Hao Wu. GN-SINDy: Greedy Sampling Neural Network in Sparse Identification of Nonlinear Partial Differential Equations. arXiv preprint arXiv:2405.08613, 2024. URL https://arxiv.org/abs/2405.08613.

Paper details

vlnc:W_vkw3jfgf
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Computer Science

Not recorded for this paper

View statement
Looking for related papers…
Contents