Energy-dissipative evolutionary Kolmogorov-Arnold networks for complex PDE systems

Guang Lin, Guang Lin 0001, Changhong Mou, Jiahao Zhang

Abstract

We introduce evolutionary Kolmogorov-Arnold Networks (EvoKAN), a novel framework for solving complex partial differential equations (PDEs). EvoKAN builds on Kolmogorov-Arnold Networks (KANs), where activation functions are spline based and trainable on each edge, offering localized flexibility across multiple scales. Rather than retraining the network repeatedly, EvoKAN encodes only the PDE's initial state during an initial learning phase. The network parameters then evolve numerically, governed by the same PDE, without any additional optimization. By treating these parameters as continuous functions in the relevant coordinates and updating them through time steps, EvoKAN can predict system trajectories over arbitrarily long horizons, a notable challenge for many conventional neural-network based methods. In addition, EvoKAN integrates the scalar auxiliary variable (SAV) method to guarantee unconditional energy stability and computational efficiency. At individual time step, SAV only needs to solve decoupled linear systems with constant coefficients, the implementation is significantly simplified. We test the proposed framework in several complex PDEs, including one dimensional and two dimensional Allen-Cahn equations and two dimensional Navier-Stokes equations. Numerical results show that EvoKAN solutions closely match analytical references and established numerical benchmarks, effectively capturing both phase-field phenomena (Allen-Cahn) and turbulent flows (Navier-Stokes).

Introduction

Solving complex, time-dependent partial differential equations (PDEs) is a cornerstone of scientific computing. While Physics-Informed Neural Networks (PINNs) have gained popularity, they often require global spatio-temporal optimization, leading to computational bottlenecks and difficulties in maintaining temporal causality for long-horizon predictions. Traditional approaches like Evolutionary Deep Neural Networks (EDNNs) attempt to update network weights dynamically but are often limited by the spatial inflexibility of standard Multi-Layer Perceptrons (MLPs), which struggle with sharp gradients and chaotic dynamics.

To address these challenges, we present Evolutionary Kolmogorov-Arnold Networks (EvoKAN). This framework achieves highly accurate, long-horizon physical predictions by training a spline-based network only on a system's initial state and mathematically evolving its parameters forward in time, bypassing the massive computational overhead of global retraining.

Methodology

The EvoKAN framework integrates three core mathematical and architectural innovations:

  1. Kolmogorov-Arnold Networks (KANs): Unlike standard MLPs that use fixed activation functions at nodes, KANs employ learnable, spline-based activation functions on the edges. This architecture provides localized flexibility and smoothness, which is essential for representing complex spatial domains and sharp physical gradients.
  2. Evolutionary Parameter Updating: The network is trained once to map spatial coordinates to the initial condition at $t=0$. Subsequently, the network's parameters ($\mathcal{K}$) are treated as time-dependent continuous functions. At each time step, the algorithm calculates a gradient based on the governing PDE and updates the parameters numerically, allowing the network to "evolve" without further machine learning optimization.
  3. Scalar Auxiliary Variable (SAV) Method: To ensure numerical stability over long horizons, the SAV approach is integrated. This method guarantees unconditional energy stability—a critical requirement for gradient flows—and simplifies the computation by solving decoupled linear equations with constant coefficients at each step.
Schematic of the EvoKAN framework showing the evolution of KAN parameters over time.

Experimental Evaluation

EvoKAN was evaluated against high-accuracy spectral methods on several standard physical benchmarks:

  • Allen-Cahn Equations (1D and 2D): These classical phase-field models describe interface dynamics. In 1D tests, EvoKAN maintained an $L^2$ error of approximately $10^{-4}$ and remained stable even with small interface thickness parameters ($\epsilon$). In 2D, it successfully captured curvature-driven interface shrinkage and domain coarsening.
  • 2D Incompressible Navier-Stokes Equations: In fluid dynamics benchmarks, EvoKAN accurately tracked vortex patterns and pressure distributions. Notably, while standard EDNNs using MLPs failed to reproduce unstable fluid solutions at lower viscosities ($\nu=0.01$), EvoKAN's superior spatial representation allowed it to succeed.

Discussion and Future Directions

The "train once, evolve forever" paradigm of EvoKAN bypasses the need for global optimization, acting as a continuous, differentiable numerical solver. The localized nature of B-splines in KANs proves more effective than standard perceptrons for capturing steep gradients in physical phenomena. Furthermore, the integration of the SAV method provides built-in stability, preventing the accumulation of errors that often plague neural PDE solvers.

Future research will focus on:

  • Adaptive Refinement: Implementing strategies to automatically refine the spline-based architecture in regions with high complexity.
  • Complex Geometries: Extending the framework to handle non-periodic boundary conditions and complex real-world domains.
  • Data Assimilation: Integrating EvoKAN with real-time data and Reduced Order Models (ROMs) for applications like climate modeling.

Summary updated

BibTeX
@article{lin2025energy,
  author = {Guang Lin and Guang Lin and Changhong Mou and Jiahao Zhang},
  title = {{Energy-dissipative evolutionary Kolmogorov-Arnold networks for complex PDE systems}},
  journal = {Journal of Computational Physics},
  year = {2025},
  doi = {10.1016/j.jcp.2025.114326},
}
Figure