A coupled Kolmogorov-Arnold Network and Level-Set framework for evolving interfaces

This paper proposes a physics-informed framework that couples Kolmogorov-Arnold Networks (KANs) with a level-set formulation to efficiently and accurately solve complex moving boundary problems, such as temperature distribution and interface evolution, without requiring experimental data.

The Gist

Imagine you are trying to track the exact moment an ice cube melts in a glass of water. It sounds simple, but if you wanted to write a computer program to predict exactly where the boundary between ice and water is at every microsecond, it becomes a mathematical nightmare.

This paper introduces a new, smarter way for computers to "visualize" and predict these moving boundaries (like melting ice or solidifying metal) using a cutting-edge AI architecture called KANs.

Here is the breakdown of how they did it, using everyday analogies.

1. The Problem: The "Shifting Sand" Dilemma

In physics, problems where a boundary moves (like a melting front) are called Stefan Problems.

Traditional AI models (called MLPs) try to solve this by building a massive, heavy "digital brain" with millions of connections. Imagine trying to map the exact shape of a moving wave by using a giant, heavy, clunky Lego set. Because the wave is always changing, the Lego set has to be enormous to keep up, which makes it slow, expensive, and prone to making mistakes where the water meets the air.

2. The Solution: The "Smart Rubber Band" (KANs)

Instead of using those heavy, rigid "Lego" connections, the researchers used Kolmogorov-Arnold Networks (KANs).

Think of a traditional AI like a series of stiff wooden beams connected by bolts. To change the shape, you have to add more and more beams.

A KAN, however, is like a series of smart, flexible rubber bands. Instead of just being "on" or "off," these rubber bands can stretch, bend, and change their own shape to fit the curve of the wave perfectly. Because the "connections" themselves are flexible, you don't need a massive structure. You can use a tiny, lightweight framework that is much more "expressive" and accurate.

3. The Tool: The "Invisible Ghost Line" (Level-Set Method)

To keep track of where the ice ends and the water begins, they used something called the Level-Set Method.

Imagine you are looking at a mountain range through a thick fog. You can't see the actual ground, but you can see "contour lines" (like on a hiking map) that show you the elevation.

The researchers told the AI to create a "digital map" of these contour lines. The "interface" (the melting edge) is simply the place where the elevation is exactly zero. By tracking this "zero line," the AI can handle complex shapes—like a melting ice cube turning into a puddle—without getting confused when the shape changes or splits apart.

4. The "Physics Teacher" (Physics-Informed Learning)

Usually, AI learns by looking at millions of pictures (data). But in this paper, they didn't give the AI any pictures of melting ice. Instead, they gave it a Physics Textbook.

They created a "Loss Function," which acts like a strict teacher. Every time the AI makes a guess, the teacher checks it against the laws of physics:

• "Does this temperature follow the rules of heat transfer?"
• "Is the melting speed consistent with how much heat is being applied?"

If the AI's guess violates the laws of physics, the "teacher" gives it a failing grade, and the AI adjusts its "rubber bands" until it obeys the laws of nature.

The Result: Small, Fast, and Accurate

The researchers tested this on 1D and 2D models (simple lines and circles). The results were impressive:

• Efficiency: While a traditional AI might need 120,000 "parts" to solve the problem, this KAN model did it with only 640. It’s like winning a heavyweight boxing match while weighing as much as a feather.
• Accuracy: Even though it was tiny, it was incredibly precise at predicting exactly where the melting edge was and how the temperature was changing.

In short: They found a way to teach a very small, very flexible AI to master complex, moving physical boundaries just by teaching it the rules of science, rather than showing it a million examples.

Technical Summary

Technical Summary: A Coupled Kolmogorov-Arnold Network and Level-Set Framework for Evolving Interfaces

1. Problem Statement

The paper addresses the challenge of solving moving-boundary (Stefan-type) problems, which model phase-change processes like melting and solidification. These problems are mathematically complex because they involve coupled transport Partial Differential Equations (PDEs) where the domain itself evolves over time.

The authors identify two primary limitations in existing Physics-Informed Neural Network (PINN) approaches based on Multi-Layer Perceptrons (MLPs):

• Data Dependency: Standard PINNs often require measurement data to stabilize training in nonlinear heat transfer problems.
• Spectral Bias & Complexity: MLPs struggle to resolve high-frequency variations and sharp gradients at interfaces, necessitating deep, computationally expensive architectures that lack interpretability.

2. Methodology (The KANLS Framework)

The authors propose a novel methodology termed KANLS, which integrates Kolmogorov-Arnold Networks (KANs) with the Level-Set (LS) method.

• Architecture (KANs): Unlike MLPs that use fixed activation functions at nodes, KANs use learnable univariate spline-based functions on the network edges. This allows for higher expressive power with significantly shallower and more compact architectures.
• Interface Representation (Level-Set): To handle complex, multi-dimensional geometries (where interfaces might undergo topological changes), the moving boundary $\Gamma(t)$ is represented implicitly by a smooth scalar field $\phi(x, t)$. The interface is defined where $\phi = 0$.
• Physics-Informed Loss Function: The framework is trained by minimizing a composite loss function $L$ that enforces:
  • Governing PDEs: Diffusion equations for both solid ($u_s$) and liquid ($u_\ell$) phases, using smooth masking functions (error functions) to separate the domains.
  • Interface Conditions: Temperature continuity (equilibrium) and the Stefan condition (latent heat balance/velocity).
  • Level-Set Dynamics: An advection equation to evolve the interface and an Eikonal regularization term ($|\nabla\phi| = 1$) to maintain $\phi$ as a signed distance function.
• Adaptive Resampling: To improve accuracy near the moving boundary, the authors implement an interface-focused collocation resampling strategy (Algorithm 1), which concentrates training points in regions where $\phi$ is near zero.

3. Key Contributions

• Hybrid Framework: The first integration of KANs with the Level-Set method for solving Stefan problems.
• Parameter Efficiency: Demonstrates that KANs can achieve high accuracy using a fraction of the parameters required by MLP-based PINNs.
• Data-Free Solving: Proves that the framework can accurately reconstruct complex interface dynamics and temperature fields using only physics-informed residuals, without requiring experimental measurement data.
• Robustness to Geometry: Extends the solution from 1D analytical benchmarks to 2D curved (Frank-type) interfaces.

4. Results

The framework was validated against analytical solutions in 1D and 2D:

• 1D Stefan Problem: The model accurately captured temperature profiles and interface evolution. Notably, the KAN-based model used only 640 parameters, compared to approximately 120,000 parameters required by a comparable MLP architecture.
• 2D Stefan Problem: The model successfully simulated a circular interface expanding in a 2D domain. It accurately tracked the interface radius $R(t) \propto \sqrt{t}$ and resolved sharp thermal gradients.
• Convergence: Error metrics (MAE, MSE, $R^2$) showed that increasing the number of collocation points improves accuracy, with diminishing returns observed after 8,000 points, indicating effective convergence.

5. Significance

This work represents a significant step forward in the application of "next-generation" neural architectures to classical physics. By leveraging the mathematical properties of KANs, the authors have developed a method that is more compact, more interpretable, and more efficient at capturing the sharp discontinuities inherent in phase-change problems. This methodology holds great potential for scaling to more complex, multi-phase, and multi-dimensional physical systems in engineering and materials science.