Abstract
The design of neutral inclusions—objects that do not perturb an external physical field—is a fundamental challenge in material science with applications in cloaking and composite material design. This paper introduces CoCo-PINNs (Conformal mapping Coordinates Physics-Informed Neural Networks), a framework that integrates complex analysis with deep learning. By embedding conformal mapping and Fourier series expansions into the PINN architecture, the authors address the instability and "credibility gap" of traditional neural networks in solving ill-posed inverse problems. The method demonstrates superior accuracy, consistency, and stability across complex, non-convex geometries.
The Challenge of Neutral Inclusions
In material science, a neutral inclusion is an object inserted into a medium that does not perturb the external field, effectively making it "invisible" to sensors or external flows. While mathematically straightforward for simple shapes like circles or ellipses, designing these inclusions for arbitrary, complex shapes is an inverse problem: given the desired effect (zero perturbation), one must find the specific interface conditions or boundary functions that achieve it.
Traditional Physics-Informed Neural Networks (PINNs) often struggle with these problems. Because they are "geometry-blind," they treat all points in space equally and can easily converge to suboptimal solutions that satisfy the loss function mathematically but fail to represent physical reality—a phenomenon the authors call the "credibility gap."
Methodology: Embedding Geometric Intelligence
CoCo-PINNs introduce three critical innovations to the standard PINN framework:
- Conformal Mapping: Utilizing the Riemann mapping theorem, the framework transforms the exterior of a complex shape into a simple circular domain. This provides a natural coordinate system for sampling collocation points, allowing the network to "understand" the boundary geometry.
- Fourier Series Representation: Instead of approximating the interface function directly, the network learns the coefficients of a truncated Fourier series. This constrains the solution to smooth, well-behaved functions, turning an ill-posed problem into an over-determined, stable one.
- Dual-Field Training: The authors prove that an inclusion neutral for two orthogonal fields (e.g., $x$ and $y$ directions) is neutral for all linear fields. Training on two fields simultaneously ensures universal neutrality.
Key Results and Benchmarks
The authors evaluated CoCo-PINNs on four geometric benchmarks: a square, a non-convex "fish" shape, an asymmetrical "kite," and a high-frequency "spike."
- Superior Accuracy: CoCo-PINNs achieved significantly lower errors than classical PINNs. For the "fish" shape, the credibility error was nearly an order of magnitude lower.
- Consistency and Stability: In repeated experiments, CoCo-PINNs converged to the same solution every time, whereas classical PINNs showed high variance. The method remained stable even as conductivity contrasts in the medium changed.
- Versatility: The framework successfully solved both "coated inclusion" problems (where a layer is added to achieve neutrality) and "imperfect interface" problems (where the boundary itself has specific resistance properties).
Conclusion
CoCo-PINNs represent a significant step forward in combining geometric function theory with deep learning. By providing a mathematically grounded way to handle complex boundaries, the framework offers a robust tool for the inverse design of advanced materials. Future work may explore extending these concepts to 3D geometries using methods like Deep Signed Distance Functions (DeepSDF) and addressing the interactions within multi-inclusion composites.
Summary updated