🔢 mathematics

An inertial minimal-deformation-rate framework for shape optimization

This paper proposes a robust numerical framework that couples a second-order inertial flow with a minimal-deformation-rate mesh motion strategy and surface-diffusion regularization to accelerate convergence and preserve mesh quality in PDE-constrained shape optimization and Willmore-driven surface hole filling.

Original authors: Falai Chen, Buyang Li, Jiajie Li, Rong Tang

Published 2026-02-02

📖 5 min read🧠 Deep dive

CC0 1.0

Original authors: Falai Chen, Buyang Li, Jiajie Li, Rong Tang

Original paper dedicated to the public domain under CC0 1.0 (http://creativecommons.org/publicdomain/zero/1.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to sculpt a piece of clay to match a specific, perfect shape. In the world of engineering and science, this "clay" is a mathematical shape (like the wing of a plane or a blood vessel), and the "perfect shape" is the one that works best for a specific job, like reducing wind resistance or carrying electricity efficiently.

This paper presents a new, smarter way to sculpt that clay. It solves two major headaches that scientists have faced for a long time:

Getting stuck: Sometimes, the clay gets stuck in a "flat spot" where it thinks it's done, but it's actually still far from the perfect shape.
The clay tearing: As you push and pull the clay to change its shape, the mesh (the grid of tiny triangles used to model the clay) can get stretched, twisted, or tangled until it breaks, forcing you to start over.

Here is how the authors' new method works, using simple analogies:

1. The "Momentum" Trick (Inertial Flow)

The Problem: Imagine you are pushing a heavy shopping cart down a hill. If you stop pushing exactly when the cart hits a flat patch, it stops immediately. In math, this is like a "first-order" method: it stops moving as soon as the slope looks flat, even if it's just a tiny bump before the real valley. It gets stuck in "local minima" (small, sub-optimal dips) and never finds the best solution.

The Solution: The authors add inertia (momentum) to the process. Think of it like a skateboarder. Even if the ground flattens out for a second, the skateboarder keeps rolling because they have built-up speed.

How it works: The math gives the shape a "push" based on its past movement. Even when the shape seems to have stopped improving, this momentum carries it over the flat spots and small bumps, helping it roll all the way down to the very bottom of the valley (the true best shape).
The Result: The shape finds a better solution much faster and doesn't get stuck halfway.

2. The "Smart Stretch" (Minimal Deformation Rate)

The Problem: When you change the shape of an object, you have to move the "clay" inside it. A standard way to do this is to stretch the clay like a rubber sheet. But if you stretch it too much in one direction, the rubber sheet gets thin and tears, or the grid lines get tangled. This usually forces scientists to stop, erase the grid, and draw a new one (remeshing), which is slow and messy.

The Solution: The authors use a strategy called Minimal Deformation Rate (MDR).

The Analogy: Imagine you are rearranging a crowd of people in a room to fit a new shape. A "dumb" way is to just pull everyone toward the new wall, causing people in the middle to get crushed or stretched. The "smart" way (MDR) is to move everyone just enough to fit the new wall, but slide them around each other so that no one gets squished. The crowd changes shape, but the distance between neighbors stays comfortable.
The Result: The internal grid (the mesh) stays neat and high-quality throughout the whole process. The computer never has to stop and redraw the grid, saving time and preventing errors.

3. The "Smoothing Cream" (Surface Diffusion)

The Problem: Sometimes, you start with a very rough, jagged piece of clay (like a box with sharp corners). If you try to smooth it out, the sharp corners can cause the math to crash or the grid to tangle immediately.

The Solution: The authors add a "smoothing cream" called surface diffusion.

The Analogy: Think of it like applying a gentle heat to the clay. It doesn't change the overall size of the object, but it melts away the sharp, dangerous spikes and smooths out the bumps.
The Result: This allows the computer to start with rough, ugly shapes (even ones that shouldn't work mathematically) and gently smooth them out so the main sculpting process can begin without crashing.

4. The "Patch Job" (Hole Filling)

The authors also applied this to fixing holes in surfaces.
Imagine you have a car body with a hole in it, and you need to fill it in so it looks perfectly smooth.

The Old Way: You might just fill the hole so the edges touch (like taping a piece of paper over a hole). It fits, but if you run your hand over it, you feel a sharp edge where the paper meets the car.
The New Way: The authors use a method that ensures the new patch doesn't just touch the car; it flows into it perfectly. The angle of the new patch matches the angle of the car exactly.
The Result: The hole is filled with a patch that is mathematically smooth and seamless, and the grid used to build that patch remains perfect, even if the starting guess for the patch was messy.

Summary of Results

In their computer experiments, this new "Inertial MDR" framework:

Finds better shapes (lower "cost" or better performance) than old methods.
Gets there faster because it doesn't get stuck on flat spots.
Keeps the grid clean without needing to stop and redraw it.
Works on rough, jagged starting shapes that would usually cause other methods to fail.

In short, they built a sculpting robot that has momentum to keep going, a smart hand to stretch the material without tearing it, and a smoothing tool to handle rough starts, allowing it to create perfect shapes from imperfect beginnings.

Technical Summary: An Inertial Minimal-Deformation-Rate Framework for Shape Optimization

Problem Statement
The paper addresses two persistent challenges in PDE-constrained shape optimization and geometric surface reconstruction:

Slow Convergence and Premature Stagnation: Standard first-order shape gradient flows often converge slowly, particularly in flat energy landscapes where the shape gradient vanishes, leading to stagnation at suboptimal configurations.
Mesh Deterioration: During geometric evolution, standard interior velocity extensions (e.g., harmonic extension) frequently cause severe mesh distortion or element entanglement, necessitating frequent remeshing. This issue is exacerbated when dealing with non-smooth (e.g., cornered) or non-convex initial geometries, which can induce solver failures.

Additionally, the paper addresses the specific challenge of Willmore-driven surface hole filling, where standard mean curvature flow often fails to enforce $G^1$ continuity (tangent plane alignment), resulting in visually discontinuous transitions.

Methodology
The authors propose a unified numerical framework that couples three distinct components:

Inertial Second-Order Flow:
Instead of a standard first-order descent, the framework employs a second-order evolution equation for the domain velocity $\tilde{w}$ . This introduces a "kinetic memory" term (inertia) scaled by $\epsilon_0$ and a time-dependent damping term $\eta(t)$ . The governing equation is:
$\epsilon_0 a(\partial^\bullet_{tt} \tilde{X}, \partial^\bullet_t \tilde{X}, v) + \eta(t) b(\partial^\bullet_t \tilde{X}, v) = -dJ(\Gamma(t); v)$
This formulation allows the system to traverse flat regions and escape shallow local minima by utilizing accumulated momentum, while the damping term ensures energy dissipation and suppresses oscillations. The mechanical energy $H(t)$ , comprising the objective functional and kinetic energy, is proven to satisfy a dissipation property ( $dH/dt \leq 0$ ).
Minimal Deformation Rate (MDR) Mesh Motion:
To preserve mesh quality without remeshing, the framework decouples the geometric shape evolution from the mesh velocity extension. The normal component of the boundary velocity is fixed by the inertial flow. The tangential boundary motion and the extension into the interior domain are determined by minimizing the instantaneous deformation-rate energy:
$\min_{w} \frac{1}{2} \int_{\Omega(t)} |\varepsilon(w)|^2 \quad \text{subject to} \quad w \cdot n|_{\Gamma(t)} = \tilde{w} \cdot n$
where $\varepsilon(w)$ is the symmetric strain-rate tensor. This constrained minimization yields a linear variational problem that maintains high mesh quality even under large deformations, outperforming standard harmonic extensions.
Surface Diffusion Regularization (BGN Framework):
To handle non-smooth or non-convex initial geometries, the normal velocity is regularized by adding a surface diffusion term. This is implemented via the Barrett–Garcke–Nürnberg (BGN) finite element framework. The modified normal velocity includes a term proportional to the surface Laplacian of the mean curvature ( $\Delta_\Gamma H$ ). The regularization parameter $\alpha$ is adapted dynamically based on the $L^2$ -norm of the inertial velocity, ensuring strong smoothing in early stages to resolve singularities while diminishing as equilibrium is approached.

Key Contributions

Inertial-MDR Coupling: The paper presents the first integration of second-order inertial dynamics with the Minimal Deformation Rate mesh motion strategy for PDE-constrained shape optimization. This coupling accelerates convergence while simultaneously preserving mesh quality, eliminating the need for remeshing.
Robustness via BGN Regularization: The work introduces surface-diffusion regularization within the BGN framework for PDE-constrained shape optimization. This approach effectively smooths geometric singularities (corners, reentrant boundaries) and preserves enclosed volume, enabling the optimization of non-smooth and non-convex initial domains that typically cause failure in standard methods.
Extension to Willmore Hole Filling: The inertial MDR framework is extended to Willmore-driven surface hole filling. Unlike parametric approaches, this method uses a second-order inertial formulation to explicitly promote $G^1$ continuity (tangent plane alignment) and high-order geometric regularity. It handles incompatible initial data (where the initial conormal does not match the exterior surface) by weakly enforcing boundary conditions, allowing the conormal to relax naturally during evolution.
Linear Discrete Schemes: The authors derive fully discrete, linear numerical schemes for both the inertial flow and the MDR extension. These schemes are computationally efficient, requiring only a single linear solver per time step, and are proven to be uniquely solvable via the Lax–Milgram lemma.

Numerical Results
The framework was validated on four benchmark problems:

Shape Reconstruction: In both 2D and 3D, the inertial BGN-MDR method achieved significantly lower objective values (e.g., $10^{-7}$ vs. $10^{-3}$ for classical $H^1$ flow) and converged faster. It successfully reconstructed smooth shapes from non-convex and non-smooth initial domains (e.g., L-shaped regions) where classical flows stagnated.
Stokes Drag Minimization: The method successfully optimized obstacle shapes to develop sharp leading/trailing tips while maintaining smooth surfaces elsewhere. It preserved mesh quality in 3D simulations where classical methods produced distorted meshes.
Eigenvalue Minimization: For minimizing Laplace eigenvalues under volume constraints, the method produced smooth boundaries and quasi-uniform meshes, avoiding the element clustering and non-smooth interfaces observed in classical $H^1$ gradient flows.
Surface Hole Filling: In reconstructing holes on a car surface with incompatible initial data, the inertial-MDR scheme produced a quasi-uniform triangulation and a smoother surface with better $G^1$ continuity compared to standard BGN schemes, which resulted in distorted elements.

Significance and Claims
The paper claims that the proposed Inertial MDR framework offers a robust, efficient, and mesh-preserving alternative to classical first-order shape optimization methods. Its primary significance lies in:

Accelerated Convergence: By leveraging inertial effects, the method overcomes the stagnation issues inherent in flat energy landscapes.
Mesh Preservation: The MDR strategy ensures that high-quality meshes are maintained throughout the evolution, avoiding the computational cost and complexity of remeshing.
Geometric Robustness: The integration of surface diffusion via the BGN framework allows the method to handle challenging initial geometries (non-smooth, non-convex) that typically cause numerical instability.
Versatility: The framework is applicable to a broad class of problems, including PDE-constrained optimization and geometric variational problems like Willmore flow, providing a unified approach for both shape design and surface reconstruction.

The authors conclude that while the method demonstrates superior performance in numerical experiments, future work is required to provide a rigorous convergence analysis of the fully discrete schemes and to extend the framework to broader classes of geometric flows.