Introduction
Computing the vibration modes, or eigenfunctions, of physical objects is a cornerstone of engineering and physics. Traditionally, these modes are calculated for a single, fixed geometry. However, in design and optimization, we often need to understand how these modes change as a shape is modified. Shape Space Spectra introduces a novel framework to compute the eigenfunctions of an entire family of shapes simultaneously using neural fields. By treating eigenfunctions as continuous functions of both spatial coordinates and shape parameters, this method enables rapid physical simulations and differentiable shape optimization.
The Challenge of Shape-Space Spectra
In traditional workflows, changing a shape—such as widening a bridge or deepening a guitar body—requires remeshing and re-running a computationally expensive eigendecomposition. This "one-off" approach hinders inverse design (e.g., finding a shape for a specific sound) and real-time interactive simulation.
A significant technical hurdle in representing eigenfunctions across a continuous "shape space" is eigenvalue crossing. As a shape evolves, the ordering of modes (e.g., which is the "first" vs. "second" mode) can swap based on their eigenvalues. Standard training methods that find modes sequentially encounter "kinks" and discontinuities at these crossing points, which break gradient-based optimization.
Methodology: Continuous Neural Eigenfunctions
The authors propose a discretization-agnostic approach using Neural Fields. A Multi-Layer Perceptron (MLP) is trained to take a spatial coordinate ($x$) and a shape parameter ($g$) as input, outputting the corresponding eigenfunction value.
To overcome the eigenvalue crossing problem and ensure smooth transitions across the shape space, three key innovations are introduced:
- Joint Training: All modes are learned simultaneously.
- Shape-Dependent Causal Sorting: Modes are dynamically reordered during training based on current eigenvalues, preserving the topological structure of the eigenfunctions even when their dominance hierarchy shifts.
- Causal Gradient Filtering: A filtering mechanism ensures a strict one-way causal chain of orthogonality, preventing higher-order modes from interfering with the optimization of lower-order ones.
Key Insights
- Smoothness over Dominance: By allowing modes to swap places in the eigenvalue hierarchy, the eigenfunctions remain smooth and analytic across the shape space, which is critical for optimization.
- Differentiability: Representing eigenfunctions via a neural network allows for direct computation of derivatives of physical properties (like pitch) with respect to shape parameters, facilitating gradient-based design.
- Generalization: Once trained, the model can predict vibration modes for unseen shapes within the learned latent space (zero-shot prediction).
Applications and Results
The method was evaluated across diverse datasets, including parametric CAD models, ShapeNet categories, and animal mesh interpolations.
- Simulation Speedup: The method achieved a 32.4x speedup in reduced-space simulations compared to full-space physics, with minimal error (2.4% MSE).
- Efficient Sound Synthesis: For tasks like optimizing bell thickness for specific notes, the method was 40x faster than previous state-of-the-art differentiable solvers, reducing hours of computation to minutes.
- Robotic Locomotion: Shape optimization enabled a walking robot to move 18x faster by using learned eigenfunctions as actuation signals.
Limitations and Future Work
While powerful, the current implementation primarily focuses on Neumann boundary conditions. Future work could integrate techniques like Physics-Informed Neural Networks (PINNs) to handle Dirichlet conditions. Additionally, like many neural field methods, the MLP faces "spectral bias," making it challenging to capture very high-frequency modes. Finally, while inference is nearly instantaneous, the initial training phase for complex shape spaces remains a significant precomputation cost.
Summary updated