Reproduce “Sparse POD Mode Selection and Manifold Dimensionality Reduction with Neural Networks” on OpenResearch

About this paper

Tomoki Koike, Prakash Mohan, Marc T. Henry de Frahan, Elizabeth Qian, Julie Bessac

SparseModesNet integrates Proper Orthogonal Decomposition (POD) modes with a residual neural network decoder and a LassoNet-inspired sparse selection mechanism, enabling accurate and interpretable dimensionality reduction for high-dimensional dynamical systems. It reduced reconstruction error by 51-78% on turbulent channel flow data compared to existing polynomial manifold methods, utilizing a minimal set of selected modes.

Key insights

• Jointly optimizing mode selection and nonlinear mapping through hierarchical sparsity regularization can lead to superior reconstruction accuracy and more interpretable reduced models than methods that decouple these steps.
• A residual decoder architecture with a specialized neural network (Π-Net) and orthogonality constraints allows for highly expressive nonlinear approximations while maintaining a physically meaningful basis of POD modes.
• For complex, chaotic, or advection-dominated systems, the expressivity of the nonlinear manifold mapping is crucial, sometimes overriding the impact of precise mode selection, particularly when the underlying dynamics are highly nonlinear.

Summary by alphaXiv.