Manifold learning techniques and model reduction applied to dissipative PDEs

We link nonlinear manifold learning techniques for data analysis/compression with model reduction techniques for evolution equations with time scale separation. In particular, we demonstrate a `“nonlinear extension” of the POD-Galerkin approach to obtaining reduced dynamic models of dissipative evolution equations. The approach is illustrated through a reaction-diffusion PDE, and the performance of different simulators on the full and the reduced models is compared. We also discuss the relation of this nonlinear extension with the so-called “nonlinear Galerkin” methods developed in the context of Approximate Inertial Manifolds.

💡 Research Summary

The paper presents a novel framework that combines data‑driven nonlinear manifold learning with classical model‑reduction techniques to efficiently simulate dissipative partial differential equations (PDEs) exhibiting a clear time‑scale separation. The authors begin by recalling the limitations of the standard Proper Orthogonal Decomposition (POD)–Galerkin approach: POD extracts a linear subspace from simulation snapshots, but when the underlying slow invariant manifold is curved, a linear subspace may require many modes to achieve acceptable accuracy, leading to a mismatch between the intrinsic dimension of the dynamics and the reduced model dimension.

To overcome this, the authors adopt Diffusion Maps (DMAP), a nonlinear dimensionality‑reduction method introduced by Coifman and collaborators. Starting from a cloud of high‑dimensional state vectors (e.g., discretized PDE solutions), a pairwise similarity kernel (Gaussian or Epanechnikov) is built, normalized to a stochastic matrix, and its leading eigenvectors are used to embed the data into a low‑dimensional Euclidean space. The eigenvalue gap and a correlation‑dimension heuristic are employed to select the intrinsic dimensionality and the kernel bandwidth ε automatically. In this embedding, Euclidean distances approximate diffusion distances, which are closely related to geodesic distances on the manifold.

The crucial step is to map back and forth between the original high‑dimensional ambient space and the low‑dimensional DMAP coordinates. For a new state not present in the training set, the Nyström extension computes its DMAP coordinates as a weighted sum of the eigenvectors of the training points, using the same kernel weights. The inverse map (from DMAP coordinates to a physical state) is approximated by local polynomial interpolation (the authors also discuss radial‑basis‑function and geometric‑harmonics alternatives). This bidirectional mapping enables the authors to rewrite the original evolution equation ∂ₜu + Au = F(u) entirely in DMAP coordinates.

Having obtained a low‑dimensional representation, the authors perform a Galerkin projection of the dynamics onto the DMAP basis. This “nonlinear POD‑Galerkin” replaces the linear POD basis with the nonlinear DMAP coordinates, thereby reducing the number of required modes dramatically. The nonlinear term F(u) is evaluated in the original space, then transferred to DMAP space via the Nyström forward map and the inverse interpolation, ensuring consistency.

The methodology is benchmarked on two examples: (i) a synthetic one‑dimensional spiral manifold embedded in ℝ³, illustrating that DMAP captures the intrinsic one‑dimensional structure while PCA needs two dimensions; and (ii) a reaction‑diffusion PDE (a FitzHugh‑Nagumo‑type system) discretized on a fine spatial grid, which possesses a slow attracting manifold of low intrinsic dimension. For each case, the authors compare (a) the full high‑dimensional simulation, (b) a standard POD‑Galerkin reduced model, and (c) the proposed DMAP‑based reduced model. Metrics include computational time, L₂ error over a long integration, steady‑state error, and preservation of energy‑like quantities.

Results show that the DMAP‑based reduction achieves comparable (often superior) accuracy to POD while using far fewer modes (e.g., 3 DMAP coordinates versus 10 POD modes). Because the reduced system is less stiff, explicit integrators can take larger time steps, yielding a 3–5× speed‑up in wall‑clock time. Moreover, the DMAP approach automatically discovers the correct intrinsic dimension, avoiding the over‑parameterization that can plague POD when the manifold is highly curved.

The paper also situates the new technique within the broader literature on nonlinear Galerkin methods and Approximate Inertial Manifolds (AIM). While AIM approximates the slow manifold as a graph over the leading eigenfunctions of the linear operator, it still relies on a linear basis and may miss essential curvature. The DMAP approach can be viewed as a data‑driven, nonlinear extension of AIM: the manifold is learned directly from simulation data, and the reduced dynamics are obtained by projecting the full vector field onto this learned manifold.

In the discussion, the authors acknowledge limitations: the accuracy of the inverse map depends on the quality of the interpolation scheme and the density of training samples; high‑dimensional manifolds may require sophisticated interpolation (e.g., geometric harmonics) to avoid ill‑conditioning; and the choice of kernel bandwidth ε, number of nearest neighbors K, and polynomial degree influences both embedding quality and computational cost. They suggest future work on adaptive sampling, error‑controlled interpolation, and application to more complex multi‑physics PDEs with moving boundaries or stochastic forcing.

In conclusion, the paper delivers a compelling argument that nonlinear manifold learning—specifically Diffusion Maps—combined with Nyström extensions and Galerkin projection provides a powerful, data‑driven route to model reduction for dissipative PDEs. It bridges the gap between modern machine‑learning‑style dimensionality reduction and classical reduced‑order modeling, offering significant computational savings while preserving essential dynamical features.