Neural-POD: A Plug-and-Play Neural Operator Framework for Infinite-Dimensional Functional Nonlinear Proper Orthogonal Decomposition

Changhong Mou, Binghang Lu, Guang LinarXiv 2026 · DOI

TL;DR

This paper introduces Neural-POD, a novel framework that uses neural networks to learn "proper orthogonal decomposition" (POD) modes. This allows for more flexible and robust representations of complex physical systems, overcoming limitations of traditional linear POD, especially when dealing with varying resolutions, parameters, or nonlinear dynamics.

What Problem Does This Paper Address?

Many scientific simulations, particularly in fields like fluid dynamics, generate massive datasets. To make these datasets manageable and enable faster predictions, researchers often use Proper Orthogonal Decomposition (POD) to find a low-dimensional representation. However, traditional POD has significant limitations: its learned "modes" are tied to the specific resolution and parameters of the training data, making them brittle and difficult to generalize to new scenarios without retraining. This "discretization dependence" hinders the development of AI for science models that can reliably transfer across different simulation setups and applications. The authors aim to create a more robust and adaptable POD framework that learns continuous, resolution-invariant bases.

Methods

The core innovation is Neural-POD, which replaces the linear basis functions learned by traditional POD with nonlinear functions parameterized by neural networks. Instead of relying on Singular Value Decomposition (SVD) of snapshot data, Neural-POD iteratively learns these basis functions by minimizing a reconstruction loss between the data and the neural representation. This process is analogous to the Gram-Schmidt orthogonalization procedure, ensuring the learned modes are orthogonal.

Key features of Neural-POD include:

  • Plug-and-Play: The learned neural bases can be easily integrated into existing reduced-order modeling (ROM) frameworks (like Galerkin projection) and operator learning architectures (like DeepONets).
  • Resolution-Independence: The learned bases are continuous functions in function space, allowing them to be evaluated on any grid, regardless of the training resolution.
  • Flexible Optimality: The training loss can be defined in various norms (e.g., L², L¹, H¹), allowing the learned bases to better capture specific types of dynamics (e.g., sharp gradients with L¹).
  • Parameter Generalization: Neural-POD learns a mapping from parameters (like viscosity) to basis functions, enabling generalization to unseen parameter regimes without retraining.

Datasets

The authors use data generated from simulations of:

  • 1D Burgers' equation: Used extensively for training and testing the framework's ability to handle varying viscosity and resolution.
  • 2D Navier-Stokes equations: Used to demonstrate the framework's applicability to more complex, high-dimensional problems.

The paper does not introduce new benchmark datasets but rather applies Neural-POD to existing simulation data for these well-known PDEs.

Key Insights and Ideas

  • Beyond Linearity: The most significant insight is that by moving from linear POD modes to nonlinear neural bases, the framework can capture complex, emergent structures in physical systems that linear methods miss.
  • True Generalization: Neural-POD's ability to learn continuous, parameter-dependent bases is crucial for building AI models that can reliably adapt to different simulation settings without constant retraining. This is a major step towards more generalizable AI for science.
  • Flexibility in Loss Functions: The ability to choose different norms (L¹, L²) for training allows for tailoring the learned bases to specific features of the data, such as sharp gradients, which is not possible with traditional SVD-based POD.
  • Bridging ROM and Operator Learning: Neural-POD acts as a unifying component, providing a robust, interpretable basis for both traditional ROMs and modern operator learning frameworks, enhancing their capabilities.

Key Results

  • Improved Reconstruction: Neural-POD consistently achieved better reconstruction accuracy compared to traditional POD, especially for out-of-distribution parameter values and at coarser resolutions. For example, in the Burgers' equation, Neural-POD showed strong generalization to unseen viscosity values, outperforming POD even when POD was provided with extra snapshots.
  • Resolution Independence Demonstrated: The framework successfully reconstructed spatiotemporal fields for the Burgers' equation at unseen viscosity parameters using different latent dimensions (r=2, 4) and resolutions, showcasing its robustness.
  • Effective Integration: Neural-POD was successfully integrated into both Galerkin ROMs and DeepONet architectures, demonstrating its "plug-and-play" nature and enhancing their performance. For instance, Neural-POD-DeepONet showed lower testing errors than POD-DeepONet at coarser resolutions.
  • L¹ vs. L² Norms: The choice of norm significantly impacted the learned bases. L²-optimal bases were smoother and globally accurate, while L¹-optimal bases better captured localized sharp transitions and nonlinear features.

Limitations and Future Work

  • Computational Cost: Training neural networks can be computationally expensive, although the authors suggest strategies like choosing one network as trainable to mitigate this.
  • High-Dimensional Problems: While demonstrated on 2D Navier-Stokes, the authors suggest extending Neural-POD to more challenging high-dimensional problems like turbulent 3D flows.
  • Physics-Informed Integration: Future work could involve integrating physics-informed neural networks (PINNs) to enforce physical constraints during basis construction, improving interpretability and extrapolation.
  • Data Assimilation and Digital Twins: The framework could be coupled with data assimilation techniques for real-time state estimation and used to build adaptive surrogate models for digital twins.
  • Geometric Variations: Extending the framework to handle geometrical variations would further broaden its applicability.

Summary updated

BibTeX
@misc{mou2026neural,
  author = {Changhong Mou and Binghang Lu and Guang Lin},
  title = {{Neural-POD: A Plug-and-Play Neural Operator Framework for Infinite-Dimensional Functional Nonlinear Proper Orthogonal Decomposition}},
  year = {2026},
  eprint = {2602.15632},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2602.15632},
}
Figure