Solving and Learning Navier-Stokes Dynamics with the Neural Basis Method
Abstract
Physics-informed machine learning has become a major direction in predictive modeling of fluid dynamics, combining governing physics with learned representations. Yet prevailing formulations neither guarantee that small PDE-residual losses imply small solution errors nor scale effectively with increasing problem size. Here we present the neural basis method (NBM) as a new approach for Navier--Stokes dynamics. NBM is built on a ``representation-then-projection'' principle: fixed neural bases define the representation space, while projection of the governing equations determines the expansion coefficients. The same bases define common coordinates for the parametric solution manifold, recasting operator learning as a parameter-to-coefficient mapping and enabling scalability by removing explicit spatial dependence. For incompressible flow, NBM yields a velocity-only formulation whose residual provides an a posteriori velocity-error estimate, establishing a certified, data-free training objective for coefficient learning. Theoretical analysis and numerical experiments demonstrate the accuracy, efficiency, generalization, and scalability of NBM, with substantial improvements over baselines.
The following analysis is generated by AI and may contain errors. It is separate from the authors’ abstract and is not peer review.
Research summary
The paper presents a rigorous and novel framework, the Neural Basis Method, that combines fixed neural bases with projection-based PDE solving to achieve certified, scalable, and accurate solutions for Navier-Stokes dynamics. By establishing a direct link between PDE residuals and solution errors, it addresses a critical limitation in physics-informed machine learning, offering a robust alternative to soft-penalty approaches.
Methodology
The paper introduces the Neural Basis Method (NBM) for solving incompressible Navier-Stokes equations. The core innovation is a "representation-then-projection" framework where fixed neural networks generate divergence-free basis functions, and the PDE solution is represented as a linear combination of these bases. By projecting the governing equations onto this fixed basis space, the problem reduces to solving for expansion coefficients. A key methodological contribution is the derivation of a velocity-only formulation by eliminating the pressure term via a discrete Leray projection, which simplifies the system and allows for a certified, data-free training objective based on the PDE residual. The authors provide a rigorous convergence analysis decomposing error into neural approximation, discretization, and algebraic solution components, and prove that the residual provides an a posteriori error estimate.
Results & evaluation
The experiments validate NBM on analytical benchmarks, boundary-driven flows, and shear-layer dynamics. The results demonstrate high accuracy, with sub-percent velocity errors in parametric settings, significantly outperforming baseline neural operators. The paper reports substantial speedups (orders of magnitude) in both training and inference compared to baselines and traditional solvers like OpenFOAM. The residual-based loss is shown to align the training landscape with velocity error, providing a mechanistic explanation for the performance gains.
Reproducibility
The paper provides detailed mathematical formulations and theoretical proofs. However, specific implementation details, hyperparameters, and code availability are not explicitly provided in the truncated text, which may limit immediate reproducibility without access to the supplementary materials or code repository.
Limitations
The method is primarily demonstrated on 2D incompressible Newtonian flows. Extension to 3D and compressible or non-Newtonian flows is mentioned but not fully detailed. The computational cost of generating and storing the neural basis functions could be a bottleneck for very high-dimensional parameter spaces. The reliance on fixed bases may limit adaptability to highly complex, multi-scale features compared to fully adaptive neural operators.
Broader impact
This work bridges the gap between classical numerical methods and deep learning, offering a rigorous, physics-consistent framework for PDE solving. The certified error estimates and scalability benefits make it highly relevant for scientific computing applications requiring reliable and efficient surrogate models, such as uncertainty quantification and design optimization. The paper presents a rigorous and novel framework, the Neural Basis Method, that combines fixed neural bases with projection-based PDE solving to achieve certified, scalable, and accurate solutions for Navier-Stokes dynamics. By establishing a direct link between PDE residuals and solution errors, it addresses a critical limitation in physics-informed machine learning, offering a robust alternative to soft-penalty approaches.
Author of this paper? Send a correction.