Rotational Equivariant Graph Neural Networks Via Local Eigenbasis Transformations

Strategic Insights -

Researchers at Technical University Munich have developed a new method for rotational equivariance in graph neural networks that achieves high predictive accuracy while maintaining a smaller computational footprint. The method, which transforms vector features and spatial neighborhood information into local, node-specific vector bases, is designed to address the need for architectural changes in neural networks to guarantee rotational equivariance. In numerical studies, the researchers found that their approach reduced computational costs by up to an order of magnitude, leading to increased inference times by up to an order of magnitude.

Key Takeaways:

  • The researchers introduced a new method for rotational equivariance in graph neural networks that achieves high predictive accuracy while maintaining a smaller computational footprint.
  • The method transforms vector features and spatial neighborhood information into local, node-specific vector bases, which enables the network to receive full physical and geometric information of the neighborhood without costly computations.
  • The approach is designed to address the need for architectural changes in neural networks to guarantee rotational equivariance, which is a common symmetry of partial differential equations, including the Navier-Stokes equations governing fluid phenomena.
  • The method is experimentally investigated in three different scenarios, ranging from an advection problem to incompressible Navier-Stokes flow around an ellipse and to a highly complex transonic cylinder flow scenario.
  • The results show that the method achieves comparable accuracy to state-of-the-art approaches while reducing computational costs by up to an order of magnitude.
  • The method consistently improves upon a data-augmentation baseline, with an error reduction of 25.3%.
  • The approach is 1.6 times faster than the next best model that guarantees equivariance.

Statistics:

  • Computational costs are reduced by up to an order of magnitude.
  • Inference times are increased by up to an order of magnitude.
  • The method achieves comparable accuracy to state-of-the-art approaches.
  • Error reduction is 25.3% compared to the data-augmentation baseline.
  • The method is 1.6 times faster than the next best model that guarantees equivariance.

Sources:

  • NewsRx. Investigators at Technical University Munich (TU Munich) Discuss Findings in Fluids Physics (Rotational Equivariant Graph Neural Networks Via Local Eigenbasis Transformations). Journal of Engineering. October 27, 2025; p 1285.
  • Rotational Equivariant Graph Neural Networks Via Local Eigenbasis Transformations. Physics of Fluids, 2025;37(8).