IKNO
A graph neural operator model based on isotropic kernels that strictly satisfies SO(3) equivariance at the architectural level. It enables rotation-robust dynamics modeling for three-dimensional physical systems such as the Navier-Stokes equations, with approximately 1/5.6 the parameter count of Tensor Field Networks.
Timeline (last 90 days)
The method proposed in the paper is named the Invariant Kernel Neural Operator with Conditioned Isotropy (IKNO), and it is a compact graph model.
IKNO constructs local interactions using scalar quantities that remain invariant under rotation and vector directions that rotate together with the data.
IKNO achieves strict SO(3) equivariance within numerical precision, rather than approximate rotational invariance.
IKNO matches Tensor Field Networks (TFN) in prediction accuracy while having only approximately 1/5.6 the number of parameters of TFN.
Compared to graph simulators with a comparable number of parameters, IKNO performs better.
IKNO retains only the necessary invariant scalars and l=1 vector features, bypassing the overhead of higher-order tensors while achieving equivalent equivariance guarantees.
The paper's validation is currently focused on three-dimensional Navier–Stokes dynamics scenarios, and whether it can be generalized to a broader range of vector-valued PDE problems remains to be investigated in future research.
All Facts (7)
The method proposed in the paper is named the Invariant Kernel Neural Operator with Conditioned Isotropy (IKNO), and it is a compact graph model.
50%UnverifiedIKNO constructs local interactions using scalar quantities that remain invariant under rotation and vector directions that rotate together with the data.
50%UnverifiedIKNO achieves strict SO(3) equivariance within numerical precision, rather than approximate rotational invariance.
50%UnverifiedIKNO matches Tensor Field Networks (TFN) in prediction accuracy while having only approximately 1/5.6 the number of parameters of TFN.
50%UnverifiedCompared to graph simulators with a comparable number of parameters, IKNO performs better.
50%UnverifiedIKNO retains only the necessary invariant scalars and l=1 vector features, bypassing the overhead of higher-order tensors while achieving equivalent equivariance guarantees.
50%UnverifiedThe paper's validation is currently focused on three-dimensional Navier–Stokes dynamics scenarios, and whether it can be generalized to a broader range of vector-valued PDE problems remains to be investigated in future research.
50%