KAIST's EqGINO Fixes the Rotation Blind Spot Breaking 3D Physics AI

Physics simulations are expensive. Whether you are designing a car body, modeling airflow around a turbine blade, or simulating heat transfer through a complex part, the gold standard is still classical numerical solvers like CFD, which can take hours or days on GPU clusters. Neural operators have emerged as a fast alternative, learning to map input geometry directly to output fields like pressure or velocity. But there is a persistent, frustrating flaw: rotate the geometry at test time, and many of these models fall apart.

EqGINO, a new framework from researchers at KAIST, directly attacks this problem. It introduces equivariance (the property that rotating the input produces a correspondingly rotated output) into the Fourier Neural Operator pipeline, without the computational explosion that has made this combination impractical in 3D.

Why Models Break When You Rotate the Input

To understand the problem, you need to understand how most 3D PDE surrogates work. They take a complex geometry, like a car mesh, and learn to predict a physical field over it. The catch is that training data is almost always collected in a canonical orientation. Deep learning surrogates for 3D PDEs often fail to generalize across geometric transformations because they depend heavily on specific coordinate systems. In practice, this means a model trained on upright car geometries will produce garbage predictions if you feed it the same car rotated 90 degrees.

Two major families of models have tried to fix this:

• Equivariant networks (like SE(3)-transformers or tensor field networks) bake in geometric symmetries by construction. They typically rely on local operations in the spatial domain, making the global receptive field -- essential for PDE dynamics -- computationally expensive. PDEs are inherently non-local: pressure at one point depends on conditions far away.
• Fourier Neural Operators (FNOs) solve the global receptive field problem by working in frequency space, using the Fast Fourier Transform to capture long-range dependencies efficiently. FNO draws inspiration from the Green's function method and directly optimizes the kernel within the Fourier frequency domain, which has been demonstrated to be an efficient means of reducing computational cost and performing global convolution. But enforcing 3D rotational equivariance inside an FNO requires spectral group convolutions, whose cost scales catastrophically in 3D.