[R] Neural PDE solvers built (almost) purely from learned warps

A research team has introduced Flower, a novel neural PDE solver that operates almost entirely through learned coordinate warps, eliminating traditional Fourier layers, attention mechanisms, and most spatial convolutions. The architecture predicts displacements at each spatial location and samples features from these warped coordinates—the only mechanism for spatial interaction. Built on a U-Net scaffold with transformer-inspired components like multiple warp heads and skip connections, Flower achieves linear computational scaling with grid points, making it particularly efficient for 3D problems.

Technically, the model replaces conventional spatial mixing operations (like those in Fourier Neural Operators or Vision Transformers) with pointwise warp predictions. This approach naturally handles phenomena like characteristics in conservation laws and high-frequency wave propagation along rays. In benchmarks from The Well suite, Flower outperforms comparable-scale models including FNOs, convolutional UNets, and ViTs, while using significantly simpler primitives.

The research originated from attempts to create lightweight versions of Fourier Integral Operators that could be stacked as neural network layers. Through iterative refinement, the team eliminated most components, finding that warping alone provided sufficient expressive power for PDE solutions. This work bridges concepts between neural networks and kinetic equations, suggesting new directions for efficient scientific computing architectures.