Size is Not the Solution: Deformable Convolutions for Effective Physics Aware Deep Learning
The study introduces Deformable Physics-Aware Recurrent Convolutions (D-PARC) to improve physics-aware deep learning by overcoming the rigidity of standard CNNs. D-PARC uses deformable convolutions inspired by Hybrid Lagrangian-Eulerian methods to autonomously concentrate computational resources in high-strain regions, outperforming larger models and fixed kernels.