New Neural Operator Scales 3D Physics Simulations to Million-Node Meshes

A new neural operator, named PGD-NO (Precomputed Geometry Decomposition Neural Operator), has been introduced to address the significant memory consumption and single-node bottlenecks that limit existing neural PDE solvers in engineering simulations. Traditional architectures struggle with high-resolution meshes due to VRAM constraints on a single compute unit. PGD-NO tackles this by shifting the computational burden of geometric encoding to an initial, deterministic pre-computation phase. The core innovation involves an iterative geometry decomposition algorithm that extracts "geometry tokens," effectively separating the process of feature extraction from the actual solution querying. This architectural design allows for linear memory scalability, which is crucial for handling extremely large datasets. The researchers demonstrated that PGD-NO can achieve high-fidelity learning on meshes with over 10 million nodes, a scale previously unachievable for many existing methods due to memory exhaustion. Beyond its scalability, PGD-NO maintains competitive predictive accuracy across various industrial benchmarks. It also offers intrinsic interpretability through its attention mechanisms, providing insights into the model's decision-making process. This advancement represents a robust and efficient solution for next-generation industrial design applications that require large-scale, high-fidelity physics simulations.