Title: Edge-Deployable Graph Neural Operator Enables Real-Time Sensing of Inaccessible Physical Fields

Abstract:

Deriving real-time insights into inaccessible interior physical fields from limited boundary data remains a persistent challenge in scientific machine learning, with significant implications for safety-critical monitoring in various engineering sectors. While current neural operators deliver high precision, they have yet to adequately address the constraints of deploying models on embedded edge platforms. This study presents VIRSO (Virtual Irregular Real-Time Sparse Operator), the inaugural neural operator featuring a distinctive spatial-spectral architecture specifically engineered for hardware deployment at the edge.

VIRSO utilizes a spectral-spatial decomposition that aligns directly with hardware execution characteristics. It learns a nonlinear mapping from sparse, geometrically disjoint boundary inputs to spatially continuous interior multiphysics fields defined on irregular unstructured meshes. This process is split into two independent pathways: a compute-bound graph spectral pathway and a memory-bandwidth-bound spatial-aggregation pathway, both of which are individually evaluated on datacenter and embedded accelerators.

This architectural approach yields a 29-fold reduction in the inference energy-delay product compared to a standard graph-operator baseline, dropping from 206 J·ms to 7.0 J·ms on an NVIDIA H200. Furthermore, it facilitates embedded inference at a rate of 17.0 samples per second on an NVIDIA Jetson Orin Nano, operating at a board-level power consumption of just 7.06 W, with no modifications required.

Additionally, a mesh-density-adaptive graph construction method, termed V-KNN, decreases the number of graph edges by 34% while simultaneously enhancing accuracy. In testing across three benchmarks featuring reconstruction ratios between 47:1 and 156:1, VIRSO attained mean relative $L_2$ errors under 1%, utilizing fewer parameters than traditional operator baselines. It also provides an inference speedup of approximately $10^4$ times over high-fidelity reference solvers. To the best of our knowledge, this work marks the first demonstration of a single-digit-watt neural operator, highlighting hardware co-design as an essential component for operator-based inference and offering a viable route for real-time deployment.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC