Physics enhanced graph transformers improve antenna design prediction accuracy
Physics-Augmented Graph Transformers for Patch-Antenna Forward and Inverse Design
Machine Learning
Summary
Predicting how patch antennas emit signals usually requires slow, detailed simulations. The authors developed a new method that uses graph transformers combined with physics knowledge to predict antenna radiation patterns more efficiently and accurately. Their approach also helps design antennas by reversing the prediction process, generating antenna shapes that produce desired radiation. This method works well on large datasets and can even generalize to new antenna shapes without extra training.
What this means in practice
- •For antenna engineers: Speed up the design and analysis of patch antennas using physics-augmented graph transformers for accurate radiation prediction without costly simulations.
- •For wireless hardware developers: Generate patch antenna designs conditioned on desired signal patterns to optimize wireless device performance faster than traditional methods.$Commercial implications: Enables selling antenna design tools that create efficient patch antennas for devices based on radiation requirements, reducing development time and cost.
Authors
Avi Epstein, Snir Nehemia, Haim Suchowski, Lior Wolf
Abstract
Full-wave electromagnetic (EM) simulation enables accurate patch-antenna analysis but is computationally expensive for large-scale forward prediction and inverse design. We present a mesh-native, physics-augmented graph-learning framework that treats radiation-pattern prediction as signal reconstruction on an irregular surface mesh. For the forward problem, a GPS graph transformer is trained with Physics-Augmented Intermediate Supervision (PAIS), an auxiliary node-level objective that predicts complex surface currents, the physical intermediate linking geometry to radiation. PAIS improves multiple GNN backbones at no inference-time cost, while shuffled-current and non-physical controls show the gain comes from physical correspondence. Direction-conditioned decoding and a differentiable radiation-integral consistency loss further exploit this structure. On an 80,000-sample CST benchmark, GPS+PAIS reaches MSE 0.17 / PSNR 19.67, generalizes to a PCA split, and transfers zero-shot to canonical patches. For inverse design, surrogate-filtered diffusion beats nearest-neighbor retrieval by 32% relative MSE.