2026 URSI GASS · Kraków, Poland · 15–22 August 2026

A Data-Free Patch Antenna Generation Framework via Two-Stage Gradient Exploration and BiScaleNorm

Peng-Yu Chian, Ting-Mao Chen, Kuo-Hung Cheng*, Wei-Wen Wu, Alan Liu, Shih-Cheng Lin, Sheng-Fuh Chang

National Chung Cheng University, Chiayi, Taiwan · *ieelsl@ccu.edu.tw

U.R.S.I.

General Assembly &
Scientific Symposium
Session B10 · Abstract #1679

P - B10-01

1Motivation

Designing microstrip patch antennas for mmWave 6G applications is a high-dimensional, non-convex optimization problem.

Evolutionary algorithms (e.g. GA) rely on expensive stochastic, iterative EM simulations — prohibitively time-consuming.
Deep-learning approaches need massive pre-collected datasets that are costly to build and fail to adapt to new specifications.

Goal: a data-free framework that needs neither offline datasets nor brute-force search.

2Proposed Framework (MPGN)

We propose the Metal Pattern Generation Network (MPGN) — a fully data-free generator that produces binary metal patterns directly, eliminating dependence on offline EM datasets and costly brute-force searches.

An online differentiable surrogate, the Gradient Estimation Network (GEN), trains alongside the generator and predicts antenna performance in real time — providing gradient guidance that cuts the required HFSS simulations by orders of magnitude (≥ 100×) versus evolutionary iterations.

3BiScaleNorm

Binary metal-pattern generation suffers from a “gradient barrier” and severe gradient vanishing. BiScaleNorm independently scales positive and negative activations using running extremes (M, m):

(1) x̂ = xM, if x > 0 x|m|, if x < 0

This keeps activations out of saturation regions, preserving active gradients for hard binarization functions (e.g. Sign) and enabling stable training of discrete patterns.

4Two-Stage Exploration

Stage I — BiScaleNorm drives rapid convergence to a local optimum (achieves basic resonance).
Stage II — a periodic mutation mechanism escapes local basins, ensuring monotonic performance improvement (deepens S₁₁).

Avoids local optima while guaranteeing the design never regresses.

5Contributions

A fully data-free framework — no expensive offline EM datasets.
BiScaleNorm — resolves gradient saturation in binary optimization.
Two-Stage Exploration — lets the model escape local optima.

6Results — 5G n257 (26.5–29.5 GHz)

Validated on the 5G n257 band, the framework generated a double-layer patch antenna:

Patch antenna metal pattern — **Fig. 1(a)** Generated patch-antenna metal pattern.

Reflection coefficient S11 — **Fig. 1(b)** Stage II deepens S₁₁ below **−10 dB** across the band; Stage I reaches only basic resonance.

Realized gain — **Fig. 1(c)** Flat realized gain **> 4 dB** across the passband.

BiScaleNorm vs. baselines: superior convergence stability vs. standard normalization (e.g. LayerNorm), which failed to reach valid resonance due to gradient vanishing.

Conclusion & Generalizability

A data-free, gradient-guided generation framework that designs valid mmWave patch antennas with orders-of-magnitude fewer EM simulations. It is robustly generalizable — successfully re-targeted to other frequency bands with minimal configuration changes, demonstrating viability as a generic solution for diverse mmWave components.