PatchData: A Dataset of Pixelated Patch Antenna Topologies and Simulated Radiation Patterns for Deep Learning-Driven Optimizations

Description

PatchData is an open-access dataset of pixelated microstrip patch antenna topologies and their corresponding simulated electromagnetic responses, developed to support data-driven antenna design, topology optimization, surrogate modeling, and deep learning applications. The dataset contains 2,398 simulated antenna configurations generated using an 8 × 10 binary pixel grid, in which conductive and etched regions are stochastically generated and subsequently conditioned through connectivity and morphological constraints to ensure physically viable and electrically active antenna structures. Each instance establishes a direct correspondence between a high-resolution representation of the antenna topology and its frequency-domain reflection coefficient (S11) response obtained through full-wave electromagnetic simulations within an automated MATLAB–CST Studio Suite co-simulation workflow. The dataset encompasses different resonance behaviors, including single-band, dual-band, tri-band, and ultra-multiband configurations, with electromagnetic responses evaluated over the 1–8 GHz frequency range. The repository is organized to facilitate both direct research use and progressive learning. It includes the complete raw dataset of paired antenna topologies and electromagnetic responses, as well as a segmented subset specifically prepared for introductory or warm-up exercises, enabling users to become familiar with the data structure, visualization, preprocessing, and analysis procedures before working with the complete collection. In addition, the repository provides Python scripts and supporting resources for data inspection, spectral analysis, visualization, and deep-learning-based antenna topology reconstruction. The structured organization of the data enables researchers and students to independently access antenna geometries and their corresponding electromagnetic responses, reproduce the analyses presented in the associated research, and develop alternative computational approaches. PatchData is intended to serve as a reproducible benchmark for machine learning, artificial intelligence, surrogate modeling, inverse antenna design, electromagnetic response prediction, and automated topology optimization, providing a common data resource for exploring the application of data-driven methods to pixelated microstrip antenna design.

Steps to reproduce

The standardized workflow for deploying PatchData comprises the following phases: 1. Acquisition and Extraction: Users must first download the compressed archive file from the official Mendeley Data landing page. Once downloaded, the package must be unzipped into the local working directory. The internal directory is strictly presented into two primary functional divisions: the “raw dataset repository”, the "Segmented dataset", and the “automated programmatic suite”. 2. Understanding the Dataset Structure: Inside the raw dataset folder, users will find distinct sub-folders systematically segregated by their multi-band resonant behavior (Single-Band, Dual-Band, Tri-Band, and Ultra-Multiband). Each sub-folder contains paired data instances consisting of the high-resolution binary antenna topology saved as a “.png” spatial matrix and its corresponding frequency-domain continuous response archived in “.xlsx” format as an S11 scattering parameter vector. 3. Using the Programmatic Suite: This folder provides ready-to-use Python scripts optimized for deep learning workflows, which should be executed through the main.py script. This program uses the following files in sequential order: - random_case_visualizer.py: An inspection routine that randomly selects and plots paired instances of antenna geometry images (_Vista.png) and their corresponding electromagnetic responses (_S11.xlsx) for rapid database auditing. - spectral_centroid_calculator.py: A routine that processes the raw continuous S11 data files within each folder, applying a consensus filter based on cutoff frequencies to compute and index both the vector-valued arithmetic means (Spectral Centroid y(S11) and the closest real physical instance (Representative Pattern y ̂(S11) for each class. - decoder_CNN.py: Contains the complete compilation script to train, validate, and test the Deep Convolutional Decoder architecture using a hybrid spatial loss function (80% SSIM / 20% MAE) and stratified partitioning to synthesize binary antenna matrices from 1D S_11 frequency sweeps.

Institutions

• Instituto Tecnológico Metropolitano

  Antioquia, Medellín
• Institución Universitaria Pascual Bravo

  Antioquia, Medellín

Categories

Licence