⚡ electrical engineering

Machine Learning Assisted Inverse Design of Pixelated mmWave Patch Antennas

This paper presents a machine learning-assisted framework for the inverse design of pixelated mmWave patch antennas in the 22–30 GHz band, utilizing an XGBoost classifier to filter non-resonant patterns, a hybrid CNN-BiLSTM surrogate model for accurate S11 prediction, and a latent space optimization approach to automatically generate antenna structures matching desired specifications.

Original authors: Nadeem Rather, Holger Claussen, Lester Ho

Published 2026-08-25

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Nadeem Rather, Holger Claussen, Lester Ho

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Modern wireless networks are pushing into a part of the radio spectrum known as the millimeter-wave band. This range of frequencies, sitting between 22 and 30 billion cycles per second, offers the high speeds needed for next-generation communication, but it is notoriously difficult to work with. At these speeds, the tiny antennas required to send and receive signals must be precisely shaped to resonate, or vibrate, at the correct frequency. If the shape is even slightly off, the signal fails to connect. Traditionally, engineers design these antennas using standard shapes like rectangles or circles, but these fixed forms limit what is possible. They restrict the ability to fine-tune the antenna for specific needs, leaving a vast landscape of potential shapes unexplored.

The challenge lies in the sheer number of possibilities. If one were to break the surface of an antenna into a grid of tiny squares, deciding whether each square is metal or empty, the number of possible combinations becomes astronomically large. Testing every single combination using computer simulations would take years, as each simulation requires several minutes of computing time to solve complex physics equations. Furthermore, most random patterns of metal squares simply do not work; they create broken islands of metal that cannot carry an electrical signal, rendering them useless. This creates a needle-in-a-haystack problem where finding a working design is inefficient and slow.

A team of researchers at the Tyndall National Institute in Ireland has developed a new way to navigate this complexity using machine learning. Instead of trying to guess the right shape or testing every possibility, they built a system that learns from experience to design antennas automatically. Their approach, described in a recent study, focuses on the 22 to 30 gigahertz range, a critical band for 5G and future wireless technologies. The researchers created a digital model of an antenna surface made of 437 tiny square pixels. Some of these pixels are metal, and some are empty, but the metal must form a continuous path from the power source to the edge of the antenna to function.

To train their system, the researchers first generated about 6,000 random patterns of these pixels and ran them through a high-fidelity physics simulator. The results were telling: only about 40 percent of these random designs actually worked, meaning they could resonate at the desired frequency. The other 60 percent were failures, creating a dataset that was heavily skewed toward useless shapes. To fix this, the team trained a computer program to act as a filter. This program learned to look at a new, untested pattern and predict whether it would work before the expensive simulation was even run. By using this filter to select only the promising patterns, they were able to generate an additional 4,000 successful designs, raising the proportion of working antennas in their training data to over half.

With this improved collection of working examples, the researchers trained a second, more sophisticated computer model. This model acts as a fast-forward simulator. Instead of running the slow, heavy physics calculations, it looks at the pattern of metal pixels and instantly predicts how the antenna will behave across the entire frequency range. It learns not just the general shape of the signal, but the specific dips and peaks that indicate a good connection. Crucially, this model was taught to pay special attention to the depth and position of the signal's resonance, ensuring it could accurately predict the most important features of a working antenna.

The final step was to reverse the process. Usually, engineers start with a shape and ask, "What does this do?" The researchers wanted to start with a goal, such as "I need an antenna that works at 25 gigahertz," and ask the computer to find the shape that achieves it. To do this, they used a mathematical technique that compresses the complex antenna shapes into a simpler, hidden space of numbers. In this compressed space, the computer could use a standard optimization method to search for the perfect shape. It adjusted the hidden numbers, checked the result against the target frequency using the fast-forward model, and repeated the process thousands of times. To ensure the result was physically possible, the system included a rule that forced the metal pixels to stay connected, preventing the creation of broken, non-functional designs.

The results of this automated design process were impressive. The team asked the system to create four different antennas, each targeting a different frequency within the 22 to 30 gigahertz band. The computer generated unique, complex pixel patterns for each request. When these digital designs were tested again in the slow, high-precision physics simulator, they worked exactly as predicted. In three of the four cases, the difference between the predicted frequency and the actual simulated frequency was less than 0.2 gigahertz, and the strength of the signal was nearly identical to the prediction. Even in the case with the largest difference, the antenna still met the requirement of working within the target band.

This work demonstrates that machine learning can effectively bridge the gap between abstract mathematical goals and physical hardware design. By filtering out bad ideas early and learning to predict performance quickly, the system can explore a design space that would be impossible for a human to navigate manually. The researchers note that while these designs have been successfully simulated, the next step is to build physical prototypes and test them in the real world to see how they perform outside the computer. If successful, this method could allow engineers to automatically create custom antennas for any specific need, reconfiguring them on the fly to adapt to changing network conditions without the need for manual redesign.

Technical Summary: Machine Learning Assisted Inverse Design of Pixelated mmWave Patch Antennas

Problem Statement
The deployment of 5G networks has increased demand for compact, reliable millimetre-wave (mmWave) antennas. While conventional microstrip patch antennas are widely used, their fixed geometries (rectangular, circular, etc.) limit the exploration of alternative topologies that could offer superior impedance matching or bandwidth. Pixelated antennas, which discretize the radiating surface into a grid of binary conductive/non-conductive cells, offer a vast design space. However, this approach transforms antenna synthesis into a high-dimensional combinatorial optimization problem ( $2^{300}$ possible layouts for a 300-pixel grid). Evaluating these layouts requires full-wave electromagnetic (EM) simulations (e.g., CST Microwave Studio), which are computationally expensive. Furthermore, randomly generated pixel patterns rarely yield functional antennas, leading to highly imbalanced datasets where the majority of samples are non-resonant, making efficient surrogate model training difficult.

Methodology
The authors propose an end-to-end machine learning (ML) framework comprising three integrated phases:

Classifier-Guided Dataset Augmentation:
- Initial Data: An initial dataset of ~6,000 pixel patterns was generated using a "structured-random" strategy. This approach draws from seven predefined shape families (e.g., rect block, split ring) to ensure topological diversity while enforcing constraints: at least 20 active pixels and a single connected metal region with a continuous path to the feed.
- Imbalance Issue: Only ~40% of these initial patterns achieved resonance ( $|S_{11}| \leq -10$ dB) in the 22–30 GHz band.
- Filtering: An XGBoost binary classifier was trained to distinguish resonant from non-resonant patterns before simulation. This classifier was used to screen a new pool of random patterns, selecting only those predicted to be resonant for full-wave simulation.
- Augmentation: This process added 4,000 high-quality samples, increasing the proportion of resonant designs in the combined 10,000-sample dataset from 40% to 52%.
Forward Surrogate Modeling (CNN–BiLSTM):
- Architecture: A hybrid deep learning model maps a $19 \times 23$ $19 \times 23$ binary pixel grid to the complex $S_{11}$ $S_{11}$ response across 801 frequency points (22–30 GHz).
  - Spatial Encoder: Three Conv2D blocks with residual connections extract geometric features.
  - Frequency Projection: Linear layers reshape features into a sequence.
  - Temporal Modeling: A 2-layer Bidirectional LSTM (BiLSTM) models inter-frequency correlations.
- Loss Function: A physics-guided composite loss was employed to emphasize resonance accuracy:
  $L = \lambda_1 L_{Re/Im} + \lambda_2 L_{dB} + \lambda_3 L_{depth} + \lambda_4 L_{freq} + \lambda_5 L_{slope}$
  This includes terms for complex $S_{11}$ MSE, magnitude (Huber loss), resonance dip depth, resonance frequency (via soft-argmin), and spectral smoothness.
- Training: The model was trained on the augmented dataset with weighted sampling to handle remaining class imbalance and data augmentation (flips).
Latent-Space Inverse Design:
- Dimensionality Reduction: A Convolutional Autoencoder (AE) was trained to compress the $19 \times 23$ binary patterns into a 64-dimensional latent space. This constrains the search to the manifold of physically valid, connected geometries.
- Optimization: The inverse design problem is formulated as gradient descent within the 64-dimensional latent space. The objective minimizes the difference between the surrogate-predicted $S_{11}$ and a target specification.
- Constraints: A connectivity projection operator ensures the generated binary pattern maintains a continuous electrical path to the feed. A straight-through estimator allows gradients to flow through the discrete binarization step.
- Robustness: The optimization employs a K=4 dropout ensemble of the surrogate and cyclic frequency shifts to reduce sensitivity to model bias.

Key Results

Dataset Efficiency: The XGBoost pre-filter successfully increased the density of resonant designs in the training set from 40% to 52%, validating the utility of classifier-guided augmentation.
Surrogate Accuracy: The CNN–BiLSTM forward model achieved a Mean Absolute Error (MAE) of 1.548 dB and an RMSE of 2.735 dB on the test set. For resonant samples, the mean error in predicting the resonance frequency was 0.724 GHz.
Inverse Design Validation: Four distinct antenna designs were generated targeting different frequencies within 22–30 GHz (22–23 GHz, 25 GHz, 26.5–27.5 GHz, and 27–28 GHz).
- All four designs, when simulated in CST, achieved $|S_{11}| \leq -10$ dB within their target bands.
- Frequency errors between the surrogate prediction and CST simulation ranged from 0.18 GHz to 0.76 GHz.
- Three of the four designs showed depth errors ( $\Delta|S_{11}|$ ) below 1.5 dB, consistent with the model's test-set performance.

Significance and Claims
The paper claims to present a novel, fully automated pipeline for the inverse design of pixelated mmWave patch antennas. Its primary significance lies in addressing the data inefficiency inherent in pixelated antenna design through classifier-guided dataset augmentation, thereby reducing the computational cost of generating high-quality training data. Furthermore, the integration of a physics-aware composite loss function and a latent-space optimization strategy enables the generation of fabrication-ready, connected antenna geometries that meet specific $S_{11}$ specifications without requiring exhaustive full-wave simulations during the design phase. The authors assert that this approach demonstrates the feasibility of automatically designing and reconfiguring antenna structures for the 5G mmWave spectrum.

Future Work
The authors outline three directions for future research:

Fabrication and measurement of the inverse-designed prototypes to quantify the simulation-to-hardware gap.
Extension of the framework to include radiation patterns, gain, and efficiency as additional design targets.
Adaptation of the pipeline for two-port bandpass filter synthesis, requiring prediction of the full scattering matrix.

Technical Summary: Machine Learning Assisted Inverse Design of Pixelated mmWave Patch Antennas

More like this