In today’s world, automation takes place in almost any industry.
Now the time has come to rethink the design process of
antennas and leveraging computational power
and new technologies like machine learning.
 - ©


Simulating antennas is computationally expensive and time consuming. It requires experts with long-time experience and in-depth knowledge in the domain of electrodynamics.
Based on the given constraints the expert chooses a topology whose radiation pattern is already known and providing a good baseline design approach. From there it is an iterative process where the expert analyzes the results of the simulation and adjusts the antenna’s parameters, in order to improve its performance, eventually leading to the specified radiation characteristics.
After an extensive process of parameter variations, it may well be that the initial guess of the topology was insufficient to reach certain performance goals and one has to start over again.


Machine learning is known for its capabilities to surpass human experts in certain tasks like image recognition, language translation, playing games and solving other tasks. Machine learning models are trained on structured data and learn patterns in the data rather than being programmed explicitly like rule-based systems. 

Artificial Neural Networks

Some machine learning models like Artificial Neural Networks, drawing inspiration from the interconnected network of neurons in the brain, are known for their ability to approximate functions which also can be highly non-linear. This ability relates to mapping input features to output targets. For example, in image classification the color values of an image are mapped to a class label we associate with the content of the given image.
 - ©

Possible Utilization of Machine Learning

Antennas are designed by defining parameters like the peak gain or directivity, polarization, frequency of operation and the required radiation pattern or coverage, just to name a few. All parameters are interrelated and can not be arbitrarily defined. Usually, an expert needs to know various antenna topologies and their respective radiation patterns to start with an initial design and further optimizing the geometry of the antenna until it yields the specified radiation pattern and resonant behavior.

Utilizing machine learning to alleviate the expert from that initial topology search, radiation patterns of commonly known and used antenna topologies could be accumulated to a dataset. The neural network model would be trained in a supervised fashion in order to predict the most suitable base topology class given a certain radiation characteristic. From the predicted base topology, the expert can continue to fine tune the antenna design. Of course, the antenna design is not only constrained by the radiation pattern it should exhibit but also by restrictions on geometry and materials used for example. These further
restrictions oppose a difficulty on the model to propose a suitable design as certain constraints might render the model’s proposal void.

Training of Neural Networks

Neural networks are known to need many examples in order to generalize properly. The more representative data samples the model can learn from during training, the better it can capture the relationship between the antenna parameters and the resulting radiation pattern. The goal is to have a model which has learned the underlying functional relationship in order to generalize to new unseen examples. Considering various geometrical and material combinations in addition to the other parameters that influence the antenna’s radiation pattern, it requires a huge amount of sampled data in order to represent the underlying relationships between all combinations. This seems not to be an effective way, also one has to consider that with the emergence of new materials, for example, the model has to be retrained entirely from scratch.
 - ©

Evolutionary Algorithms

Another approach, inspired by biology, could be used to generate randomized populations of design proposal within the given constraints and evaluate their performance. If none of proposed designs reaches the set performance threshold to be accepted, defined by a ‘fitness-function’, the best performing designs of the current generation gets randomly combined and permutated to yield the next generation of design proposals to be evaluated.

This process repeats until the performance threshold is reached. This method comes from the family of evolutionary algorithms and was already pursued to some extend in [1]. The drawback of this approach is that no knowledge is used to generate the antenna design and therefore relying on computational heavy search for every new query.

(Reference: [1] Hornby, Greg & Globus, Al & Linden, Derek & Lohn, Jason. (2006). Automated Antenna Design with Evolutionary Algorithms. Collection of Technical Papers - Space 2006 Conference. 1. 10.2514/6.2006-7242.)


Combining these two ideas to alleviate their drawbacks seems like a fruitful endeavor to be pursued to build systems which support engineers and scientist in creating wireless technology.

Here at PIDSO we are always on the frontier of pushing the boundaries of current technologies and finding novel ways to build better solutions.