Machine Learning–Based Computational Framework for Graphene–Silicon Twin-Port Band-Notched Wideband Antenna in Optical Regime

The world of wireless communication is constantly racing toward faster speeds and greater data capacity. To keep up with the demands of future technology, scientists are looking beyond the radio waves we use today and into the terahertz spectrum. This region of the electromagnetic spectrum sits between microwaves and light, offering the potential for data transfer rates that are currently unimaginable. However, harnessing these frequencies is difficult. Signals at this level tend to fade quickly and get absorbed by the air, requiring specialized equipment to send and receive them effectively. The key to unlocking this potential lies in designing tiny, efficient antennas that can operate in this high-frequency regime without losing power or clarity.

In this pursuit, researchers have turned to a combination of advanced materials and smart computing. One promising material is graphene, a single layer of carbon atoms that is incredibly thin and conductive, which can be tuned to change how an antenna behaves. Another approach involves using silicon, a common material in electronics, shaped into specific forms to guide waves efficiently. The challenge, however, is that designing these antennas is a slow and complex process. Traditional computer simulations, which act as virtual wind tunnels for testing how waves interact with metal and dielectric structures, take a long time to run. To speed things up, a team of researchers from several Indian institutions has developed a new method that combines these physical designs with machine learning, a form of artificial intelligence that learns from data to make predictions.

The researchers set out to build a two-port antenna, essentially a device with two separate connection points, using a silicon base coated with a thin layer of graphene. They designed the antenna to operate within a specific wide range of frequencies, roughly between 2.35 and 3.8 terahertz. To make the device even more useful, they needed to create a "notch," or a deliberate gap, in its performance. This gap would block a specific range of frequencies between 2.75 and 3.15 terahertz, preventing interference from other signals that might be operating in that same band. They achieved this by adding circular metallic rings and a printed line to the surface of the antenna. Furthermore, to ensure the two ports did not interfere with each other, they incorporated a special pattern on the ground layer, known as a defected ground structure, which acts like a barrier to keep the signals separate.

To understand how this design worked, the team used powerful computer software to simulate the behavior of electromagnetic waves. These simulations revealed that the silicon dielectric rod, shaped with a modified ring slot, was responsible for the wide operating range. The simulations showed that the antenna successfully created two distinct bands of operation: one from 2.2 to 2.65 terahertz and another from 3.3 to 3.7 terahertz. The circular rings and printed line effectively blocked the unwanted frequencies in the middle, while the defected ground structure kept the coupling between the two ports extremely low, reducing it to below -25 decibels. This low level of interference is crucial for a device to function as a reliable multi-port system. The graphene coating added a layer of flexibility, allowing the antenna's frequency to be tuned by adjusting its electrical potential, much like turning a dial to change the station.

Recognizing that running these complex simulations for every possible design variation would take too long, the researchers introduced a machine learning framework to predict the antenna's performance. They generated a massive dataset by slightly changing the dimensions of the antenna's parts in the simulation software, creating over 125,000 unique data points. They then trained two different types of machine learning models on this data: a deep neural network, which mimics the layers of the human brain, and a random forest, which uses a collection of decision trees. The goal was to see if these models could accurately predict how the antenna would perform, specifically its ability to reflect signals, without needing to run the full, time-consuming simulation every time.

The results of this computational experiment were encouraging. Both machine learning models proved capable of predicting the antenna's behavior with high accuracy, achieving success rates well above 80 percent. The deep neural network provided slightly more precise predictions than the random forest model, correctly estimating the performance across the frequency spectrum. When the researchers compared the machine learning predictions to the actual simulation results, the numbers matched closely, confirming that the artificial intelligence had learned the underlying physics of the design. This suggests that in the future, engineers could use these models to design and optimize antennas in a fraction of the time it currently takes, simply by letting the computer learn from previous examples.

The final verification of the design came from cross-checking the results with different simulation tools, which confirmed that the antenna maintained stable radiation patterns and high efficiency. The device showed a gain, or ability to focus energy, of around 3.5 to 4.5 decibels, which is a strong performance for this type of application. The researchers also checked the diversity parameters, which measure how well the two ports work together in a fading environment, and found that the signals remained uncorrelated and effective. By combining a novel physical design using silicon and graphene with a smart machine learning prediction system, the team has created a blueprint for a high-speed antenna that could one day support the terabit-per-second data speeds required for the next generation of wireless communication.