Graphene MIMO Antenna Reaches Terahertz Speeds With Machine Learning Design

Engineers have long chased the terahertz band, the stretch of spectrum between 0.1 and 10 THz that promises the enormous bandwidths a future 6G network would need for terabit-per-second links. The catch has always been the antenna itself: at these frequencies, conductor and dielectric losses eat into efficiency, and packing multiple antennas close enough for a compact device invites crippling mutual coupling.

The design, developed by a team led by Md. Ashraful Haque and Shorab Hossain Somrat, centers on a radiating patch carved with three connected plus-shaped slots. The central slot is mirrored on the left and right sides of the patch and rotated outward, a geometry the authors say pushes surface currents away from the feed and creates two closely spaced resonant modes. In the final configuration, simulated in CST Studio Suite, those resonances land at 3.8125 THz and 4.0725 THz with return losses of −53.62 dB and −41.18 dB respectively — deep dips that indicate most of the input power is accepted by the antenna rather than bounced back to the source.

The antenna is small by terahertz standards. Its overall footprint is 140 × 140 square micrometers, with a patch measuring 90 × 130 square micrometers, and the final MIMO layout spans 1.77 wavelengths by 4.57 wavelengths. The radiating patch and ground plane are 3-micrometer-thick graphene, chosen after the team compared aluminum, copper and graphene patch materials; graphene produced the deepest return losses and a gain exceeding 10 dB, while copper managed only −20 dB and −16 dB at 4.2 and 4.71 THz and aluminum a single shallow resonance at 4.7 THz. The substrate is an 8-micrometer layer of polyimide with a dielectric constant of 3.5 and a loss tangent of 0.0027, selected over silicon and alumina after both alternatives produced four unwanted resonances with narrow bandwidths.

Getting two antennas to coexist at these frequencies required a decoupling strategy. The team built three MIMO configurations with identical elements spaced 80 micrometers apart edge to edge: one with an inverted element, one with orthogonal elements, and one arranged back to back. The best of the three, the back-to-back layout, still only managed isolation of −24.8 dB. Adding a wall between the elements changed the picture. A copper wall improved isolation to −28 dB but shrank the operating bandwidth, while the proposed graphene wall pushed mutual coupling down to −31.87 dB across a wide band — a comparison the authors present as evidence that the material itself, not merely the presence of a barrier, drives the improvement.

The reported MIMO metrics are striking on paper. Peak gain reaches 11.38 dB, radiation efficiency 84.73 percent, and the envelope correlation coefficient, a measure of how independently the two elements radiate, stays below 0.000249 across the band, corresponding to a diversity gain above 9.998 dB. Channel capacity loss, which quantifies how much element correlation erodes theoretical MIMO throughput, remains below the commonly recommended 0.5 bits per second per hertz limit, peaking near 0.45 bps/Hz between 3.75 and 4.05 THz and dropping to 0.095 bps/Hz near 5.1 THz. Simulated radiation patterns show stable directional behavior, with a main lobe of 23.1 dB(V/m) on the φ = 0° plane at 3.8125 THz.

Beyond the electromagnetic model, the authors distilled the antenna into an equivalent lumped RLC circuit and simulated it in Keysight’s Advanced Design System. Parallel RLC branches represent the two resonant modes, additional branches reproduce the wideband response, and a shunt branch models the crosstalk introduced by the decoupling wall. After iterative tuning of the component values, the circuit’s reflection coefficient tracked the CST full-wave simulation closely, which the team offers as circuit-level validation of the design — an analysis step they note is missing from most comparable published terahertz MIMO antennas.

The machine learning component is where the work claims its most practical contribution. The researchers swept six design parameters — substrate thickness, feed width, ground thickness, box length, inset width and patch width — across defined ranges, generating 120 simulated data points. They split the data 80/20 into training and test sets and trained five regression algorithms: RANSAC, Extra Trees, Gradient Boosting, Decision Tree and XGBoost. XGBoost came out on top, explaining 88.70 percent of the variance in antenna efficiency with a variance score of 90.16 percent, and posting the lowest mean squared error of the five at 11.07 percent. RANSAC fared worst, with an R² of just 50.60 percent. The team then used the winning model to shrink the antenna volume by 15.1 percent while preserving matching, gain, efficiency, isolation and diversity performance.

Simulated-versus-predicted comparisons for 24 test points show the model tracking measured efficiency values between roughly 81 and 85 percent with only small deviations — for example, a simulated 81.42 percent against a predicted 81.16 percent at the first test point. The authors frame this as evidence that regression models can replace laborious parametric sweeps in terahertz antenna design, where each full-wave simulation is computationally expensive and the design space is large and nonlinear.

Important caveats remain. No physical prototype was built. The team states plainly that their laboratory lacks the specialized micro- and nanofabrication and terahertz measurement facilities needed to manufacture and characterize a micrometer-scale graphene antenna, and that graphene with terahertz-grade properties is not readily available commercially. All reported figures come from CST full-wave simulation and ADS circuit modeling, and the authors acknowledge that performance is highly sensitive to fabrication tolerances in slot geometry, feed dimensions, element spacing and the decoupling structure — small errors could shift the resonant frequencies and degrade matching, gain and isolation. Prototype fabrication, terahertz measurements and tolerance analysis are deferred to future work.

Even as a simulation-only result, the study lands in a competitive field. In a comparison table against recently reported terahertz MIMO designs, the proposed antenna’s combination of 1.33 THz bandwidth, −31.87 dB isolation and sub-0.00025 ECC stands out, since prior designs achieving very high isolation or very low correlation typically did so over bandwidths narrower than 0.04 THz, while wideband competitors traded away decoupling performance. Whether those simulated numbers survive contact with a real graphene fabrication line is the question the authors themselves have left open — but the design offers a concrete, circuit-validated starting point for the terahertz antenna hardware that 6G visions will eventually demand.

Subject of Research: Simulation-based design of a graphene terahertz MIMO antenna with decoupling walls and XGBoost-assisted optimization

Article Title: Regression machine learning-assisted graphene MIMO antenna design with decoupling structures for micrometer wave communication

Article References: Haque, M. A., Somrat, S. H., Billah, M., Tiang, J.-J., Singh, N. S. S., Nahas, M., & Zakariya, M. A. (2026). Regression machine learning-assisted graphene MIMO antenna design with decoupling structures for micrometer wave communication. Results in Optics, Article 101201. https://doi.org/10.1016/j.rio.2026.101201

Image Credits: AI Generated

DOI: 10.1016/j.rio.2026.101201

Keywords: graphene, MIMO antenna, terahertz, 6G, XGBoost, machine learning, decoupling wall, polyimide substrate, RLC equivalent circuit, CST simulation, envelope correlation coefficient, radiation efficiency