ML-Aided Antenna Design

ML-Aided Antenna Design is a machine learning system that predicts the S-parameter (S11) performance of a dual-band double T-shaped monopole antenna directly from its geometric design parameters, replacing time-consuming full-wave EM simulations with fast, accurate surrogate models. Three ML techniques — Lasso regression, K-Nearest Neighbours, and Artificial Neural Networks — are trained on 223 CST simulation samples and used to sweep over 2 million candidate designs to identify the optimal parameter set.

This project replaces the traditional trial-and-error EM simulation workflow with trained ML surrogate models. The reference antenna — a printed dual-band double T-shaped monopole operating at 2.4 GHz and 5.2 GHz for WLAN — has five variable geometric parameters (l21, l22, w1, w2, w). For each combination, CST Studio Suite produces 35 S11 sample points across 2–5.6 GHz. A Figure of Merit (FOM), defined as the sum of absolute S11 values across both target bands, condenses each design's full S11 curve into a single performance scalar. Three independently implemented ML models learn to map the five parameters → 35 S11 values → FOM, allowing the optimal design to be found by exhaustive ML-based search rather than exhaustive simulation.

The workflow has three stages: data collection, model training, and design optimisation. In data collection, 223 antenna configurations were simulated in CST Studio Suite to obtain S11 plots; FOM values were computed from those plots. In model training (80/20 train/test split, five-fold cross-validation), three MATLAB models are fitted: a Lasso regression with quadratic feature expansion (20 features, elastic-net α = 0.75, λ by 1-SE rule); a KNN regressor with k = 5 (Euclidean distance); and an ANN with one hidden layer of 10 neurons trained via Levenberg–Marquardt backpropagation (70/15/15 split). In design optimisation, all three models are queried over approximately 2.1 million parameter combinations and the set achieving the highest predicted FOM is reported as the recommended design.

• Predicts full S11 frequency response (2–5.6 GHz, 35 sample points) from five geometric dimensions in milliseconds vs. minutes per CST simulation
• Identifies the globally optimal design across more than 2.1 million candidate combinations — infeasible with EM simulation alone
• ANN achieves R² = 0.9489 (MSE = 0.742); Lasso R² = 0.7732; KNN R² = 0.7495
• Lasso sensitivity analysis: FOM shift < 2 dB across all tested coefficient perturbations
• All three models validated against a physically fabricated and lab-measured antenna
• FOM-based scalar scoring enables rapid ranking without inspecting individual S11 curves

Data pipeline: generate.m randomly samples the five design parameters within defined ranges to produce 223 configurations. Each is simulated in CST Studio Suite for 35 S11 values; FOM is computed by summing |S11(f)| at the target frequency points.

Lasso (`Lasso.m`): Raw inputs expanded to 20 quadratic features. MATLAB lasso() called with Alpha=0.75 and CV=5; Index1SE lambda selected. Sensitivity analysis (sensitivity_test.m) perturbs individual coefficients and re-evaluates FOM.

KNN (`knn.m`): Data normalised and split 70/30. fitcknn called with NumNeighbors=5; Euclidean distance for neighbour lookup.

ANN (`ANN_code.m`, `ANN_test.m`): Created via MATLAB's Neural Fitting App (fitnet). Hidden layer size = 10; training function = trainlm. Exported as myNeuralNetworkFunction.m for reuse.

Optimisation sweep: All three trained models evaluated over a fine-grained grid of ~2.1 million design points; the parameter set with maximum predicted FOM is reported.

No live demo. See the MATLAB source files (Lasso.m, knn.m, ANN_code.m) and the final report for full methodology, results, and comparison with CST simulations and physical measurements.

MATLAB · CST Studio Suite · MATLAB Neural Network Toolbox · Statistics and Machine Learning Toolbox (lasso, fitcknn)

• Surrogate modelling cuts antenna evaluation time from minutes (EM simulation) to milliseconds
• ANN achieves R² = 0.9489 — 95% of S11 variance explained by five geometric parameters
• Lasso model validated against random perturbation: FOM shift < 2 dB across all tested coefficient changes
• All three models validated against a physically fabricated and lab-measured antenna
• Over 2.1 million candidate designs searched to identify optimal parameters