Predictive modeling of antenna performance using machine learning for IoT applications

Abstract The Internet of Things (IoT) is driving the demand for compact, energy-efficient antennas capable of supporting massive, low-power wireless connectivity. Traditional antenna design workflows, which rely on full-wave electromagnetic (EM) simulations and physical prototyping, are computationally expensive and time-consuming—limiting rapid prototyping and scalability. This study proposes a novel simulation-augmented machine learning framework that eliminates the need for external EM solvers during the early design phase. A custom Python-based antenna simulator is developed to analytically approximate performance metrics—gain, bandwidth, and efficiency—based on geometric and material parameters, enabling fast and scalable synthetic data generation. This dataset is used to train and evaluate a wide spectrum of regression models, including Random Forest, Gradient Boosting, Multi-Layer Perceptron Neural Networks, Adaptive Lasso with Cross-Validation (ALR-HT), Supervised Kernel Thinning (SKT), RankUp (transformed boosting), FastQR via Bayesian Ridge, Generalized Linear Models (GLM) using Poisson Regression, Bayesian Regression, and Adaptive Deep Ensemble Regression (ADR). Performance is benchmarked using R-squared (R $$^2$$ ), Mean Squared Error (MSE), and Mean Absolute Error (MAE). Results show that the neural network and the deep ensemble (ADR) achieve the highest gain accuracy (R $$^2$$ $$\approx $$ 0.998), Random Forest and RankUp lead on bandwidth (R $$^2$$ = 0.97 and 0.95), and Gradient Boosting and RankUp are strongest for efficiency (R $$^2$$ $$\approx $$ 0.99). Feature importance analysis reveals dielectric constant as the dominant predictive variable. The proposed framework offers a novel, scalable, and simulation-light approach to antenna optimization, significantly accelerating the design pipeline for next-generation IoT applications.