<p>For vehicular communication systems, adaptive high-efficiency antenna structures are required to provide efficient communication within dynamic V2X environments that are prone to high speeds, Doppler effect, multipath fading, and interference. Therefore, this research work aims to design and experimentally verify a reconfigurable two-port slot-loaded microstrip MIMO antenna of size 5 × 5 mm² along with Defected Ground Structure (DGS) and PIN diodes.However, in contrast to conventional antennas using a structural optimization approach, the proposed antenna system utilizes a deep learning-based adaptive control mechanism to directly improve the performance of the antenna by predicting the channel variations and proactively selecting the optimal reconfiguration states. The framework consists of a Maximum-Entropy Regularized Decision Transformer (MERDT) and a Spatial–Temporal Graph Convolutional Network (ST-GCN), with the ST-GCN used to capture spatial–temporal patterns of vehicular mobility and channel correlation, and the MERDT used to generate long-horizon reconfiguration policies that dynamically adjust frequency tuning, beam direction, and radiation characteristics. This allows for the elimination of mutual coupling effects, enhanced impedance matching stability, reduced packet loss, and high spectral efficiency under dynamic V2X environments. Moreover, an Improved Binary Portia Spider Optimization (IB-PSO) algorithm optimizes the antenna geometry as well as the learning model’s hyperparameters in a combined fashion. This leads to faster convergence as well as a balanced trade-off between electromagnetic and intelligent performance. The experimental/simulation results show that the ECC value is less than 0.02, while the diversity gain is approximately 10&#xa0;dB. The radiation efficiency reaches up to 93%, while the peak gain is as high as 7.65 dBi. The return loss is as low as − 21.5&#xa0;dB, while the VSWR drops to as low as 1.12.</p>