<p>Machine Learning adoption proves successful in educational methodology development for creating tailored learning systems that enhance Science, Technology, Engineering, and Mathematics (STEM) education for students with diverse characteristics. Standard educational institutions encounter difficulties in method adaptation because student populations demonstrate varied skill levels and learning approaches. This research develops an adaptive learning framework that leverages Transformer-Based Models, Graph Neural Networks (GNNs), and Reinforcement Learning (RL) under high-performance computing (HPC) infrastructure to optimize STEM education personalization. The system utilizes a dataset of 12,000 student interaction records, comprising features such as quiz scores, assignment submissions, time spent on learning platforms, and activity logs, to dynamically modify educational content and pathways in real time. HPC resources are efficiently utilized for both model training and inference. Experimental results show that HPC-enhanced Transformer models (BERT, GPT, and T5) effectively predict student learning needs with RMSE values of 0.412, 0.398, and 0.405, MAE values of 0.328, 0.315, and 0.322, and <i>R</i>² scores of 0.87, 0.89, and 0.88, respectively. GNN-based recommendation models (GCN and GAT) achieved over 85.6% and 87.3% accuracy, with content relevance scores exceeding 82.1% and 84.5%, and student satisfaction levels of 80.3% and 82.7%. HPC-enhanced RL approaches (DQN and PPO) demonstrated improved engagement through completion rates of 72.4% and 76.8% and learning outcome improvements of 14.2% and 18.7%. Overall, the proposed framework capitalizes on supercomputing to deliver real-time adaptive learning, significantly enhancing student engagement, efficiency, and retention in STEM education.</p>