Schools in the hot-arid climates of underserved Iranian regions face significant cooling challenges and high energy consumption, leading to prolonged thermal discomfort. These issues primarily stem from inadequate infrastructure and inefficient mechanical HVAC systems. To address this, the present study employs a Machine Learning framework to optimize school building design for energy efficiency and occupant comfort. A detailed model of a two-classroom school in Yazd, Iran, was developed and simulated using EnergyPlus and jEPlus, generating over three million design scenarios. Key design parameters, including orientation, insulation thickness, wall type, shading, and HVAC setpoints, were varied to minimize two objectives: annual cooling energy and hours of thermal discomfort. Six machine learning algorithms were evaluated to predict these performance metrics, from which a fine-tuned Gradient Boosting model emerged as the most accurate predictor (R2 = 0.9971). The analysis confirmed that building orientation and cooling setpoint temperature were the most influential factors for both objectives. The study lies in its specific application which is optimizing schools in Iran's underserved, arid regions, a context largely overlooked by prior research. This work, therefore, provides a framework tailored to the unique climatic and infrastructural challenges facing these schools. The analysis of all scenarios identified an optimal configuration with a 172 degrees orientation, 10 cm of external insulation, a 0.83 m overhang, and a 28.3 degrees C cooling setpoint. This optimized design reduces the annual cooling load to 1898.4 kWh and discomfort hours to 1547.3, achieving a practical balance between energy efficiency and student well-being.