This thesis addresses the attitude maneuvering and stabilization problem of underactuated spacecraft equipped with fewer than three independent internal actuators. Such systems exhibit fundamental controllability and stabilization limitations and are particularly sensitive to modeling errors and environmental disturbances. To tackle these challenges, a control architecture is proposed that combines a nonlinear model predictive controller (NMPC), a physics-informed neural network (PINN)--based residual torque compensator, and a Lyapunov-based supervisory safety mechanism. The NMPC serves as the baseline controller and is responsible for constraint handling, underactuation-aware maneuver planning, and nominal closed-loop stability under practical actuator limits. To mitigate performance degradation caused by model mismatch--primarily inertia uncertainty and unmodeled dynamic couplings--a PINN is trained offline to estimate residual disturbance torques using simulation data.