Unmanned Aerial Vehicles (UAVs) operating in icing environments face critical mission risks due to aerodynamic degradation and sensor occlusion from ice accretion. This paper presents a modular, real-time failure prediction framework that integrates CFD-informed surrogate models with a certifiable Simulink-based control system. Using physics-derived simulations on both a NACA 0012 airfoil and an EO/IR sensor dome, neural regressors were trained to predict lift/drag degradation and visual occlusion under varying atmospheric conditions. The validated models, achieving sub-5% error, are integrated into a real-time Simulink architecture designed for onboard execution, enabling live risk evaluation, autonomous alert triggering, and manual override control. The framework supports future integration with mission management interfaces and fault-tolerant UAV stacks, offering a scalable foundation for AI-enhanced survivability across defence, aerospace, and autonomous flight systems.