To address the problem of limited accuracy in single-task learning for airfoil icing prediction and the difficulty in simultane-ously obtaining icing morphology and aerodynamic characteristics, a multi-task learning based airfoil icing prediction framework called MTAF (Multi-Task Airfoil Former) is proposed. The framework employs an encoder-decoder architecture that extracts deep representations of airfoil geometry and flight conditions through a shared feature encoder, realizes feature fusion and task interaction via a task correlation module, and accomplishes icing prediction and aerodynamic coefficient prediction respectively through dedicated prediction heads. The performance of Transformer and ConvNeXt backbone net-works within this framework is comparatively studied. Based on training and testing with 625 NACA 2412 airfoil samples, results demonstrate that the Transformer architecture achieves pixel accuracy of 99.79% and coefficient of determination for 0.990 6 in icing prediction tasks, significantly outperforming 99.50% and 0.977 2 of ConvNeXt. For aerodynamic coefficient prediction, Transformer achieves coefficient of determination values of 0.975 4 for lift coefficient and 0.987 4 for drag coefficient, which are comparable to 0.973 3 and 0.990 1 of ConvNeXt, respectively, with each architecture showing distinct advantages. The MTAF framework maintains stable prediction performance across different icing types and various flight conditions, providing a high-precision prediction tool for airfoil icing engineering applications.