As one of the main methods to study ice formation of aircraft， numerical simulation introduces a lot of parameter uncertainties when calculating ice formation， which affects the accuracy and reliability of the numerical simulation. It is important to develop methods of uncertainty quantification and quantify the uncertainty scientifically for evaluating numerical simulation results. To solve the problem of high-dimensional input-output that is difficult to be solved by traditional parameter uncertainty quantification methods， an ice shape prediction proxy model is proposed based on the proper orthogonal decomposition and error back-propagation neural network. The proxy model is proved to have high accuracy and excellent generalization ability under single input and double input parameters by taking the droplet median size and temperature as examples. Finally， on the basis of ice shape calculated by the proxy model with Monte Carlo sampling， the icing range is established by criteria 2σ. It is found that the uncertainty of droplet median size mainly affects the ice angle growth of glaze ice， while the superposition of temperature and droplet median size uncertainty affect the frost ice thickness. This study establishes a method for the subsequent impact analysis of multi-icing conditions and provides ideas for the uncertainty quantification of multi-dimensional input-output.