Arctic sea ice continues to shrinking with the global warming,creating a significant chance to shipping on the Arctic routes.However,some issues still hinder the opening of Arctic shipping routes.Especially,the navigation safety is threatened by the uncertainty of sea ice prediction due to the complex physical mechanism of sea ice change.In recent years,deep learning algorithms,having excellence in tackling nonlinear fitting problems,have shown increased evidence of potential to address the problem of sea ice prediction.This paper provides a critical review of existing deep learning methods developed for sea ice prediction.First,the deficiency of numerical model in sea ice prediction was analyzed,leading out the advantage of deep learning methods.Second,the limitations of adopting deep learning methods only were pointed out in light of the characteristics of sea ice.Finally,the possible paths to adopting specialized knowledge of meteorology and oceanography into deep learning methods so as to provide a better prediction model were discussed.