The multimode waveguide is essential for future mode-division-multiplexing optical interconnection systems to further improve data capacity. However, the complex waveguide design process based on numerical methods is time consuming and requires a lot of computational effort. By using machine learning trained models, one can find the rules from the observed data (sample) and use the learned rules (model) to predict the unknown or unobservable data quickly. In this paper, to accelerate the multimode waveguide design, several regression models consisting of fully connected and long short-term memory layers are employed to predict the effective refractive indices from the fundamental mode to fourth-order TE mode with various waveguide geometries. For air cladding waveguides, the percentages of eligible data whose prediction errors are less than 10 − 3 of different modes are 89.95%, 88.10%, 82.29%, 75.83%, and 71.19%. For S i O 2 cladding ones, they are 95.40%, 92.81%, 90.90%, 81.99%, and 86.39%. Based on these models, two kinds of mode multiplexers are designed, and coupling efficiency of higher than 85% is achieved.