This work develops a data-driven multi-fidelity topology design (MFTD) method for designing fins in a latent heat thermal energy storage tube. The high-fidelity simulation resolves the actual solid-liquid phase change process using the enthalpy method, while the low-fidelity topology optimization (TO) simply considers the natural convection with Darcy flow. The above MFTD method is integrated into the framework of evolutional algorithm, and the variational autoencoder is introduced to generate new offspring. Fins for accelerating the melting and solidification processes at different Grashof number (Gr) are designed. It is found that when the fin volume fraction is low, the melt designs exhibit strong heterogeneity due to the strong convection, while the solidification designs are almost isotropic. Along with the increase of the fin volume fraction, the fins are first getting longer, then having more branches or sub-branches and finally becoming thicker. The superiority of the present data-driven MFTD method to the gradient-based direct TO method for solving optimization problems with strong multimodality has been demonstrated in this work. Results find that compared with the designs from direct TO, the MFTD melt design can further reduce the melting time by at least 27 % and 20 % at Gr = 3.3 × 103 and Gr = 3.3 × 104 respectively, and the MFTD solidification design can further shorten the solidification time by at least 9 %.