The flow field of molten steel in the mold play an important role in the slab quality, and rapid and accurate prediction of flow-field is an important foundation of process optimization and on-line control for continuous casting. Because high computational cost of computational fluid dynamics (CFD) simulations for flow field in a slab continuous-casting mold cannot satisfy the demand of the rapid evaluation, an attention-gated dual-head U-Net model is proposed on the base of small-sample on a two-dimensional flow-field. First, three-dimensional steady-state flow field is obtained by OpenFOAM, and the related two-dimensional flow-field is obtained by interpolation. Next, a data extension method based on the statistical characteristics of turbulent kinetic energy is developed to expand the flow-field dataset according to the relationship between turbulent kinetic energy and fluctuating velocity. Furthermore, an attention-gated dual-head U-Net model with a multi-channel input integrating the signed distance function (SDF) and inlet velocity masks is proposed to adaptively focus on key flow regions, which include the jet from the nozzle and the upper and lower recirculation zones. The results show that, after training, the inference time for a flow-field is less than 1 s. Under the conditions of a fixed nozzle angle and various casting rates, the velocity mean absolute error (MAE) is on the order of 10^-3 m/s. Under the unseen nozzle-angle and casting-rate conditions, the velocity MAE remains on the order of 10^-2 m/s. These results demonstrate that the data extension method based on turbulent-kinetic-energy statistics combined with the attention-gated dual-head U-Net model can enable rapid reconstruction of molten-steel flow fields in the mold, and provide a feasible approach for subsequent multi-physics digital-twin modeling of continuous casting.