Topology optimization in fluid flow and heat transfer applications is inherently multi-objective, as competing performance measures such as pressure loss and thermal efficiency must be balanced. Exploring different objective weightings generally requires repeated full topology optimization runs, resulting in rapidly increasing computational costs even for closely related design preferences. To reduce this cost, recent studies have investigated end-to-end learning approaches that directly predict optimized topologies from initial designs. However, in multi-objective settings, such methods often exhibit limited generalization, as reliable coverage of the design space typically demands dense sampling of objective weightings and repeated retraining. To address these limitations, this work proposes a trajectory-based learning framework that exploits the information contained in full topology optimization processes. Rather than learning direct mappings from initial to final designs, a conditional U-Net model is trained on complete optimization trajectories to predict incremental topology updates. The model is conditioned on the current design state, a normalized evolution stage, and an objective-weighting parameter, enabling the learned evolution dynamics to adapt continuously across different objective settings. The network is trained exclusively on topology optimization trajectories corresponding to minimum temperature and minimum pressure loss. Despite this restricted training set, the learned evolution dynamics enable the generation of physically consistent and structurally reasonable configurations under arbitrary weightings between temperature and pressure objectives, allowing controlled interpolation of multi-objective trade-offs without explicit multi-objective retraining.