Deep learning reduced-order models (DL-ROMs) provide an efficient alternative for unsteady flow prediction, yet most existing approaches rely on spatially uniform loss formulations that fail to distinguish dynamically dominant regions. In vortex-dominated flows, this mismatch leads to systematic error accumulation within vortex cores and shear layers, ultimately degrading long-horizon prediction stability. To address this limitation, a vorticity-guided local weighting reduced-order modeling method based on a three-dimensional U-shaped convolutional neural network architecture is proposed. A physics-guided weighting strategy is incorporated at the loss-function level, adaptively emphasizing vortex-dominated regions during training without modifying the network structure or increasing model complexity. This design targets the primary source of error growth in unsteady flows while preserving computational efficiency. Numerical results on canonical cylinder wake flows show that the proposed method reduces long-horizon prediction error by nearly 40% compared with baseline DL-ROMs and maintains coherent flow structures over time. Additional cross-Reynolds-number tests are conducted to evaluate numerical generalization behavior under increasingly complex wake dynamics.
He et al. (Thu,) studied this question.