Abstract
<title>Abstract</title> <p>Data-driven fluid modelling now spans application-specific surrogates, multi-physics pretraining, and Earth-system foundation models. Nevertheless, many CFD predictors are evaluated within a single application family or on canonical configurations with fixed or simplified geometries and boundary conditions. Governing parameters are often held fixed or supplied explicitly, while substantial shifts in flow family or physical regime frequently require task-specific training or fine-tuning. We present Generating fLOws frOm Data (GLoOD), a proof-of-concept autoregressive model combining a visual encoder-decoder with temporal self-attention over semantic snapshot embeddings. Using pressure and velocity fields alone, without explicit Reynolds number or geometric descriptors, GLoOD is trained on synthetic laminar Lattice-Boltzmann obstacle flows and evaluated without retraining on circles, geometry-shifted squares, lid-driven cavities, and von Karman vortex shedding. Paired identity-backbone ablations isolate the contribution of temporal sequence modelling. Across aggregate losses, drag diagnostics, line profiles, fieldwise velocity decompositions, and wake-lag tests, temporal self-attention consistently improves prediction over direct previous-frame forwarding. Transfer remains non-trivial under moderate shifts but deteriorates under stronger changes in boundary-value problem and temporal organisation. The SWIN visual encoder-decoder is more robust than its UNet counterpart.</p>