CFD2VEC

A self-supervised pretraining architecture for 3-D steady and time-averaged flows, built as a general foundation model: one encoder is pretrained by masked field modelling on solved cases; the learned representation then serves field prediction at arbitrary query points, solver warm start (super-fidelity), frozen-embedding probes and case retrieval. The core is independent of any solver or dataset: solvers and data sources are plug-ins.

• Source code: https://github.com/Sensifai-BV/cfd2vec