Dataset of 22 million centrifugal compressors with their performance predicted by a 1D mean-line analysis model. The train* and test* files are the ones used in the case study. The data is stored as parquet (Arrow) format. File descriptions: dated_real: Simulations of the 14 known geometries. dated_real_map: Simulations along a grid (to draw the compressor map) of the 14 known geometries. dated_real_perturbation_1M1: 1st Batch of 1M perturbed real compressors. dated_real_perturbation_1M2: 2nd Batch of 1M perturbed real compressors. dated_uniform_19M: Batch of 19M uniformly sampled compressors. dated_uniform_compact_1M: Batch of 1M uniformly sampled compressors on a reduced operating range. test_* and train_* contain the data to reproduce the case study of the associated paper. The dataset was generated to discuss guidelines for generating synthetic dataset for engineering design. However, it can also be used to train pre-design tools for centrifugal compressors. More information can be found on Github.