Physical state-anchored spatiotemporal learning for full-field fatigue prediction and microstructure inverse design in polycrystals: data

This dataset record contains versioned archives of author-generated 201-frame CPFEM tensors for training, validation and testing; withheld 500-cycle continuation data; the frozen 625-design/2,500-RVE inverse-design topology bank; source data supporting the main and Supplementary figures and tables; the 24 prospective CPFEM verification input decks and losslessly extracted response arrays; and complete E1/E5 timing evidence. Author-generated numerical data are CC BY 4.0, while author-developed code members included in Core Evidence are BSD-3-Clause and identified member-by-member in the deposition manifest. Third-party calibration reference numerical data are not redistributed and must be obtained from their cited original sources. The record contains six ZIP archives: Core Evidence, Inverse Bank, Mainline Test, Mainline Validation, Mainline Train, and Figure and Table Source Data. The first five provide the underlying research data and evidence; the sixth provides the direct numerical data for the main and Supplementary figures and tables.