The invention relates to a
crystal variation self-coding generation method based on feature decoupling and
mask optimization, a medium and a program product. The
crystal variation self-coding generation method comprises the following steps: extracting atomic absolute position features, unit
cell parameter features, component features and
mask features of a
crystal; preprocessing the features; performing learning training on the variational auto-
encoder model to enable the variational auto-
encoder model to learn internal representation of crystal features; and generating features by using the trained variational auto-
encoder model, decoding the generated features by generating codes, splicing component vectors and element coordinates to obtain a corresponding relation, and finally generating a new
crystal structure. Compared with the prior art, the
feature extraction method and the
model architecture have the advantages that the understandability of the features and the accuracy of description are improved, the
model learning training is more efficient, the key information of the
crystal structure can be more accurately captured, and a
solid foundation is provided for generating a new
crystal structure.