Heres one paper that I can immediately think of, https://arxiv.org/abs/1409.7495. The authors use a synthetic dataset to select and enginer features of a “real” dataset. Not sure if this is what you are looking for but could be a step in the right direction.