“…However, the realization of controllable and highly efficient self-focusing is still an ill-posed inverse problem. Traditional inverse design method of varied line-spaced grating was usually based on intuitive considerations, systematic fine-tuning (grid searching) or other traditional optimization method [9], which would cost expensively both computation resources and time. A small change of its target properties will require a recalculation and re-optimization of the model which will consume a huge number of resources.…”