Fig.  3.

Experimental proof of convergence of the Neuroute embedded with the istGAN. (A) The morphological evolution of the Neuroute at different stages prior to convergence is depicted. Operationally, the istGAN is embedded in the experimental device at different training stages, and the Neuroute generator is regulated according to the given phase output, aligning with the experimental outcomes of the multi-target area as illustrated in (B). The normalized electric field distribution surrounding the local area of multi-targets with the CRLB parameters are employed to visually observe the convergence property of the istGAN. (C) Loss function of the forward model of the istGAN; the average training mean square error (MSE) converged to less than 0.00021 after 2,000 steps. (D) Loss function of the inverse model of the istGAN based on the pre-trained forward model; the average training loss converges to less than 0.0024 after 2,000 steps.