Hybrid inverse design of mode-locked fiber lasers

Dispersion in mode-locked fiber lasers (MLFLs) crucially shape the output spectra, underscoring the necessity for precise dispersion control. However, the direct impact of fine dispersion variations on output spectra has not been revealed yet. Here, a hybrid inverse design strategy is proposed, combining the fully-connected neural network (NN) with the genetic algorithm (GA) to predict the net second-order dispersion (SOD) of the laser cavity given the desired target spectrum. The NN trained with simulation data performs direct inverse design, which utilizes the target spectrum as input and predicts net SOD values. These predictions are further refined through GA iterative optimization with a numerical laser model for precise adjustments on the net SOD of cavity. The hybrid inverse design strategy manifests superior accuracy for arbitrary customized target spectrum, reaching a ∼35-fold lower average design error compared to indirect inverse design using GA alone. Moreover, the indirect inverse design of the hybrid approach converges in merely 2 generations, whereas searching with GA alone are unable to fall below the preset threshold even reaching the maximum of 10 generations. This approach is expected to build a new workflow for the inverse design of MLFLs.