Fig. 2: ALFRED design idea and application example.
a Swarm parallel optimizer that searches the multidimensional space of solutions with an array of tentative geometries that use a cooperative scheme based on self and social interactions. b Predictor for each tentative solution: a convolutional neural network (CNN) extracts features from an input geometry and predicts the spectral response s0(ω) through a series of fully connected networks trained on different material thicknesses that are automatically selected by the switch (S). c–f Design example of a polarizing beam splitter. c Target response that maximizes the normal transmission for TE (green) and normal reflection for TM (blue) polarizations with a flat first-order derivative at the target design wavelength λ0 = 900 nm. d Cost function for progressive iterations of ALFRED (blue: particles, red: mean value). e Designed structure with a grey unit cell and brown nanoresonators. f FDTD simulations of (e) for TE and TM polarizations at λ = λ0
