Learning the nonlinear plasma-mirror response: Fast surrogate modeling and optimization of ellipticity of attosecond pulses

Abstract

Optimizing the polarization of attosecond emission from relativistic laser–foil interactions is hindered by a non-convex response surface and the computational cost of particle-in-cell (PIC) scans. We present a surrogate-assisted framework that learns the nonlinear mapping from incident polarization to the ellipticity of the reflected attosecond pulse from an ultrathin plasma foil. A Fourier-embedded multilayer perceptron is trained on a 1D PIC dataset (fixed pulse energy) to predict bandpass-defined ellipticity and is then coupled to a global optimizer (dual annealing) whose proposals are verified by PIC. The surrogate attains a test MAPE of about 4% and evaluates new points ≈ 103 × faster than PIC, enabling rapid exploration of parameter space. Using N ≤ 10 surrogate-seeded candidates, PIC verification yields ellipticity gains of ≳ 2% over the best value present in the training scan; in a second setting (longer pulse and thicker foil), improvements of ≈ 3% are obtained with a dozen PIC validations. The optimum occurs off the circular-drive line (a 0, y ≠ a 0, z), highlighting non-intuitive structure in the control landscape. By reducing the number of costly simulations by an order of magnitude or more, this approach makes practical the routine design of circular or elliptical attosecond sources and transfers readily to other stiff, nonlinear simulation-driven optimization tasks.