Efficient data-driven polarization learning for attosecond science and nonperturbative nonlinear optics

Abstract

This paper is devoted to the computation of atomic/molecular polarization (dipole moment) or acceleration in the context of attosecond science and with preliminary application to nonperturbative nonlinear optics. Specifically, dipole moments and dipole accelerations are efficiently learnt for continuous sets of physical parameters using neural networks trained from a finite number of solutions to parameterized Time Dependent Schrödinger equations computed with classical numerical methods. We then propose an application to a Maxwell-Schrödinger system modeling the macroscopic propagation of intense and short laser pulses in a gas, and show that polarization learning allows for an important improvement of the computational efficiency. Some experiments and analytical results illustrate the proposed strategy.