X-ray-induced atomic transitions via machine learning: A computational investigation

Intense x-ray free-electron laser pulses can induce multiple sequences of one-photon ionization and accompanying decay processes in atoms, producing highly charged atomic ions. Considering individual quantum states during these processes provides more precise information about the x-ray multiphoton ionization dynamics than the common configuration-based approach. However, in such a state-resolved approach, extremely huge-sized rate-equation calculations are inevitable. Here we present a strategy that embeds machine-learning models into a framework for atomic state-resolved ionization dynamics calculations. Machine learning is employed for the required atomic transition parameters, whose calculations possess the computationally most expensive steps. We find for argon that both feedforward neural networks and random forest regressors can predict these parameters with acceptable, but limited accuracy. State-resolved ionization dynamics of argon, in terms of charge-state distributions and electron and photon spectra, are also presented. Comparing fully calculated and machine-learning-based results, we demonstrate that the proposed machine-learning strategy works in principle and that the performance, in terms of charge-state distributions and electron and photon spectra, is good. Our work establishes a first step toward accelerating the calculation of atomic state-resolved ionization dynamics induced by high-intensity x rays.