Artificial intelligence applied to high harmonic generation driven by structured laser pulses

Abstract

[EN] In this thesis we present fundamental research at the intersection of High Harmonic Generation (HHG) and Artificial Intelligence (AI), demonstrating the potential of combining both fields to model and understand ultrafast physical phenomena. Throughout this work, the major result was to accurately and efficiently describe both the microscopic quantum dynamics and the macroscopic propagation effects of HHG within the development of an AI-based model trained with exact results from the Three-Dimensional Time-Dependent Schrödinger Equation (3D-TDSE). This approach drastically reduces computational costs (from months to hours or even minutes) while retaining the quantum-level precision, that enables us further exploration of the HHG process.

This hybrid AI-3D-TDSE model is capable of predicting the atomic dipole acceleration as a function of the amplitude and phase of the driving laser field. Beyond its simplicity, once trained, the neural network effectively replaces the 3D-TDSE at every point of the generating medium. Afterwards, the HHG emissions can be propagated to the far field through the integral solution of Maxwell’s equations.

The model has been successfully validated in theoretical and experimental contexts. It was first employed to simulate and analyze HHG driven by structured laser beams of vortices with different Orbital Angular Momentum (OAM), accurately reproducing the spatial intensity distributions, OAM spectra, and temporal emission profiles of synthesized attosecond pulses. Moreover, it has been used as theoretical backup of experiments generating HHG with Hermite-Gauss (HG) driving beams, where the harmonic propagation give rise to an interference pattern between the different lobes of the HHG beams formed in the near field.

Subsequently, the model was extended to mixed gaseous media, revealing a regime of coherent interference between species that acts as a natural spectral filter. These results highlight the potential of the developed framework for control the HHG emission, along with the attosecond pulse characteristics.

The results presented here show that integrating AI into strong-field physics not only accelerates computation but defines other paths to follow in the standard simulation methodologies. Although this model is not general, it is perfectly suited to compute macroscopic HHG driven by structured laser pulses. Looking ahead, future efforts may include extending the model to incorporate non-linear propagation of the driving beam in dense media, accounting for dispersion from both neutral atoms and free electrons, both affected by the gas pressure. In such a case, a comprehensive analysis over the required dataset must be performed, as the number of free parameters increases abruptly. Another direct extensions of the model could be the addition of the wavelength and polarization of the driving beam, giving rise to a further generalization of the methodology here developed.

From another perspective, the development of AI tools in HHG and attosecond science can be use to develope diagnostic techniques to characterize the properties of the Extreme Ultraviolet (EUV) emission in experiments. Attosecond light science faces several challenges in the development of characterization techniques that can retrieve spatiotemporally the properties (polarization, intensity, and phase) of attosecond pulses. AI may open the route for in situ measurements that can rely on accurate predictions.