SUPR
ML-based diagnostics of laser-matter interactions
Dnr:

NAISS 2024/5-126

Type:

NAISS Medium Compute

Principal Investigator:

Arkady Gonoskov

Affiliation:

Göteborgs universitet

Start Date:

2024-03-27

End Date:

2025-04-01

Primary Classification:

10303: Fusion, Plasma and Space Physics

Secondary Classification:

10306: Accelerator Physics and Instrumentation

Allocation

Abstract

The project is a continuation of the ongoing activity dedicated to the development of ML (machine learning)-based methodologies for experimental diagnostics and theoretical analysis in the research area of high-intensity laser-plasma physics. Over the past few years, several striking examples demonstrated the potential of solving long-standing problems in this area and opening entirely new pathways in both theoretical and experimental studies. These include methods for reconstructing the interaction scenarios with limited diagnostics as well as for retrieving indicative features, symmetries, self-similarities and other peculiarities. In this context ML-based tools are expected to establish a new paradigm for understanding and mastering complex systems hardly accessible for traditional analytical methods. Apart from the general interest, further studies in this area are strongly motivated by the necessity of developing a systematic basis for retrieving information in modern experiments on the high-intensity laser facilities in Sweden (the Relativistic Attosecond Physics Laboratory at Umeå University and the Lund Laser Centre) and worldwide (ELI-NP, ELI-beamlines, the Central Laser Facility at RAL, the CORELS laser in South Korea and others). Despite the ever increasing capabilities of the numerical tools, several well-known difficulties have re-emerged with greater severity along the path to using large-scale simulations for experimental design and analysis at these facilities: the initial conditions for the interactions are not well characterized and vary from experiment to experiment, while the diagnostics produce a limited output that is difficult to disentangle. In the continuation of this activity we intend to benefit from our recent finding - achieving exact energy conservation within explicit kinetic plasma simulation [see Gonoskov JCP 502, 112820 (2024)]. We intend to use the computational resources of NAISS to generate synthetic data for training ML models to infer difficult-to-measure parameters that are key to understanding and controlling laser-plasma interactions. The obtained data will be used for modern approaches including transfer learning with the final goal of creating reliable ML-based diagnostics for experiments in laser-plasma physics. In particular, we intend to develop ML-based diagnostics for the current experiments on High-Harmonic Generation (HHG) and Vacuum Laser Acceleration (VLA) of electrons being conducted at the Relativistic Attosecond Physics Laboratory at Umeå University.