Artificial Intelligence for the Future Photon Science
Closing the loop between theory and experiment in photon science
Laser-plasma accelerators make large, conventional particle accelerators more compact, less costly, as well as increase broad availability and access in science, industry and medicine. However, comprehension of the involved physics requires sophisticated and computationally demanding algorithms for simulation and reconstruction. This Helmholtz AI young investigator group aims to loop between theory and experiment by researching data-driven digital twinning techniques that stimulate theoretical comprehension as well as experimental validation of the very complex dynamics involved in laser-particle acceleration.
We research and apply recent surrogate modelling techniques such as Physics-informed Neural Networks for acceleration of state-of-the-art Particle-In-Cell simulations (PIConGPU) and identification of unmodelled dynamics from data (PDE identification & learning). That exploration requires fast reconstruction of experimental data, e.g. X-ray scattering data, by reliable neural networks. The latter revolutionize the way scattering experiments are carried out by fast and reliable data analysis leveraging large amounts of training data and injection of prior knowledge into the inference procedure of neural networks.