Sprecher
Beschreibung
In this talk, we will explore the current development of PIConGPU in machine learning-based simulations for plasma acceleration and highlight three key applications. These projects mark significant advances in the integration of AI and advanced data workflows into plasma physics research with PIConGPU. They illustrate not only our current methods, but also our vision for future in-transit AI-assisted simulation analysis.
First, we present automated optimizations of large three-dimensional particle-in-cell (PIC) simulations to find parameters consistent with self-truncated ionization injection experiments during laser wakefield acceleration. By combining Bayesian optimization with the PIConGPU framework and automating the entire parameter tuning workflow with Snakemake, we efficiently achieve excellent convergence with experimental measurements.
Second, we present radINN, an invertible neural network designed to use emitted radiation spectra to quantify injected charge. Trained with synthetic data from PIConGPU simulations, radINN excels at identifying injection processes and provides valuable insights for improving future experimental diagnostic.
Finally, we present a new heterogeneous streaming workflow that streams plasma simulation data directly into a machine learning application, without writing any data to disk. This approach leverages openPMD and ADIOS2 to facilitate real-time model training during data transfer, overcoming the challenges of petabyte-scale data storage and I/O bottlenecks. We demonstrate this workflow by tackling the inverse problem of predicting particle dynamics from radiation in a PIConGPU simulation of the Kelvin-Helmholtz instability.
