Scalable Parameter Space Optimization for HPC Simulations in Laser-Ion Acceleration

Kurzfassung

Laser-ion acceleration is a promising method for accelerating ion bunches with unique properties relevant to a wide range of applications. However, it faces challenges due to the non-linear and complex physics that prevent optimization for higher energies. This study uses a Gaussian process model as a surrogate to efficiently explore and optimize parameters for generating energetic ions in the RTF-RPA mechanism. The model was initially developed using data from a semi-analytical solution and subsequently trained with particle-in-cell simulations. To optimize the process of generating training data, the model was interrogated to determine the best sampling strategy adapted to the already existing data. A comparative analysis benchmarks random and Halton sampling with adaptive sampling. Various target functions for the adaptive sampling were tested, compared, and evaluated using statistical methods and metrics for convergence. The adaptive sampling algorithm and simulation interface have been implemented and integrated into a high-performance computing environment. This demonstrates improved exploration and optimization efficiency for high-energy ions through parallel automation and Gaussian process-guided adaptive sampling.