Computational science for accelerators

Computational science is the study of numerical methods and computing resources to address scientific problems. Modern accelerators require some aspect of computational science. This includes the use of high-performance computing for

• simulations of the performance of the FRIB linear accelerator (linac);
• computational modelling of beam-material interaction, including the charge-stripping and target system;
• machine-learning applications in the FRIB linac, such as accelerator tuning, anomaly detection, and fast particle identification;
• simulation and modeling of general accelerator research and development such as beam physics and design of accelerator components.

High-performance computing

In addition to on-site computing resources, researchers at FRIB take advantage of high-performance computing (HPC) resources at the Institute for Cyber-Enabled Research at Michigan State University, and the world-leading HPCs of the U.S. Department of Energy, such as the supercomputer at the National Energy Research Scientific Computer Center and the exascale HPC Frontier Supercomputer.

In additional to accessing the world-leading HPC hardware, the researchers at FRIB also design efficient algorithms that can efficiently run on the cutting-edge hardware architectures of exascale computation, including

• the auto differentiation algorithms in beam dynamics studies;
• fast algorithm for the Interactions of charged particles in accelerators.

FRIB accelerator modeling and high-level controls

The operation and optimization of the FRIB linac requires precise computer models and control frameworks. The researchers at FRIB have developed general software toolkits that have been benchmarked with FRIB accelerators and are readily used by the broader accelerator community.

• Beam envelope simulations code: FRIB-Linear-Accelerator-Modeling-Engine (FLAME)
• High-level control framework phantasy

Machine learning and artificial intelligence

Researchers at FRIB use algorithms such as neural networks, Gaussian processes, clustering algorithms, and dimensionality reduction to

• perform fast particle identification;
• enable automated beam-tuning and optimization for the FRIB linac;
• establish virtual diagnostics for determining beam distribution and beam loss control;
• detect anomaly in accelerator components.