Machine learning, a type of artificial intelligence, significantly accelerates computational tasks and enables new technologies in a wide range of areas such as voice and image recognition, self-driving cars, stock market trading, and medical diagnostics.
Before tackling a specific task Machine learning Algorithms usually need to be trained with existing data. This allows the algorithm to make quick and accurate predictions about future scenarios. But what if you don’t have the data available for training and your work is brand new?
Currently, researchers at the SLAC National Accelerator Laboratory at the Department of Energy have demonstrated that machine learning can be used to optimize particle accelerator performance by teaching algorithms the basics. Physics The principles behind accelerator Operation-No prior data is required.
“Injecting physics into machine learning is a very hot topic in many fields of study. Materials science, Environmental science, Battery research, Particle physics Others “. Physical Review Accelerator and Beam.. This is one of the first examples of using physics-based machine learning in the accelerator physics community.
Educate AI in physics
Accelerator is powerful machine It energizes electron beams or other particles for use in a wide range of applications such as basic physics experiments, molecular imaging, and radiation therapy for cancer. To get the best beam for a particular application, the operator needs to adjust the accelerator for best performance.
For large particle accelerators, this can be very difficult as there are so many components that need to be adjusted. To complicate matters, not all components are independent. In other words, adjusting one can affect the settings of another component.
Recent research at SLAC has shown that machine learning can greatly support human operators by accelerating the optimization process and finding useful accelerator settings that no one has ever thought of. Machine learning can also help diagnose the quality of a particle beam without interfering with the particle beam, as other techniques usually do.
For these steps to work, researchers first had to train machine learning algorithms using data from previous accelerator operations, computer simulations that make assumptions about accelerator performance, or both. However, they also found that using information from physical models in combination with available experimental data could dramatically reduce the amount of new data needed.
New research shows that if you know enough about physics to explain how accelerators work, you don’t really need previous data.
The team used this approach to tune SLAC’s SPEAR3 accelerator to power the lab’s Stanford Synchrotron Radiation Source (SSRL). Using information obtained directly from physics-based models, the researchers say, the results were as good as training the algorithm with real-world archived data.
“Our results are the latest highlight of the progressive promotion at SLAC to develop machine learning tools for adjusting accelerators,” said SLAC Staff Scientist Joe Duris, Principal Investigator of the Study. Told.
Predict the unknown
That doesn’t mean that existing data is useless. Even if you bring down your physics, they are still useful. In the case of SPEAR3, researchers were able to further improve the machine learning model based on physical information by combining it with actual data from the accelerator. The team is also applying methods to improve the tuning of SLAC’s Linac Coherent Light Source (LCLS) X-ray lasers. It is one of the most powerful sources of X-rays on Earth, and archived data are available from previous experiments.
When the SLAC crew turns on LCLS-II next year, the possibilities for new methods will be fully revealed. This superconducting upgrade to LCLS includes a brand new accelerator that requires you to determine the optimal settings from the beginning. The operator may find it useful to have an AI by his side who has already learned some basics of accelerator physics.
Adi Hanuka et al, Gaussian process based on physical models for online optimization of particle accelerators, Physical Review Accelerator and Beam (2021). DOI: 10.1103 / PhysRevAccelBeams.24.072802
SLAC National Accelerator Laboratory
Quote: AI is the particle accelerator performance obtained from https://phys.org/news/2021-07-ai-physics-optimize-particle.html on July 29, 2021 (July 29, 2021). Learn physics to optimize
This document is subject to copyright. No part may be reproduced without written permission, except for fair transactions for personal investigation or research purposes. The content is provided for informational purposes only.
AI learns physics and optimizes particle accelerator performance
Source link AI learns physics and optimizes particle accelerator performance