image: This system utilizes machine learning algorithms to optimize the operation of particle accelerators, reducing manual intervention and enhancing precision in real-time control. By integrating virtual accelerators and reinforcement learning, the system achieves seamless transition from simulation to real-world applications, enabling efficient and adaptive accelerator management.
Credit: ©Science China Press
Under the robust drive of the AI for Science frontier, artificial intelligence is reshaping research paradigms at an unprecedented pace while profoundly transforming daily life. In recent years, with the rapid advancement of AI technology, autonomous driving has gained significant attention. But can a massive scientific apparatus with tens of thousands of components, such as a particle accelerator, achieve stable operation through a similar "autonomous driving" technology? The answer is yes. Scientists are leveraging machine learning to enable intelligent control of particle beams, opening up new possibilities for commissioning and operating high-power, high-intensity accelerators.
Particle accelerators are essential tools for exploring the structure of matter and fundamental physical laws, requiring extremely high operational precision. Traditionally, the tuning and operation of accelerators have relied heavily on manual intervention, consuming substantial human resources and significantly increasing the time cost of research. The introduction of machine learning provides a transformative solution to these challenges. By training intelligent controller, machine learning can drastically reduce manual intervention, enhance operational efficiency, and unlock new possibilities for equipment control.
However, the path to realizing this technology is fraught with theoretical and technical challenges. For instance, the dynamics of accelerators are exceptionally rapid, and available observational data primarily represent steady-state conditions, failing to capture the full dynamic evolution process. This characteristic renders traditional nonlinear dynamical control theories inadequate for direct application. Additionally, as accelerators are systems with extremely high degrees of freedom and spatiotemporal evolution, only partial variable information is observable, complicating the design and tuning of controllers. Furthermore, due to the high cost of obtaining real-world accelerator operational data, relying on virtual accelerators for offline training has become a practical alternative. Yet, achieving seamless transfer from virtual to real-world accelerators remains a significant technical hurdle.
A recent study, published in Science China: Physics, Mechanics & Astronomy, offers innovative solutions to these issues. The research, conducted by He Yuan’s team from the Institute of Modern Physics, Chinese Academy of Sciences, in collaboration with Zhao Hong’s team from Xiamen University, is titled “Machine Learning for Online Control of Particle Accelerators.” The first author of the study is Chen Xiaolong, and the corresponding authors are He Yuan and Wang Zhijun.
The team proposed a feasible pathway for “autonomous driving” of accelerators, addressing challenges at both theoretical and technical levels. On the theoretical front, they developed a control-process-based dynamic model for accelerators and introduced a time-series phase-space reconstruction technique to ensure the control system captures equivalent global information, enhancing reliability and controllability. Technically, they designed a high-precision virtual accelerator and a machine learning controller. By employing reinforcement learning algorithms, they efficiently processed massive data generated by the virtual accelerator, enabling offline training of the controller and its seamless transition to real-world applications.
Supported by these breakthroughs, the team achieved the first-ever global trajectory adaptive control of the CAFe2 superconducting segment, encompassing 42 degrees of freedom, which is now integrated into routine operations. This marks the first application of AI technology in complex accelerator systems within China, setting a significant milestone for machine learning applications in this field.
This accomplishment lays a solid foundation for further advancements in intelligent control technologies for accelerators. Future research is expected to expand the applicability of these theories and methods while developing more efficient machine learning algorithms, driving particle accelerator technologies to new heights.
Journal
Science China Physics Mechanics and Astronomy
Method of Research
Experimental study