Luca Scomparin

Information

Institute: IPE
Room: 207
Phone: 25650
Email: Luca.Scomparin#kit.edu

Thesis topic
Embedded Deep Reinforcement Learning for the control of relativistic electron bunches on FPGA

The Helmholtz association is at the forefront of fundamental and applied science research. One of its goals is the realization of compact and cost efficient accelerators for applications in science, industry, and medicine. Increasing the availability and efficiency of these kind of systems will have far-reaching societal implications. The tuning of future particle accelerators will pose great challenges. Specifically, the increase in the dimension of the training task will in turn reduce the performance of the optimizer, in a phenomenon known as curse of dimensionality. This will have impact on thebudget of large scientific endeavors as the resource requirements for commissioning and operation will potentially become unbearable. One of the possible solutions is the use of Machine Learning (ML) techniques in order to achieve automatic control. A promising approach is the use of Reinforcement Learning (RL), where an agent is trained to interact with an environment in order to maximize the cumulative reward, based on a set of observations. Specifically, this could represent a key tool to control the dynamic occurring when performing accelerator research and development in extreme conditions. For instance, short bunches with length in the order of a few ps and charges in the nC range can be routinely achieved in synchrotron light sources. In these conditions the MicroBunching Instability (MBI) occurs. This phenomena leads to the formation of microstructures in the longitudinal phase space that interact with their own emitted Coherent Synchrotron Radiation (CSR). This leads to a strong radiation emission in the THz range, but in turn gives rise to a complicated non-linear dynamics that induces strong fluctuations in the emitted radiation power. The control of this phenomenon is currently limited. A more stable THz source would have applications ranging from material science to biology. Achieving an high degree of control of these fast phenomenas brings its own challenges. The feedback loop needs to be implemented to act in the timescales of the dynamics, in this way ruling out the use of conventional off-the-shelf hardware. Thus, a custom platform needs to be developed. A general system that allows to apply RL to control fast dynamic at an accelerator through self-learning is potentially one step towards achieving autonomous accelerators. The use of general self-learned controllers would strongly reduce the amount of human effort necessary during the commissioning and day-to-day operation of novel facilities.

Conferences and Talks

1. 11th Workshop on Longitudinal Electron Bunch Diagnostics, held in Lille, France between June 29th and July 1st 2022;
2. 10th MT ARD ST3 Meeting 2022, held in Berlin, Germany between September 7th and 9th 2022;
3. 11th International Beam Instrumentation Conference 2022, IBIC 2022, held in Kraków, Poland between September 11th and 15th 2022;
4. 8th Annual MT Meeting 2022, held in Hamburg, Germany between September 26th and 27th 2022;
5. 14th International Particle Accelerator Conference 2023, IPAC 2023, held in Venice, Italy between May 7th and 23th 2023;
6. 12th Workshop on Longitudinal Electron Bunch Diagnostics, held in Karlsruhe, Germany between June 12th and 14th 2023;
7. 11th MT ARD ST3 Meeting 2023, held in Dresden, Germany between July 5th and 7th 2023.