ML-based Control Methods
Model predictive control enabled by data-driven models for improved accelerator operation and optimization.
MPC Control Machine Learning
Overview
Modern particle accelerators require real-time control of thousands of coupled parameters. Classical PID controllers and lookup tables cannot handle the complexity of multi-input, multi-output (MIMO) systems with highly nonlinear responses. This project develops Data-0Driven Model Predictive Control (MPC) methods powered by data-driven surrogate models.
Model Predictive Control
MPC solves an online optimization problem at each control step:
subject to the system dynamics and operational constraints. The key ingredient is a fast, accurate model — which we learn from data.
Data-Driven Surrogate Models
We train surrogate models using:
- Koopman-based linear predictors for convex MPC formulations
- LSTM to handle time-delays in the system
Accelerator Applications
- Lower-level RF control of RFQ, which has significant delays
- Beam stablity control of hadron beam sources
- Beam halo management and loss prevention
- Adaptive tuning under drift and perturbations