TOWARDS CONTINUAL MACHINE LEARNING FOR PARTICLE ACCELERATORS

Machine Learning (ML) has become an essential tool in modern scientific and engineering applications, enabling predictive modeling for complex systems. Many particle accelerator facilities are adopting ML-based solutions to accelerate time-consuming optimization tasks through fast inference, and to enable low-latency anomaly prediction. However, ML models assume stationary data distribution, as such, when data distribution drifts away from the training data, ML models’ performance degrade. In particle accelerators, data drift is inevitable. These drifts can originate from either changes in the machine settings or non-measured factors such as equipment degradation. In this paper, we present an application of rehearsal based continual learning method to maintain model performance on drifting data. We present an ML surrogate to reconstruct beam current data from Spallation Neutron Source accelerator that can be used for downstream tasks such as anomaly detection. We use the data from different beam settings that demonstrate systematic known shift in the data. We demonstrate that a model trained incrementally on new data lose performance on previous data distributions due to catastrophic forgetting. In contrast, integrating rehearsal based continual learning can maintain model performance in such scenarios and limit forgetting on previous data distributions.