Outlook Towards Deployable Continual Learning for Particle Accelerators

TL;DR

This paper addresses the challenge that data drift and concept drift hamper the long-term deployment of ML in particle accelerators. It argues that deployable continual learning—an approach that updates models over time while retaining prior knowledge—offers a path to robust performance across evolving machine conditions. The authors survey existing ML applications in accelerators, review continual learning methodologies, and map them to practical use-cases (anomaly detection, optimization/control, surrogate models, and virtual diagnostics) while outlining the required infrastructure and workflows. The work highlights concrete strategies, including memory-based replay, gradient projection, meta-learning, and hybrid approaches, and provides guidance for implementing continual-learning pipelines with appropriate hardware, software, and monitoring. Overall, the paper aims to catalyze research and collaboration toward sustainable, drift-resilient ML deployments in accelerator facilities.

Abstract

Particle Accelerators are high power complex machines. To ensure uninterrupted operation of these machines, thousands of pieces of equipment need to be synchronized, which requires addressing many challenges including design, optimization and control, anomaly detection and machine protection. With recent advancements, Machine Learning (ML) holds promise to assist in more advance prognostics, optimization, and control. While ML based solutions have been developed for several applications in particle accelerators, only few have reached deployment and even fewer to long term usage, due to particle accelerator data distribution drifts caused by changes in both measurable and non-measurable parameters. In this paper, we identify some of the key areas within particle accelerators where continual learning can allow maintenance of ML model performance with distribution drifts. Particularly, we first discuss existing applications of ML in particle accelerators, and their limitations due to distribution drift. Next, we review existing continual learning techniques and investigate their potential applications to address data distribution drifts in accelerators. By identifying the opportunities and challenges in applying continual learning, this paper seeks to open up the new field and inspire more research efforts towards deployable continual learning for particle accelerators.

Figures (6)

• Figure 1: Differences between a typical surrogate model and digital twin. (a) Schematic Diagram of surrogate model system. (b) Schematic diagram of a digital twin system with two way information flow with a possibility of having human-in-the-loop.
• Figure 2: Drift observed in beam current data obtained at SNS accelerator. The y-axis represent mean of the beam current waveform per macro-pulse. Though, it is a harsh summarization of full macropulse about 100K measurements long, it clearly demonstrates the drift in the measurements.
• Figure 3: Impact on ML model prediction performance of drift due to slow variation in upstream inputs (laser distribution, rf cavity phase and amplitude) and potentially other unknown changes from a configuration change at a small accelerator. All data shown is test data, but during the period before the marker, there is corresponding training data. This shows the impact of drift on ML model predictions (in this case of the beam size). Reproduced from Edelen-MLforDesignAndControl.
• Figure 4: Data drift within same classes (no concept drift)
• Figure 5: A guide to selection of appropriate continual learning method(s) given system constraints