Smarter Beam Control with AI: The SmartISIS Project
STFC Scientific Computing’s Ada Lovelace Centre (ALC) has successfully completed SmartISIS, an innovative project designed to revolutionise beam transport optimisation at the ISIS Neutron and Muon Source. By developing intelligent surrogate models for the Medium Energy Beam Transport (MEBT) component of the ISIS injector, the project harnessed cutting-edge AI to enhance performance and operational efficiency. The new MEBT section will be installed during the 2027 long shutdown to improve the beam transport in the ISIS linear accelerator.
“Having fast-executing, accurate and uncertainty-aware surrogate models of the accelerator through the SmartISIS project will support both novel accelerator design and operations by allowing us to explore the accelerator’s operating space more rapidly. This will enable the use of modern control and optimisation techniques, such as Bayesian optimisation and reinforcement learning, to improve accelerator performance and support more stable operation. Being able to develop, train and evaluate these approaches offline using the surrogate models will also allow us to achieve results more quickly and safely before deploying them in the control room”
Kathryn Baker, Machine Learning and Controls Applications Developer, ISIS Neutron and Muon Source
The Challenge
Particle accelerators like ISIS rely on precise beam control to ensure safe and efficient operation. Conventional modelling of beam dynamics uses computationally demanding simulation tools such as the ASTRA particle tracking code, an approach that limits real-time decision-making and optimisation. The existing models cannot easily account for noise and uncertainty, which introduces inaccuracies and leaves opportunities for efficiency gains untapped.
The SmartISIS project aimed to overcome these challenges by introducing an AI-driven surrogate modelling approach. These surrogate models approximate complex simulations much faster, enabling rapid optimisation and calibration to support live accelerator operations.
Our Approach
The team created a HyperNetwork-based surrogate model trained on ASTRA simulations. Unlike static models, this surrogate integrates machine settings and beam parameters to predict beam evolution with remarkable speed. To address sharp peaks in beam loss profiles (a critical operational metric) AI-based techniques were applied, significantly improving predictive accuracy.
To complement the surrogate model, the project tested an advanced Bayesian Optimisation (BO) framework, which intelligently searches for machine settings that maximise transmission efficiency. This algorithm was trialled on the real ISIS injector’s Low Energy Beam Transport (LEBT) system and achieved consistently improved transmission, demonstrating the power of AI in real-world accelerator environments.
“By combining accelerator simulations with machine-learning surrogate models, the SmartISIS project demonstrated how complex beam behaviour could be predicted much more quickly. This could support operators in exploring machine settings more efficiently, while uncertainty estimation and Bayesian optimisation provide a route towards safer and more informed optimisation of accelerator performance.”
Jaehoon Cha, Data Scientist, Scientific Computing
Benefits
SmartISIS delivers more than just technical gains; it supports STFC’s mission to enhance national research facilities. By reducing reliance on labour-intensive calibration and diagnostics, the project improves ISIS uptime for experiments benefitting materials science, clean energy research, and healthcare innovation.
This work paves the way for applications of intelligent control systems in other STFC facilities and beyond through advanced engineering and digitisation. Globally, the project enhances UK leadership in accelerator science, supporting international collaborations in data-driven facility optimisation.
Technical Insight
At its core, the surrogate model uses latent space mapping and meta-learning techniques to capture non-linear relationships between input configurations and beam properties. Unlike traditional polynomial fittings, this approach allows robust extrapolation under noisy conditions, opening the door to the development of digital twins for accelerator systems. Plans are underway to extend the concept to model uncertainties explicitly, creating an uncertainty-aware optimisation pipeline critical for safety in high-power beamlines.
