Accelerating Bayesian Optimization via Robust Surrogates

Particle accelerators are sophisticated research tools that require expert-level input from dedicated operators to meet key performance parameters. Machine learning is poised to improve operations even for large and multi-faceted accelerator facilities. Speci_cally, the application of digital twins will allow signi_cant enhancements in machine control and uptime, while also providing a powerful resource for experimental planning by facility users. However, physics-based models of particle accelerators often cannot be run in a timely manner, even with modern hardware, for such applications. Advances in machine learning have provided the ability to train sophisticated surrogate models that encapsulate complex, nonlinear relationships. The use of surrogate models, trained on simulations and machine measurements o_ers a powerful tool that can provide rapid evaluations of a particle accelerator's state, once trained. The application of appropriately trained surrogate models can be used as part of digital twin for a particle accelerator, both as an online model and to inform novel control schemes. A system for archiving, searching, and deploying the trained surrogates at a world class particle accelerator facility will be developed. The resulting system will be used to train a surrogate model of an active particle accelerator using measurements, and where necessary, simulated data. E_cacy of the surrogate models and model management system will be demonstrated by application of the surrogate models to advanced control schemes tested on the particle accelerator. Tools for developing, managing, and deploying surrogate models have wide applicability to complex systems in industry and science. The ability to deploy surrogate models for internal use with a digital twin while also providing controlled access for external users to plan experiments will be useful for any large-scale scienti_c equipment.