Uncertainty Aware ML-based surrogate models for particle accelerators: A Study at the Fermilab Booster Accelerator Complex

TL;DR

This paper introduces a novel uncertainty-aware surrogate modeling approach for particle accelerators, addressing limitations of existing methods by capturing out-of-distribution uncertainties and enabling continuous learning at Fermilab.

Contribution

The paper proposes a new method for calibrated uncertainty estimation in deep learning models, specifically tailored for particle accelerator applications, improving over standard techniques.

Findings

Outperforms standard uncertainty estimation methods in accelerator data

Effectively captures out-of-distribution uncertainties

Supports continuous learning applications

Abstract

Standard deep learning methods, such as Ensemble Models, Bayesian Neural Networks and Quantile Regression Models provide estimates to prediction uncertainties for data-driven deep learning models. However, they can be limited in their applications due to their heavy memory, inference cost, and ability to properly capture out-of-distribution uncertainties. Additionally, some of these models require post-training calibration which limits their ability to be used for continuous learning applications. In this paper, we present a new approach to provide prediction with calibrated uncertainties that includes out-of-distribution contributions and compare it to standard methods on the Fermi National Accelerator Laboratory (FNAL) Booster accelerator complex.