Paper  Title  Page 

TUPA55  Progress Toward Improving Accelerator Performance and Automating Operations with Advanced Analysis Software  465 


Funding: Research presented in this conference paper was supported by the Laboratory Directed Research and Development program of Los Alamos National Laboratory under project numbers XXG2, XX8R and XXB6. The penetrating radiography provided by the Dual Axis Radiographic Hydrodynamic Test (DARHT) facility is a key capability in executing a core mission of the Los Alamos National Laboratory (LANL). A new suite of software is being developed in the Python programming language to support operations of the of two DARHT linear induction accelerators (LIAs). Historical data, built as hdf5 data structures for over a decade of operations, are being used to develop automated failure and anomaly detection software and train machine learning models to assist in beam tuning. Adaptive machine learning (AML) that incorporate physicsbased models are being designed to use noninvasive diagnostic measurements to address the challenge of time variation in accelerator performance and target density evolution. AML methods are also being developed for experiments that use invasive diagnostics to understand the accelerator behavior at key locations, the results of which will be fed back into the accelerator models. The status and future outlook for these developments will be reported, including how Jupyter notebooks are being used to rapidly deploy these advances as highlyinteractive web applications. 

Poster TUPA55 [1.919 MB]  
DOI •  reference for this paper ※ doi:10.18429/JACoWNAPAC2022TUPA55  
About •  Received ※ 15 July 2022 — Revised ※ 08 August 2022 — Accepted ※ 10 August 2022 — Issue date ※ 12 August 2022  
Cite •  reference for this paper using ※ BibTeX, ※ LaTeX, ※ Text/Word, ※ RIS, ※ EndNote (xml)  
TUPA65  Machine Learning for the LANL Electromagnetic Isotope Separator  490 


Funding: Los Alamos National Laboratory Electromagnetic Isotope Separator Project. The Los Alamos National Laboratory electromagnetic isotope separator (EMIS) utilizes a Freeman ion source to generate beams of various elements which are accelerated to 40 keV and passed through a 75degree bend using a large dipole magnet with a radius of 1.2 m. The isotope mass differences translate directly to a spread in momentum, dp, relative to the design momentum p0. Momentum spread is converted to spread in the horizontal arrival location dx at a target chamber by the dispersion of the dipole magnet: dx = D(s)dp/p0. By placing a thin slit leading to a collection chamber at a location xc specific isotope mass is isolated by adjusting the dipole magnet strength or the beam energy. The arriving beam current at xc is associated with average isotope atomic mass, giving an isotope mass spectrum I(m) measured in mA. Although the EMIS is a compact system (5 m) setting up and automatically running at an optimal isotope separation profile I(m) profile is challenging due to timevariation of the complex source as well as unmodeled disturbances. We present preliminary results of developing adaptive machine learningbased tools for the EMIS beam and for the accelerator components. 

DOI •  reference for this paper ※ doi:10.18429/JACoWNAPAC2022TUPA65  
About •  Received ※ 18 July 2022 — Revised ※ 07 August 2022 — Accepted ※ 08 August 2022 — Issue date ※ 10 August 2022  
Cite •  reference for this paper using ※ BibTeX, ※ LaTeX, ※ Text/Word, ※ RIS, ※ EndNote (xml)  
THXD2  6D Phase Space Diagnostics Based on Adaptive Tuning of the Latent Space of EncoderDecoder Convolutional Neural Networks  837 


We present a general approach to 6D phase space diagnostics for charged particle beams based on adaptively tuning the lowdimensional latent space of generative encoderdecoder convolutional neural networks (CNN). Our approach first trains the CNN based on supervised learning to learn the correlations and physics constrains within a given accelerator system. The input of the CNN is a high dimensional collection of 2D phase space projections of the beam at the accelerator entrance together with a vector of accelerator parameters such as magnet and RF settings. The inputs are squeezed down to a lowdimensional latent space from which we generate the output in the form of projections of the beam’s 6D phase space at various accelerator locations. After training the CNN is applied in an unsupervised adaptive manner by comparing a subset of the output predictions to available measurements with the error guiding feedback directly in the lowdimensional latent space. We show that our approach is robust to unseen timevariation of the input beam and accelerator parameters and a study of the robustness of the method to go beyond the span of the training data.  
Slides THXD2 [19.086 MB]  
DOI •  reference for this paper ※ doi:10.18429/JACoWNAPAC2022THXD2  
About •  Received ※ 18 July 2022 — Revised ※ 05 August 2022 — Accepted ※ 08 August 2022 — Issue date ※ 09 August 2022  
Cite •  reference for this paper using ※ BibTeX, ※ LaTeX, ※ Text/Word, ※ RIS, ※ EndNote (xml)  
FRXE1  Bayesian Algorithms for Practical Accelerator Control and Adaptive Machine Learning for TimeVarying Systems  921 


Particle accelerators are complicated machines with thousands of coupled time varying components. The electromagnetic fields of accelerator devices such as magnets and RF cavities drift and are uncertain due to external disturbances, vibrations, temperature changes, and hysteresis. Accelerated charged particle beams are complex objects with 6D phase space dynamics governed by collective effects such as space charge forces, coherent synchrotron radiation, and whose initial phase space distributions change in unexpected and difficult to measure ways. This twopart tutorial presents recent developments in Bayesian methods and adaptive machine learning (ML) techniques for accelerators. Part 1: We introduce Bayesian control algorithms, and we describe how these algorithms can be customized to solve practical accelerator specific problems, including online characterization and optimization. Part 2: We give an overview of adaptive ML (AML) combining adaptive modelindependent feedback within physicsinformed ML architectures to make ML tools robust to timevariation (distribution shift) and to enable their use further beyond the span of the training data without relying on retraining.  
Slides FRXE1 [34.283 MB]  
DOI •  reference for this paper ※ doi:10.18429/JACoWNAPAC2022FRXE1  
About •  Received ※ 08 August 2022 — Revised ※ 10 August 2022 — Accepted ※ 12 August 2022 — Issue date ※ 27 September 2022  
Cite •  reference for this paper using ※ BibTeX, ※ LaTeX, ※ Text/Word, ※ RIS, ※ EndNote (xml)  