Multi-fidelity machine learning methods for sputtering yield calculations relevant to magnetic fusion energy systems

Valaitis, Sonata

Permalink

https://hdl.handle.net/2142/127510 Copy

Description

Title

Multi-fidelity machine learning methods for sputtering yield calculations relevant to magnetic fusion energy systems

Author(s)

Valaitis, Sonata

Issue Date

2024-12-11

Director of Research (if dissertation) or Advisor (if thesis)

Curreli, Davide

Committee Member(s)

Vergari, Lorenzo

Department of Study

Nuclear, Plasma, & Rad Engr

Discipline

Nuclear, Plasma, Radiolgc Engr

Degree Granting Institution

University of Illinois at Urbana-Champaign

Degree Name

M.S.

Degree Level

Thesis

Keyword(s)

• multi-fidelity
• machine learning
• sputtering
• sputtering yield
• magnetic fusion energy systems
• tokamaks
• plasma-material interactions
• gradient boosting model
• artificial neural network
• Yamamura
• feature engineering
• binary collision approximation simulations
• RustBCA
• feature importance analysis
• SHAP
• plasma-facing components
• tungsten
• boron

Abstract

Sputtering of plasma-facing components in magnetic fusion energy systems is an area of significant concern in fusion research. Sputtering yield data in this regime is difficult to obtain both experimentally and computationally. A substantial database of sputtering yields for a range of fusion-relevant materials is systematically generated from empirical formulas, binary collision approximation simulations, and published experimental and calculated results. A machine learning pipeline is optimized for the sputtering yield prediction problem. Linear regression, artificial neural network, and gradient boosting models are assessed in combination with various feature engineering methods. A multi-fidelity gradient boosting tree demonstrates a gain in computational efficiency on the order of 10^6 compared with high-energy binary collision approximation simulations. The gradient boosting model accurately and robustly predicts sputtering yields for a selection of ion materials incident on tungsten and boron across a broad range of ITER-relevant incident ion energies and angles. Feature importance analysis is employed to enhance model interpretability and inform the development of a semi-empirical formula applicable for all angles of ion incidence. Generalizability of the model is assessed for unknown ion and target material parameters. The database and multi-fidelity machine learning model are made available online for web retrieval.

Graduation Semester

2024-12

Type of Resource

Thesis

Handle URL

https://hdl.handle.net/2142/127510

Copyright and License Information

Copyright 2024 Sonata Valaitis

Owning Collections