Degree Name

Department

Cell Biology

College

Life Sciences

Defense Date

2026-05-19

Publication Date

2026-06-12

First Faculty Advisor

Gus Hart

First Faculty Reader

Tyler Jarvis

Second Faculty Reader

Branton Campbell

Honors Coordinator

Amber Gonda

Keywords

Cluster Expansion, Alloy, Generalized Aliasing Decomposition, Predictive Modeling, Computational Physics

Abstract

The cluster expansion (CE) method is a powerful framework for modeling alloy systems by linking atomic configurations to thermophysical properties. However, constructing accurate and robust CE models requires careful selection of both training structures and cluster bases. In this work, we investigate how structure enumeration and cluster selection strategies influence CE performance in binary face-centered cubic alloys using a Julia-based implementation. Model performance is evaluated using empirical risk and the generalized aliasing decomposition (GAD), which provides insight into the relationship between model complexity and generalization error. We find that physically motivated selection strategies that prioritize small clusters consistently outperform approaches focused primarily on numerical stability. We also show that data-driven cluster selection methods have the potential to outperform both approaches and may provide a systematic framework for guiding CE construction. Overall, these results highlight the importance of combining physical intuition with data-driven methods in CE modeling, while demonstrating the usefulness of GAD for understanding complexity–generalization trade-offs. This work explore strategies for selecting clusters and training data in CE models and prompts further study in more complex multi-component systems.