Hyperparameter and Model Structure Selection in Lattice Structure Applications

Abstract

The intersection of data science and materials engineering presents significant opportunities for advancing the state of lattice material research using machine learning (ML). However, a gap remains in the understanding of hyperparameter tuning and model structure selection for material scientists applying graph neural networks (GNNs) to lattice materials. This thesis aims to bridge that gap by providing a guide to the processes behind model design. The goal is to ultimately improve the ability of materials engineers to better utilize GNNs in their research of both lattice materials and broader materials topics.

The study begins with an overview of lattice materials and their unique properties, followed by a detailed breakdown of the core parameters that govern the performance of neural networks, particularly GNNs. A review of recent work from several materials science research groups highlights current applications of ML techniques in the study of lattice materials. To illustrate the parameter selection process, a dummy task is introduced, showcasing how hyperparameter tuning and structure selection impacts model performance. Results from this task demonstrate the critical importance of appropriate parameter choices for accurate and efficient modeling.

My research aims to be readable by both materials engineers and data scientists, providing both a practical framework for materials engineers to apply GNNs and a good overview of the research space for data scientists who may wish to collaborate on related projects.