Paper Number
ML14 My Program
Session
AI and ML in Rheology
Title
Clustering and measurement methods for quantitative microstructural analysis using molecular dynamics data: A case study of polymer crystallization
Presentation Date and Time
October 22, 2025 (Wednesday) 4:25
Track / Room
Track 6 / Sweeney Ballroom C
Authors
- Tourani, Elyar (University of Tennessee, Chemical and Biomolecular Engineering)
- Edwards, Brian J. (University of Tennessee, Chemical and Biomolecular Engineering)
- Khomami, Bamin (University of Tennessee, Chemical and Biomolecular Engineering)
Author and Affiliation Lines
Elyar Tourani, Brian J. Edwards and Bamin Khomami
Chemical and Biomolecular Engineering, University of Tennessee, Knoxville, TN 37996
Speaker / Presenter
Khomami, Bamin
Keywords
theoretical methods; computational methods; methods; machine learning; polymer melts; polymer solutions
Text of Abstract
Accurate characterization of local symmetry, order, and connectivity is essential for linking microstructural evolution, like crystallization, to mechanical behavior in molecular dynamics (MD) simulation data. This study introduces a novel, physically informed clustering framework tailored to MD-driven data and designed to systematically identify and track evolving structural domains. Our method integrates the desired variables, specifically the measures of the crystallinity index, into a grid-based clustering architecture within a Connected Component Analysis (CCA) framework.
To address the common locality and sparsity issues in spatial datasets, we develop a de novo diffusion-based imputation technique, enabling robust reconstruction of missing or noisy local ordering information in physical systems without introducing artificial biases. Compared to conventional clustering methods, our approach demonstrates unparalleled superior efficiency, consistent precision in spatial resolution, and interpretability of clustering parameters. We assess biases, establish best practices for clustering, and perform sensitivity and error analysis in various procedures. Our framework avoids heuristic parameter tuning and ensures consistent cluster identification across diverse datasets by using physically motivated variables. The computational complexity is retained at O(n log n), matching or exceeding that of widely used algorithms while improving the sensitivity to subtle ordering transitions.
The clustering framework is applied to MD simulations of supercooling-induced and flow-induced crystallization in polyethylene melts, serving as case studies. Our method effectively captures anisotropic nuclei evolution and provides key metrics (volume, surface area, and convexity) for understanding structure-property relationships. Although focused on crystallization, this methodology is broadly applicable for studying microstructural transformations in polymeric and soft matter systems.