Calculating material properties with purely data-driven methods

Abstract

Unsupervised machine learning (ML) methods are incorporated in this work to depict correlations and investigate hidden relations between data points that refer to a diffusion coefficient dataset for the Lennard-Jones (LJ) fluid. Widely used clustering algorithms are incorporated, such as the k-means, k-medoids, and DBSCAN, to create data frames of common characteristics that apply to the fluid's state space. The symbolic regression (SR) ML method is applied, next, on each cluster and provides analytical expressions that describe the fluid's behavior under various temperature and density conditions, suggesting an alternative pathway towards diffusion coefficient calculation. © 2022 ACM.