Analysis of Self-Assembly Pathways with Unsupervised Machine Learning Algorithms

Abstract

Colloidal and nanoparticle systems display a rich and exciting phase behavior including the self-assembly of highly complex crystal structures. Nucleation and growth pathways toward crystallization have been studied both computationally and experimentally, but the mechanisms for the formation of the precritical nucleus and consequent crystal growth are yet to be fully understood. Recent advances in the application of machine learning algorithms applied to many-particle systems have led to significant breakthroughs in the ability for high-throughput analysis of phase transitions and the identification of crystal structures. We build upon these techniques to identify and analyze pathways for nucleation and growth in supercooled liquids of colloidal systems modeled with isotropic pair potentials. Our study involves the development of unsupervised machine learning models trained on spherical-harmonics-based descriptors. These models allow us to determine clusters of local environments that are present prior to and during crystallization. We analyze these environments to identify prevalent motifs and local order within the supercooled liquid prior to formation of the critical nucleus.

Machine learning and descriptor selection for the computational discovery of metal-organic frameworks

“…As the community starts undertaking bigger systems with more unlabelled data and running into high-dimensional datasets, unsupervised learning-based models will become crucial to tackle those problems. In 2019, an interesting such study was done for the colloidal systems by Adorf et al, where they applied unsupervised SVMs to identify pathways for nucleation and growth of super-cooled liquids [159]. The domains of novel MOFs synthesis and crystal structure prediction also carry similar attributes as the nucleation of super-cooled liquids and hence, the supervised learning approach of Adorf and team could also be utilised for those related domains.…”

Characterising soft matter using machine learning

“…Dimensional reduction using both linear and non-linear techniques was combined with unsupervised learning by Adorf and coworkers. 77 They went on to provide an alternative route to discovering the pathways to self-assembly, for example crystallization via nucleation. They began with a large number of descriptors including bond angles, bond lengths, spherical harmonic order parameters and the bispectrum environment descriptor.…”

Non-classical nucleation pathways in stacking-disordered crystals

“…To partition a multidimensional OP space in different volumes, each one associated with the local environment of a crystalline structure or liquid phase, we use artificial neural networks (NN). In condensed matter NN have been used for potential energy surface calculations [57,58], to construct accurate molecular force fields [59], to improve potential energy of coarse grained models for water [49], or for identification and classification of local ordered or disordered structures using supervised [60][61][62] and unsupervised [63][64][65][66][67][68] methods. Ideally, unsupervised learning allows to cluster high-dimensional OP space into sets corresponding to different structures before they have been identified [63][64][65]67].…”