Kinetic network models to study molecular self-assembly in the wake of machine learning

Self-assembly provides a powerful tool to fabricate advanced materials through the bottom-up approach. Understanding the kinetics of self-assembly would thus greatly help achieve precise control in designing elegant structures for advanced materials. However, it is difficult to elucidate the kinetic details of self-assembly at a molecular level. Kinetic network models (KNMs) based on molecular dynamics (MD) simulations have the potential to elucidate the ensemble of self-assembly pathways that would be difficult to obtain experimentally. In this article, we discuss several major challenges of applying KNMs to study self-assembly and review the recent development of machine learning approaches to address these challenges. First, it is important to consider permutations in the molecular index to properly describe self-assembled structures containing identical monomers. To address this issue, physical coordinates invariant to permutations can be adopted (e.g., aggregate size and morphology in the self-assembly of nanoparticles, and the number of rhombic and hexagonal ice in ice nucleation on a surface). More recently, self-assembled structures are described as undirected graphs, which can be further adopted in the Graph Neural Network to identify metastable states. Second, KNMs of self-assembly processes may yield numerous parallel pathways with comparable probabilities, making it difficult to understand the major mechanisms of self-assembly. To address this challenge, the path lumping algorithm is developed to group numerous transition pathways obtained from transition path theory into metastable path channels according to their kinetic similarity. Finally, reinforcement learning-based methods open a window of opportunities for the rational design of kinetic pathways to control the self-assembly process. In the wake of machine learning, we anticipate that KNMs have the potential to be widely applied in molecular self-assembly studies and facilitate the design of new materials.

Automated summary

Self-assembly is a powerful approach for fabricating advanced materials, but understanding its kinetics at a molecular level is challenging. Kinetic network models based on molecular dynamics simulations have shown potential in elucidating self-assembly pathways. This article discusses challenges in using these models and reviews the recent development of machine learning approaches to address them. It also highlights the potential of these models in molecular self-assembly studies and material design.