(169bh) Computational Investigation of Reaction Coordinate Optimality for Ice Nucleation Studies

Crystallization plays a crucial role in a myriad of processes spanning from cloud microphysics to the production of pharmaceuticals and semiconductors, underscoring its profound significance in both microscopic and macroscopic processes. As a first-order phase transition, crystallization typically proceeds through the nucleation and growth mechanism. Within this framework, nucleation is the initial step in crystallization. As such, understanding nucleation at a microscopic level, is of intense interest for advancing our ability to understand and engineer all such processes. More often than not, nucleation is the rate-limiting step in crystallization.. This makes nucleation a rare event that occurs over short timescales but only after considerably longer waiting times. Conventional molecular simulation techniques fail to sample the right configurations that successfully trespass large energy barriers characteristic of the nucleation process. As such, advanced sampling techniques are needed for efficient exploration of the nucleation pathway. One necessary ingredient of such techniques is the availability of suitable collective variables (CVs)- also known as order parameters- that accurately describe the progress of nucleation.

In this work, we use molecular dynamics simulations with forward flux sampling (FFS), and machine learning to assess conventional CVs for homogeneous and heterogeneous ice nucleation. By analyzing a collection of configurations with diverse committer probability values, we screen tens of thousands of distinct CVs that construct crystalline nuclei using different criteria that employ five cut-offs and/or decision points. In addition to standard cut-offs used for constructing bond order parameters (BOPs), we assess the efficacy of two other strategies, namely chain exclusion and the inclusion or exclusion of the first hydration shell around a crystalline nucleus. We find these techniques to not culminate in more potent CVs, although their efficacy or lack thereof depends on the key cut-offs used for constructing BOPs.