Melt crystallization is essential to many industrial processes, including semiconductor, ice, and food manufacturing. Nevertheless, our understanding of the melt crystallization mechanism remains poor. This is because the molecular-scale structures of melts are difficult to clarify experimentally. Computer simulations, such as molecular dynamics (MD), are often used to investigate melt structures. However, the time evolution of the structural order in a melt during crystallization must be analyzed properly. In this study, dimensional reduction (DR), which is an unsupervised machine learning technique, is used to evaluate the time evolution of structural order. The DR is performed for high-dimensional data representing an atom–atom pair distribution function and the distribution function of the angle formed by three nearest neighboring atoms at each period during crystallization, which are obtained by an MD simulation of a supercooled Lennard–Jones melt. The results indicate that crystallization occurs via the following activation processes: nucleation of a crystal with a distorted structure and reconstruction of the crystal to a more stable structure. The time evolution of the local structures during crystallization is also evaluated with this method. The present method can be applied to studies of the mechanism of crystallization from a disordered system for real materials, even for complicated multicomponent materials.

熔融结晶对于许多工业过程至关重要，包括半导体、冰和食品制造。然而，我们对熔体结晶机制的理解仍然很差。这是因为熔体的分子尺度结构难以通过实验阐明。计算机模拟，例如分子动力学 (MD)，通常用于研究熔体结构。然而，必须正确分析结晶过程中熔体结构顺序的时间演变。在这项研究中，降维（DR）是一种无监督的机器学习技术，用于评估结构顺序的时间演化。DR 是针对代表原子 - 原子对分布函数和结晶过程中每个周期三个最近相邻原子形成的角度的分布函数的高维数据执行的，这些数据是通过过冷 Lennard-Jones 熔体的 MD 模拟获得的. 结果表明结晶通过以下活化过程发生：具有扭曲结构的晶体的成核和晶体重建为更稳定的结构。结晶过程中局部结构的时间演变也用这种方法评估。本方法可用于研究真实材料的无序系统结晶机理，甚至适用于复杂的多组分材料。这是通过对过冷 Lennard-Jones 熔体的 MD 模拟获得的。结果表明结晶通过以下活化过程发生：具有扭曲结构的晶体的成核和晶体重建为更稳定的结构。结晶过程中局部结构的时间演变也用这种方法评估。本方法可用于研究真实材料的无序系统结晶机理，甚至适用于复杂的多组分材料。这是通过对过冷 Lennard-Jones 熔体的 MD 模拟获得的。结果表明结晶通过以下活化过程发生：具有扭曲结构的晶体的成核和晶体重建为更稳定的结构。结晶过程中局部结构的时间演变也用这种方法评估。本方法可用于研究真实材料的无序系统结晶机理，甚至适用于复杂的多组分材料。结晶过程中局部结构的时间演变也用这种方法评估。本方法可用于研究真实材料的无序系统结晶机理，甚至适用于复杂的多组分材料。结晶过程中局部结构的时间演变也用这种方法评估。本方法可用于研究真实材料的无序系统结晶机理，甚至适用于复杂的多组分材料。