Glass is inherently unstable to crystallization. However, how this transformation occurs while the dynamics in the glass stay frozen at the particle scale is poorly understood. Here, through single-particle-resolved imaging experiments, we show that due to frozen-in density inhomogeneities, a soft colloidal glass crystallizes via two distinct pathways. In the poorly packed regions of the glass, crystallinity grew smoothly due to local particle shuffles, whereas in the well-packed regions, we observed abrupt jumps in crystallinity that were triggered by avalanches—cooperative rearrangements involving many tens of particles. Importantly, we show that softness—a structural-order parameter determined through machine-learning methods—not only predicts where crystallization initiates in a glass but is also sensitive to the crystallization pathway. Such a causal connection between the structure and stability of a glass has so far remained elusive. Devising strategies to manipulate softness may thus prove invaluable in realizing long-lived glassy states.

玻璃固有地对结晶不稳定。但是，人们对这种转变是如何发生的却又不了解玻璃中的动力学以粒子尺度保持冻结状态。在这里，通过单粒子分辨成像实验，我们显示出由于冻结密度不均匀性，软胶体玻璃通过两个不同的途径结晶。在玻璃的填充不良的区域中，由于局部颗粒混洗，结晶度平稳增长，而在填充良好的区域中，我们观察到由于雪崩引发的结晶度突然跳跃-涉及数十个颗粒的合作重排。重要的是，我们表明柔软度（通过机器学习方法确定的结构顺序参数）不仅可以预测玻璃中结晶的起始位置，而且对结晶路径也很敏感。迄今为止，在玻璃的结构和稳定性之间的这种因果关系仍然难以捉摸。因此，设计柔软度的策略在实现长寿命玻璃态中可能是无价的。