Unsupervised topological learning approach of crystal nucleation

Area of Science:

• Materials Science
• Computational Physics
• Chemistry

Background:

• Crystal nucleation is vital but its atomic-level mechanisms remain unclear.
• Experimental observation of nucleation's sub-picosecond and nanometer scales is challenging.
• Classical Nucleation Theory has limitations in explaining complex nucleation pathways.

Purpose of the Study:

• To reveal the detailed structural features of crystal nucleation without prior assumptions.
• To investigate the interplay of ordering phenomena during homogeneous nucleation.
• To explore element-specific nucleation pathways beyond Classical Nucleation Theory.

Main Methods:

• An unsupervised learning approach utilizing topological descriptors from persistent homology.
• Simulations applied to monatomic metals to observe nucleation phenomena.
• Analysis of structural features at nanometer length and sub-picosecond time scales.

Main Results:

• Identified simultaneous translational and orientational ordering during homogeneous nucleation.
• Demonstrated that ordering occurs in regions with low five-fold symmetry due to strong bonding.
• Revealed element-specific nucleation pathways that differ from Classical Nucleation Theory predictions.

Conclusions:

• Unsupervised learning with topological descriptors effectively reveals nucleation mechanisms.
• Homogeneous nucleation in metals involves simultaneous ordering driven by bonding.
• Nucleation pathways are element-dependent and more complex than previously hypothesized.