Predictive crystallography at scale: mapping, validating, and learning from 1000 crystal energy landscapes

Area of Science:

• Materials Science
• Computational Chemistry
• Solid-State Physics

Background:

• Computational crystal structure prediction (CSP) is vital for materials discovery.
• CSP reveals trends and insights beyond observed crystal structures.
• Previous CSP studies were limited in scope and scale.

Purpose of the Study:

• To demonstrate the reliability and scalability of CSP for small, rigid organic molecules.
• To perform the largest survey of CSP for over 1000 organic compounds.
• To enable large-scale data generation for materials design.

Main Methods:

• Force-field-based CSP investigations.
• Analysis of over 1000 small, rigid organic molecules.
• Development of machine-learned energy potentials (neural network lattice energy correction, MACE equivariant message-passing neural network).

Main Results:

• CSP located 99.4% of observed experimental structures.
• 74% of observed structures were ranked among the most stable.
• Developed transferable machine-learned potentials improving energy rankings.

Conclusions:

• The CSP workflow is highly reliable and scalable for organic molecular crystals.
• Large CSP datasets provide broad utility and explanatory power for materials design.
• This approach facilitates insights into crystal properties and rationalizes empirical rules.