Comparing molecules and solids across structural and alchemical space.

Area of Science:

• Materials Science
• Computational Chemistry
• Machine Learning

Background:

• Evaluating material (dis)similarity is key for algorithm development in materials science.
• Structural similarity metrics are vital for compound classification, chemical space exploration, and predicting material properties.
• Smooth Overlap of Atomic Positions (SOAPs) provide invariant descriptors for atomic environments, underpinning machine-learned inter-atomic potentials.

Purpose of the Study:

• To develop a unified framework for describing the similarity of crystalline, disordered, and molecular structures.
• To introduce powerful metrics for navigating alchemical and structural complexities in materials.
• To advance machine-learning techniques for predicting molecular properties.

Main Methods:

• Combining smooth overlap of atomic positions (SOAPs) with a regularized entropy match (REMatch) approach.
• Developing metrics to describe the similarity of whole molecular and bulk periodic structures.
• Utilizing a kernel derived from SOAPs and REMatch with ridge regression for property prediction.

Main Results:

• Successfully described the similarity of both molecular and bulk periodic structures within a unified framework.
• Achieved a mean absolute error below 1 kcal mol(-1) in predicting atomization energies for organic molecules.
• Demonstrated a significant milestone in applying machine learning for accurate molecular property evaluation.

Conclusions:

• The SOAPs-REMatch framework provides powerful metrics for assessing structural and alchemical similarity.
• This approach enables efficient navigation of complex materials configuration spaces.
• The method represents a significant advancement in machine-learning-driven materials discovery and property prediction.