One Size Fits All? Development of the CPOSS209 Data Set of Experimental and Hypothetical Polymorphs for Testing...

Area of Science:

• Solid-state chemistry
• Computational materials science
• Crystallography

Background:

• Organic crystal structure prediction (CSP) methods are crucial for understanding material properties.
• A balance exists between theoretical rigor, computational cost, and accuracy in CSP.
• Experimental screening and computational methods are often complementary in polymorph discovery.

Purpose of the Study:

• To create a benchmark dataset of crystal structures and energies for 20 organic molecules.
• To evaluate the performance of various lattice energy modeling methods.
• To assess the applicability of machine learning models for CSP.

Main Methods:

• Generated sets of 6-15 crystal structures per molecule, including known polymorphs and CSP-derived structures.
• Employed electronic structure calculations and anisotropic atom-atom models for initial CSP.
• Reoptimized structures and calculated lattice energies using periodic dispersion-corrected density functional theory (DFT) and many-body dispersion (MBD) methods.
• Tested two Machine Learned Foundation Models (MACE-MP-0, MACE-OFF23) on the dataset.

Main Results:

• Compared lattice energies and structures from original CSP with DFT and MBD calculations.
• Illustrated challenges in modeling polymorphs and their relative energies across different organic molecules.
• Demonstrated significant variations in observed polymorphs even for similar molecules.
• Showcased the utility of the dataset for preliminary testing of modeling approaches.

Conclusions:

• The study provides a valuable dataset for validating CSP and lattice energy methods.
• Accurate prediction of organic polymorphs remains challenging, requiring careful consideration of theoretical models.
• Machine learning models show promise but require further development and validation for broad CSP applications.