Accelerating structure prediction of molecular crystals using actively trained moment tensor potential

Area of Science:

• Computational Chemistry
• Materials Science
• Crystallography

Background:

• Machine-learned interatomic potentials (MLIPs) show promise for inorganic crystal structure prediction.
• Predicting molecular crystal structures and polymorphs remains computationally challenging.

Purpose of the Study:

• To develop and validate a machine learning methodology for accelerating molecular crystal structure prediction.
• To assess the efficacy of Moment Tensor Potentials (MTP) combined with active learning for polymorph ranking.

Main Methods:

• Utilized Moment Tensor Potentials (MTP) for interatomic interactions.
• Employed an active learning strategy based on the maxvol algorithm for efficient data selection.
• Tested the methodology on benzene and glycine crystal structure prediction.

Main Results:

• The developed MTP accurately reproduced density-functional theory (DFT) results for benzene and glycine polymorphs.
• The active learning approach effectively guided the potential training process.
• The methodology demonstrated good agreement with DFT in ranking polymorphs.

Conclusions:

• Moment Tensor Potentials (MTP) combined with active learning offer a powerful approach to accelerate molecular crystal structure prediction.
• This methodology can significantly enhance computationally guided polymorph searches.
• The findings suggest broader applicability of MLIPs in molecular materials discovery.