Machine-Learned Potentials by Active Learning from Organic Crystal Structure Prediction Landscapes.

Area of Science:

• Computational chemistry
• Materials science
• Crystallography

Background:

• Accurate energy ranking of potential organic crystal structures is crucial but computationally expensive.
• High-level density functional theory (DFT) methods are accurate but too slow for exploring large crystal energy landscapes.
• Empirical force fields offer speed but lack the accuracy needed for reliable structure ranking.

Purpose of the Study:

• To investigate active learning for training machine-learned interatomic potentials (MLIPs) for organic crystal structure prediction (CSP).
• To develop a highly automated workflow combining active learning with CSP sampling methods.
• To improve the efficiency and accuracy of ranking crystal structure energies.

Main Methods:

• Active learning strategies were employed to train MLIPs using CSP datasets.
• A hierarchical approach was used, starting with force fields and progressing to MLIPs.
• On-the-fly training within Monte Carlo simulations was utilized to model structures beyond energy minima.

Main Results:

• The developed MLIPs, trained via active learning, achieved near-DFT accuracy in reranking large, diverse crystal structure landscapes.
• The automated workflow efficiently explored a wide range of crystal packing space.
• MLIPs significantly reduced the need for expensive DFT calculations, improving computational efficiency.

Conclusions:

• Active learning-trained MLIPs offer a computationally efficient and accurate alternative for organic CSP.
• This method enhances the reliability of energy rankings in crystal structure prediction.
• The approach is extendable for modeling structures far from lattice energy minima.