Accelerated Organic Crystal Structure Prediction with Genetic Algorithms and Machine Learning.

Area of Science:

• Computational chemistry
• Materials science
• Crystallography

Background:

• Organic crystal structure prediction (CSP) is crucial for understanding material properties.
• Accurate CSP requires efficient methods for screening and optimizing crystal candidates.

Purpose of the Study:

• To present a high-throughput, end-to-end pipeline for organic crystal structure prediction.
• To leverage neural network potentials for efficient screening and structural relaxation.
• To assess the performance of different stages of the pipeline.

Main Methods:

• A two-stage pipeline combining random search and genetic algorithm (GA) optimization.
• Use of neural network potentials for energy calculations.
• Validation on 21 molecules from CSP blind tests.

Main Results:

• Random search alone achieved ≈50% success rate.
• The full pipeline with GA optimization found matches for ≈80% of targets.
• The GA approach required 10-100 times smaller initial populations.
• A pipeline using an ANI model achieved ≈60% success rate.

Conclusions:

• The developed pipeline significantly improves the efficiency and success rate of organic CSP.
• Machine learning models can approach the accuracy of ab initio methods while maintaining efficiency.
• This tool has the potential to accelerate materials discovery.