Data-efficient machine learning for molecular crystal structure prediction.

Area of Science:

• Computational Chemistry
• Materials Science
• Machine Learning

Background:

• Machine learning (ML) combined with quantum mechanical (QM) calculations offers high accuracy at a reduced cost.
• Generating reference QM data for large periodic systems, crucial for organic crystal structure prediction (CSP), is computationally expensive.
• Accurate assessment of numerous trial structures is needed for CSP, demanding efficient and accurate methods.

Purpose of the Study:

• To develop a data-efficient ML approach for organic crystal structure prediction (CSP).
• To accurately describe intermolecular interactions, including H-bonding and many-body dispersion.
• To enable high-throughput screening of potential polymorphs with reduced computational expense.

Main Methods:

• Developed tailored Δ-ML models combining physics-based descriptions with ML.
• Enhanced long-range interactions using density functional tight binding (DFTB).
• Trained a short-range ML model on high-quality first-principles reference data.

Main Results:

• The Δ-ML models effectively screen crystal candidates, capturing subtle intermolecular interactions.
• The workflow is broadly applicable to various molecular materials.
• The method avoids the need for computationally expensive, high-level periodic calculations for reference data.

Conclusions:

• This data-efficient ML approach significantly reduces the computational cost of organic CSP.
• The method allows for the use of high-accuracy wavefunction methods in CSP.
• The presented workflow offers a powerful tool for exploring the potential polymorph landscape of molecular materials.