Accurate and efficient machine learning interatomic potentials for finite temperature modelling of molecular crystals

Area of Science:

• Computational materials science
• Chemical physics
• Machine learning applications

Background:

• Machine learning interatomic potentials (MLIPs) are transforming molecular crystal modeling.
• Calculating sublimation enthalpies, crucial for crystal stability, faces challenges with data requirements and density functional theory accuracy.

Purpose of the Study:

• To develop highly accurate and efficient MLIPs for molecular crystals.
• To reduce the number of reference structures needed for MLIP training.
• To enable reliable finite temperature and pressure calculations of crystal properties.

Main Methods:

• Leveraging foundation models in chemistry and materials science.
• Utilizing quantum diffusion Monte Carlo benchmarks for accuracy.
• Generating MLIPs with a reduced dataset (approx. 200 structures).

Main Results:

• Achieved sub-chemical accuracy for MLIPs describing molecular crystals at finite temperature and pressure.
• Reduced data requirements by an order of magnitude compared to state-of-the-art methods.
• Successfully computed sublimation enthalpies for the X23 dataset, including anharmonicity and nuclear quantum effects, with experimental accuracy.

Conclusions:

• The developed framework enables accurate MLIPs for molecular crystals with unprecedented data efficiency.
• The approach is generalizable to pharmaceutical crystals like paracetamol and aspirin.
• Accurate modeling of nuclear quantum effects and ambient conditions facilitates insights into pharmaceutical and biological systems.