Global machine learning potentials for molecular crystals.

Area of Science:

• Computational materials science
• Solid-state chemistry
• Machine learning applications

Background:

• Accurate modeling of molecular crystals is challenging due to large unit cells.
• Polymorphs often exhibit minor energy differences (approx. 1 kJ/mol), necessitating high-accuracy computational methods.
• First-principles methods like density functional theory (DFT) provide accuracy but are computationally expensive.

Purpose of the Study:

• To develop accurate and computationally efficient machine learning interatomic potentials (MLIPs) for molecular crystals.
• To create global MLIPs applicable to any molecular crystal, overcoming limitations of system-specific models.
• To bridge the gap between the accuracy of DFT and the efficiency of classical force fields.

Main Methods:

• Training global MLIPs using existing databases of DFT calculations for molecular crystals and molecules.
• Developing MLIPs that capture interatomic interactions with high fidelity.
• Utilizing large-scale DFT data to ensure broad applicability and accuracy.

Main Results:

• The developed MLIPs achieve accuracy comparable to DFT calculations for molecular crystals.
• MLIPs demonstrate superior performance compared to traditional classical force fields.
• The potentials are validated against experimental benchmarks, confirming their reliability.

Conclusions:

• Machine learning interatomic potentials offer a computationally viable alternative for accurate molecular crystal modeling.
• These MLIPs can be used universally across different molecular crystals, facilitating broader research.
• The approach significantly reduces computational cost while maintaining high accuracy, enabling large-scale simulations.