MACE-OFF: Short-Range Transferable Machine Learning Force Fields for Organic Molecules

Area of Science:

• Computational Chemistry
• Materials Science
• Biophysics

Background:

• Classical empirical force fields have limitations in accuracy and transferability for predictive modeling.
• Existing methods struggle with first-principles simulations of complex molecular systems.

Purpose of the Study:

• Introduce MACE-OFF, a novel series of short-range transferable force fields for organic molecules.
• Demonstrate the capability of machine learning force fields for accurate molecular simulations.

Main Methods:

• Developed MACE-OFF using state-of-the-art machine learning and high-level quantum mechanical reference data.
• Validated MACE-OFF on diverse gas- and condensed-phase properties, including molecular crystals, liquids, and peptides.
• Incorporated quantum nuclear effects for enhanced accuracy.

Main Results:

• MACE-OFF accurately predicts gas- and condensed-phase properties of molecular systems.
• Achieved accurate and easy-to-converge dihedral torsion scans for unseen molecules.
• Successfully simulated free energy surfaces, peptide folding dynamics, and protein dynamics.

Conclusions:

• MACE-OFF enables first-principles simulations of molecular systems with high accuracy.
• The developed force fields offer a relatively low computational cost for advanced simulations.
• Facilitates broader adoption of predictive molecular modeling in chemistry and related fields.