Machine learning transferable atomic forces for large systems from underconverged molecular fragments

Area of Science:

• Materials Science
• Computational Chemistry
• Condensed Matter Physics

Background:

• Machine learning potentials (MLPs) offer a cost-effective alternative to electronic structure calculations for atomistic simulations.
• Current MLPs often require computationally expensive, large-scale systems to generate accurate reference forces for training.
• Training MLPs typically utilizes both energies and forces from reference calculations.

Purpose of the Study:

• To develop a method for training MLPs using smaller molecular fragments, bypassing the need for large, computationally intensive systems.
• To demonstrate the transferability of MLPs trained on fragment data to extended systems.

Main Methods:

• Utilizing density-functional theory (DFT) calculations on molecular fragments to obtain reference data.
• Training second-generation high-dimensional neural network potentials (HDNNPs) using fragment-derived energies and forces.
• Validating the transferability of the trained HDNNPs on extended systems, specifically metal-organic frameworks (MOFs).

Main Results:

• MLPs trained on small molecular fragments can accurately reproduce forces in extended systems.
• The developed method significantly reduces the computational cost associated with generating training data for MLPs.
• The trained potentials exhibit excellent transferability, demonstrating the viability of the approach for realistic condensed phase environments.

Conclusions:

• Training MLPs on molecular fragments is a computationally efficient strategy for achieving first-principles accuracy.
• This approach overcomes the limitations of large system requirements for force calculations in MLP training.
• The demonstrated success with MOFs highlights the broad applicability of this method in materials science and computational chemistry.