Machine-Learning Interatomic Potentials Achieving CCSD(T) Accuracy for Systems with Extended Covalent Networks and...

Area of Science:

• Computational materials science
• Quantum chemistry
• Machine learning

Background:

• Machine-learning interatomic potentials (MLIPs) offer accurate, cost-effective atomistic simulations.
• CCSD(T) calculations provide high accuracy but are computationally expensive, especially for periodic systems.
• Existing MLIPs struggle with extended covalent networks and van der Waals (vdW) interactions.

Purpose of the Study:

• To develop a methodology for training MLIPs with CCSD(T) accuracy for extended covalent networks.
• To enable accurate large-scale atomistic simulations of systems with significant vdW interactions.
• To apply the developed MLIPs to analyze the structure and properties of covalent organic frameworks (COFs).

Main Methods:

• Utilized a Δ-learning approach with a dispersion-corrected tight-binding baseline.
• Trained MLIPs on the energy differences between CCSD(T) and the baseline.
• Incorporated vdW-bound multimers into the training set to capture dispersion interactions.
• Combined vdW-aware tight-binding with a local MLIP for CCSD(T)-level accuracy.

Main Results:

• Achieved root-mean-square energy errors below 0.4 meV/atom on training and test sets.
• Accurately reproduced electronic total atomization energies, bond lengths, vibrational frequencies, and intermolecular interactions.
• Successfully applied the method to a quasi-2D COF, analyzing its structure, interlayer binding, and hydrogen absorption at CCSD(T) accuracy.

Conclusions:

• The developed methodology provides a practical route for large-scale atomistic simulations of extended covalent networks with vdW interactions.
• This approach achieves chemical accuracy, significantly advancing the capabilities of computational materials science.
• Enables in-depth analysis of complex materials like COFs, previously limited by computational constraints.