Machine-Learned Force Fields for Lattice Dynamics at Coupled-Cluster Level Accuracy.

Area of Science:

• Computational materials science
• Quantum chemistry
• Solid-state physics

Background:

• Accurate prediction of material properties requires sophisticated computational methods.
• Density functional theory (DFT) and coupled cluster (CC) theory are common approaches for electronic structure calculations.
• Machine-learned force fields (MLFFs) offer a computationally efficient alternative for large-scale simulations.

Purpose of the Study:

• To develop and assess machine-learned force fields (MLFFs) for carbon diamond and lithium hydride solids.
• To compare the accuracy of MLFFs trained on DFT and CC potential energy surfaces.
• To investigate methods for improving MLFFs, particularly for CC data limitations.

Main Methods:

• Training MLFFs on potential energy surfaces from DFT and CC calculations.
• Calculating phonon dispersions and vibrational densities of states (VDOS) using MLFFs.
• Employing delta-learning and charge-aware MLFF approaches to address CC data limitations.

Main Results:

• MLFFs trained on CC theory show improved accuracy in predicting vibrational frequencies compared to DFT.
• Calculated phonon dispersions and VDOS from CC-trained MLFFs exhibit better agreement with experimental and ab initio results.
• Anharmonic effects on VDOS for lithium hydride were estimated using CC-level MLFFs.

Conclusions:

• MLFFs trained on higher-level theory (CC) provide more accurate vibrational properties for solids.
• Delta-learning and charge-aware approaches enhance MLFF performance with CC data.
• MLFFs are a viable tool for studying complex vibrational phenomena, including anharmonicity.