Investigation of Data Set Portability on Various Machine-Learned Interaction Potentials for Pyrophyllite Clay

Area of Science:

• Computational Materials Science
• Machine Learning in Physics
• Quantum Chemistry

Background:

• Neural network potentials (NNPs) are increasingly used for simulating material properties.
• Accurate descriptors are crucial for capturing atomic interactions in high-dimensional machine-learned interaction potentials (MLIPs).
• Understanding the impact of architecture and descriptors on NNP accuracy is essential for reliable predictions.

Purpose of the Study:

• To investigate the influence of architecture and descriptors on the accuracy of MLIPs.
• To compare the performance of different MLIPs against reference calculations.
• To assess the impact of adding dispersion corrections a posteriori.

Main Methods:

• Developed and trained four MLIPs using atom-centered symmetry functions with varying descriptors (embedding, attention masks, message passing).
• Validated MLIPs against a representative dataset and compared with PBE-D3 reference results.
• Applied Grimme's D3 dispersion correction a posteriori to an MLIP trained on PBE data.

Main Results:

• All investigated MLIPs accurately reproduced PBE-D3 energy and forces, regardless of descriptor choice.
• Similar accuracy was observed for structural parameters, exfoliation energy, and vibrational spectra across different MLIPs.
• The a posteriori D3 correction improved accuracy for elastic and exfoliation energies and enhanced stability.

Conclusions:

• The construction of a representative dataset is paramount for achieving desired accuracy in MLIPs.
• MLIPs, when trained on appropriate data, can reliably predict various material properties.
• A posteriori dispersion correction is an effective strategy to enhance MLIP accuracy for specific properties.