Machine Learning Force Field for Optimization of Isolated and Supported Transition Metal Particles.

Area of Science:

• Computational chemistry
• Materials science
• Catalysis research

Background:

• Computational modeling is crucial for advancing catalysis research.
• Developing accurate and cost-efficient simulation methods is an ongoing challenge.
• Machine learning techniques are increasingly used to derive interatomic potentials from ab initio data.

Purpose of the Study:

• To develop an energy-free machine learning calculator for predicting energies and atomic forces in metallic systems.
• To improve the accuracy and reduce the computational cost of simulations in catalysis.
• To investigate the application of combined neural networks for interatomic potential prediction.

Main Methods:

• Developed an energy-free machine learning calculator using three individually trained neural networks.
• Employed a graph neural network to predict atomic energies, achieving a mean absolute error (MAE) within 0.004 eV compared to density functional theory (DFT).
• Utilized two feed-forward networks for predicting atomic force norm and direction, achieving a MAE within 0.080 eV/Å against DFT.

Main Results:

• Successfully predicted atomic energies and forces for monometallic Pd nanoparticles, bimetallic AuPd nanoalloys, and supported Pd metal crystallites on silica.
• Demonstrated high accuracy in energy predictions (MAE < 0.004 eV) and force predictions (MAE < 0.080 eV/Å) relative to DFT.
• Showcased the interpretability of the graph neural network by revealing the physics of cohesion energy in a monometallic particle.

Conclusions:

• The developed machine learning calculator offers a computationally efficient and accurate alternative to traditional DFT methods for catalysis simulations.
• The approach of combining multiple neural networks effectively captures complex interatomic interactions.
• The method provides insights into the underlying physics of material properties, enhancing the understanding of catalytic processes.