How graph neural network interatomic potentials extrapolate: Role of the message-passing algorithm.

Area of Science:

• Materials Science
• Computational Chemistry
• Machine Learning

Background:

• Graph neural network interatomic potentials (GNN-IPs) show promise for materials modeling.
• Universal GNN-IPs trained on crystalline data often extrapolate well to surfaces and amorphous structures.
• The theoretical basis for this extrapolation capability remains unclear.

Purpose of the Study:

• To provide a theoretical explanation for the extrapolation capabilities of GNN-IPs.
• To investigate how GNN-IPs learn and predict interactions in untrained geometries.
• To identify factors influencing the extrapolation performance of universal potentials.

Main Methods:

• Demonstrated GNN-IPs' ability to capture non-local electrostatic interactions via message-passing.
• Tested GNN-IP models (SevenNet, MACE) on toy models and DFT data for electrostatic force prediction.
• Analyzed the extrapolation of Coulomb and kinetic energy terms in universal GNN-IPs.
• Investigated the impact of hyperparameters on extrapolation performance.

Main Results:

• GNN-IPs accurately predict electrostatic forces in untrained domains, learning the Coulomb interaction's functional form.
• The ability to learn non-local electrostatics, combined with GNN embedding, explains extrapolation.
• SevenNet-0 successfully extrapolates Coulomb interactions but not kinetic energy forces in untrained domains.
• Hyperparameters significantly impact extrapolation performance, with specific limitations identified.

Conclusions:

• GNN-IPs' extrapolation ability stems from learning non-local electrostatic interactions.
• The embedding nature of GNNs further contributes to this phenomenon.
• Understanding these mechanisms is key to developing more robust universal interatomic potentials.