Beyond independent error assumptions in large GNN atomistic models.

Area of Science:

• Computational Chemistry
• Materials Science
• Machine Learning

Background:

• Relative energy difference calculations are crucial for adsorption energy, catalyst screening, and reaction energy prediction.
• Density Functional Theory (DFT) effectively cancels systematic errors, but Graph Neural Network (GNN) accuracy for relative energies is uncertain.

Purpose of the Study:

• To evaluate the accuracy of GNNs in calculating relative energy differences compared to DFT.
• To investigate the nature of errors in GNN predictions and explore methods for error reduction.

Main Methods:

• Analysis of approximately 483 million DFT and GNN energy difference pairs from the Open Catalyst 2020-Dense dataset.
• Introduction of the subgroup error cancellation ratio to quantify error cancellation in chemically similar pairs.
• Proposal of the marginal correct sign ratio to assess GNN performance in distinguishing subtle energy differences.

Main Results:

• GNNs exhibit correlated errors that can be mitigated through subtraction, leading to more precise predictions.
• State-of-the-art GNNs achieve up to 77% error reduction in chemically similar subgroups, rivaling DFT's error cancellation.
• Local embedding similarity in GNNs correlates with error cancellation magnitude, suggesting improved training strategies.

Conclusions:

• GNNs demonstrate significant potential for accurate relative energy difference calculations, comparable to DFT.
• Error cancellation in GNNs enables higher accuracy and the ability to discern subtle energy variations.
• Future GNN development should focus on enhancing embedding similarity for chemically similar systems to maximize error cancellation.