Single-Point Extrapolation to the Complete Basis Set Limit through Deep Learning.

Area of Science:

• Computational chemistry
• Materials science
• Quantum chemistry

Background:

• Electronic structure calculations are computationally expensive.
• Machine learning (ML) offers a way to accelerate predictions of molecular properties.
• Basis set incompleteness error (BSIE) is a significant factor in calculation accuracy.

Purpose of the Study:

• Develop and train a graph neural network (GNN) model to correct BSIE.
• Compare ML models trained on total potential versus BSIE.
• Evaluate the generalization ability of ML models on unseen molecular systems.

Main Methods:

• Trained a GNN model on ab initio data.
• Focused on correcting BSIE between small and large basis sets (cc-pVDZ and cc-pV5Z).
• Utilized RHF and B3LYP levels of theory.

Main Results:

• ML models trained to correct BSIE generalize better than those trained on total potential.
• Ensemble models improve potential behavior with insufficient training data.
• Achieved a test error of 0.184 kcal/mol for B3LYP using the cc-pV5Z potential.
• Accurately reproduced large basis set interaction energy curves on the S66x8 dataset.

Conclusions:

• ML models effectively correct BSIE, enhancing prediction accuracy.
• Model performance depends on the similarity between training data and target systems.
• ML provides a computationally feasible approach for accurate molecular property prediction.