Predicting Molecular Energies of Small Organic Molecules With Multi-Fidelity Methods.

Area of Science:

• Computational Chemistry
• Machine Learning
• Quantum Chemistry

Background:

• Machine learning (ML) is increasingly used for predicting quantum chemical properties.
• Multi-fidelity methods like kappa-ML and Multifidelity Machine Learning (MFML) reduce computational costs for training data generation.

Purpose of the Study:

• Implement and analyze multi-fidelity ML methods for predicting electronic molecular energies at the DLPNO-CCSD(T) level.
• Evaluate model accuracy and efficiency in training data generation time-cost.
• Assess model performance on diverse molecular datasets, including atmospherically relevant, isomeric, and complex conjugated molecules.

Main Methods:

• Implementation and analysis of multi-fidelity ML methods (kappa-ML, MFML).
• Prediction of electronic molecular energies at the DLPNO-CCSD(T) level.
• Evaluation of models on small organic molecules and a public dataset.

Main Results:

• Multi-fidelity methods offer a significant reduction in computational cost for generating training data.
• Models demonstrate accuracy in predicting electronic molecular energies.
• Performance is assessed across various molecular types, including atmospherically relevant and complex structures.

Conclusions:

• Multi-fidelity ML methods are efficient and accurate for predicting quantum chemical properties.
• These methods show promise for applications involving complex and atmospherically relevant molecules.
• The study provides a comprehensive analysis of model performance based on accuracy and computational cost.