Search for Correlations Between the Results of the Density Functional Theory and Hartree-Fock Calculations Using...

Area of Science:

• Computational Chemistry
• Materials Science
• Machine Learning

Background:

• Supramolecular structures are crucial in various chemical and biological processes.
• Accurate prediction of their properties is essential for understanding and designing new materials.
• High-level quantum chemistry calculations are computationally expensive.

Purpose of the Study:

• To develop machine learning models for predicting high-level quantum chemistry (B3LYP-D4/def-TZVP) outputs from computationally cheaper (HF-3c) inputs.
• To evaluate the performance of different machine learning models in predicting various quantum chemical descriptors.
• To identify the most predictable properties and the best-performing models for supramolecular systems.

Main Methods:

• A dataset of 1031 supramolecular structures (dimers, trimers, tetramers) was compiled.
• Six quantum chemistry descriptors were calculated using both HF-3c and B3LYP-D4/def-TZVP methods: Gibbs energy, electronic energy, entropy, enthalpy, dipole moment, and band gap.
• Linear, tree-based, and neural network machine learning models were trained and evaluated.

Main Results:

• Statistical analysis revealed good correlations for energy-related properties (Gibbs energy, electronic energy, entropy, enthalpy) between the two methods.
• Dipole moment and band gap showed weaker correlations.
• The best predictive models were LASSO (linear), XGBoost (tree-based), and single-layer perceptron (neural networks), with energy properties achieving the highest prediction accuracy.

Conclusions:

• Machine learning models can effectively predict B3LYP-D4/def-TZVP properties from HF-3c outputs for supramolecular structures.
• Energy-related properties are more amenable to accurate prediction than dipole moments.
• This approach offers a computationally efficient alternative for estimating high-level quantum chemistry results.