Learning Molecular Conformational Energies Using Semilocal Density Fingerprints.

Area of Science:

• Computational Chemistry
• Machine Learning
• Quantum Mechanics

Background:

• Density Functional Theory (DFT) relies on electron density and its gradients.
• Current DFT methods have limitations in accuracy for certain molecular properties.
• Machine learning (ML) offers potential for improved accuracy in chemical predictions.

Purpose of the Study:

• To develop a general theoretical framework for transforming semilocal electron density descriptors into ML-compatible feature vectors.
• To introduce the semilocal density fingerprint (SLDF) descriptor for learning molecular conformational energies.
• To assess the accuracy and transferability of ML models based on SLDF.

Main Methods:

• Developed a framework to convert electron density (ρ(r)) and its gradients (∇ρ(r)) into fixed-size feature vectors.
• Introduced the semilocal density fingerprint (SLDF) descriptor.
• Trained ML models using SLDF features to predict molecular conformational energies.
• Evaluated model performance on benchmark datasets and tested transferability to unseen molecules and chemical systems.

Main Results:

• SLDF-based ML models achieved >100 times greater accuracy than semilocal DFT for conformational energies.
• Predictions approached spectroscopic accuracy (≈1 cm⁻¹).
• Demonstrated significant transferability, achieving 10-fold higher accuracy for unseen molecules.
• Successfully corrected DFT's description of the oxirene potential energy surface without prior training data.

Conclusions:

• The SLDF descriptor provides a robust and accurate way to represent electron density information for ML.
• SLDF-based ML models significantly outperform semilocal DFT in predicting conformational energies.
• The developed framework shows excellent transferability, enabling accurate predictions for new chemical systems and molecules.