Predicting molecular dipole moments by combining atomic partial charges and atomic dipoles

Area of Science:

• Computational Chemistry
• Machine Learning in Quantum Mechanics
• Molecular Property Prediction

Background:

• Molecular dipole moment (μ) is crucial for predicting spectra and electrostatic interactions.
• Extracting μ from quantum mechanical calculations is computationally intensive.
• Machine learning (ML) offers a potential solution for efficient μ prediction.

Purpose of the Study:

• To develop a physically informed ML model for accurate and efficient molecular dipole moment prediction.
• To combine local atomic polarization and global charge movement effects within a unified ML framework.
• To benchmark the model's performance against high-level quantum chemical methods.

Main Methods:

• Developed 'MuML' models combining symmetry-adapted Gaussian process regression for local polarization and scalar partial charges for global charge movement.
• Trained models on the QM7b dataset using coupled-cluster theory and density functional theory (DFT) computed dipole moments.
• Utilized a calibrated committee model for reliable uncertainty estimation.

Main Results:

• The combined MuML model achieved high accuracy in reproducing molecular dipole moments.
• Demonstrated excellent transferability to larger, more complex molecules, approaching DFT accuracy at reduced computational cost.
• Showcased the scalar model's superiority for large molecules dominated by charge separation.

Conclusions:

• Physics-based ML models integrating local and non-local effects significantly improve dipole moment prediction.
• The developed model offers a computationally efficient alternative to traditional quantum chemical methods.
• Highlights the need to account for both local and non-local contributions for accurate molecular dipole moment modeling, especially for condensed phases.