Alchemical and structural distribution based representation for universal quantum machine learning.

Area of Science:

• Computational Chemistry
• Quantum Machine Learning
• Materials Science

Background:

• Developing accurate predictive models for electronic properties across diverse chemical spaces is computationally intensive.
• Existing quantum machine learning (QML) models often struggle with generalization to unseen chemical environments and elements.

Purpose of the Study:

• To introduce a novel atomic representation for creating universal kernel ridge regression-based QML models.
• To enable accurate prediction of electronic properties for a wide range of chemical compounds, including extrapolation to new elements.
• To demonstrate the model's capability in predicting covalent bonding and atomization energies.

Main Methods:

• A new atomic representation based on scaled Gaussian distribution functions, incorporating structural and elemental degrees of freedom.
• Training kernel ridge regression-based QML models on diverse chemical compound datasets.
• Utilizing "alchemical extrapolation" by interpolating across the periodic table for predicting properties of untrained elements.

Main Results:

• The QML models achieved chemical accuracy for out-of-sample compounds after training on a few thousand instances.
• Demonstrated accurate prediction of covalent bonding (single, double, triple bonds) and atomization energies in organic molecules.
• Showcased competitive predictive power for various electronic ground state properties, including polarizability, HOMO-LUMO gap, and dipole moments.

Conclusions:

• The proposed atomic representation facilitates the development of universal QML models for electronic properties.
• The "alchemical extrapolation" capability allows accurate predictions for novel chemical compositions and bonding scenarios.
• The QML models offer a computationally efficient and accurate alternative to traditional methods for materials discovery and property prediction.