An orbital-based representation for accurate quantum machine learning

Area of Science:

• Computational chemistry
• Quantum machine learning
• Materials science

Background:

• Predicting electronic properties of molecules is crucial for materials discovery.
• Existing methods can be computationally expensive or lack accuracy for diverse chemical spaces.
• Quantum machine learning (QML) offers a promising avenue for accelerating these predictions.

Purpose of the Study:

• To introduce a novel electronic structure-based representation for QML.
• To enable accurate prediction of electronic properties across chemical compound space.
• To develop flexible QML models applicable to diverse molecular species.

Main Methods:

• Utilizing computationally inexpensive ab initio calculations to construct the representation.
• Developing QML models trained on datasets like QM7b, QM7b-T, QM9, and LIBE.
• Incorporating molecular geometry, charge, and spin as input features for property prediction.

Main Results:

• The QML models demonstrated high accuracy for various property labels, including total potential energy, HOMO/LUMO energies, ionization potential, and electron affinity.
• The approach successfully accounted for molecular species with different charge and spin multiplicities.
• Accurate inference of total potential energies was achieved using geometry, charge, and spin as inputs.

Conclusions:

• The proposed electronic structure-based representation offers an accurate and flexible approach for QML in chemistry.
• This method enables efficient property prediction across vast chemical spaces and diverse molecular conditions.
• The developed QML models hold significant potential for accelerating materials design and discovery.