Quantum Deep Descriptor: Physically Informed Transfer Learning from Small Molecules to Polymers.

Area of Science:

• Computational Materials Science
• Quantum Chemistry
• Machine Learning

Background:

• Materials informatics (MI) requires efficient methods for predicting material properties.
• Existing descriptors may not fully capture essential quantum-chemical characteristics.
• Transfer learning offers a promising avenue for leveraging pre-trained models.

Purpose of the Study:

• To develop a physically informed transfer learning approach for materials informatics.
• To introduce a quantum deep descriptor (QDD) derived from a quantum deep field (QDF) model.
• To demonstrate the QDD's effectiveness in predicting polymer properties.

Main Methods:

• A quantum deep field (QDF) model was trained using density functional theory (DFT) on a large molecular property database.
• The pre-trained QDF model generated a quantum deep descriptor (QDD) encoding quantum-chemical features.
• The QDD was applied to predict polymer properties like band gap and dielectric constant.

Main Results:

• The QDD effectively encodes fundamental quantum-chemical characteristics of molecules.
• QDDs pre-trained on small molecules successfully predicted properties of polymers.
• Performance was comparable or superior to existing descriptors for specific property predictions.

Conclusions:

• The proposed DFT-based, physically informed transfer learning approach is valuable for materials informatics.
• The QDD provides quantum-chemical insights into materials.
• The method facilitates efficient materials discovery and property prediction.