Discovering Molecular Insights in Organic Optoelectronics with Knowledge-Informed Interpretable Deep Learning.

Area of Science:

• Computational chemistry
• Artificial intelligence in materials science
• Machine learning for molecular discovery

Background:

• Deep learning accelerates molecular screening but often acts as a black box, limiting chemical knowledge generation.
• Current models struggle to provide interpretable insights for rational molecular design.

Purpose of the Study:

• To introduce LUMIA (Learning and Understanding Molecular Insights with Artificial Intelligence), an interpretable deep learning framework.
• To enhance molecular design and accelerate chemical discovery by integrating chemical knowledge into AI models.

Main Methods:

• Developed LUMIA, integrating chemistry-informed contrastive learning and Monte Carlo tree search (MCTS).
• Pretrained LUMIA on 1.4 million organic molecules with knowledge-informed augmentations (e.g., π-conjugation, substituent effects).
• Utilized MCTS for intrinsic interpretability and uncovering substructure-property relationships.

Main Results:

• Achieved state-of-the-art performance in organic optoelectronic property prediction.
• Identified novel substructure patterns influencing reorganization energy.
• Discovered synergistic substituent effects in pyrazole derivatives, enabling rational molecular design.

Conclusions:

• Interpretable deep learning frameworks like LUMIA are crucial for advancing chemical discovery.
• Explicitly embedding chemical knowledge transforms AI's role from prediction to insight generation.
• LUMIA facilitates rational molecular design beyond existing datasets and accelerates the discovery of novel materials.