Interpretable AI and Machine Learning Classification for Identifying High-Efficiency Donor-Acceptor Pairs in Organic...

Area of Science:

• Materials Science
• Computational Chemistry
• Artificial Intelligence

Background:

• Predicting organic solar cell (OSC) efficiency is vital for developing new donor and acceptor materials.
• Current machine learning regression models struggle to define clear design rules for high-efficiency OSCs.
• Interpretable AI offers a path to understand structure-property relationships in OSC materials.

Purpose of the Study:

• To develop a machine learning framework for identifying high-efficiency donor-acceptor pairs in OSCs based on chemical structures.
• To extract key molecular features that differentiate high- from low-efficiency OSC materials.
• To establish general design principles for novel OSC materials.

Main Methods:

• Integration of interpretable AI (Shapely values) with supervised classification models (SVM, decision trees, random forest, gradient boosting).
• Application of unsupervised machine learning (PCA with loading vectors) for validation and feature identification.
• Analysis of molecular descriptors such as van der Waals surface area and partial equalization of orbital electronegativity.

Main Results:

• Supervised models successfully identified high-efficiency donor-acceptor pairs using only chemical structures.
• Identified key molecular features, including van der Waals surface area and Moreau-Broto autocorrelation, crucial for efficiency.
• Features identified by supervised methods were a subset of those found by unsupervised analysis, confirming robustness.

Conclusions:

• The developed interpretable AI approach provides a robust method for predicting OSC material efficiency.
• The identified molecular features offer valuable design principles for creating next-generation organic solar cells.
• This framework facilitates efficient exploration of vast chemical spaces for optimal donor-acceptor materials in OSCs.