AI model embedded with chemical knowledge enables efficient experimental search of highly fluorescent covalent organic frameworks (COFs) to avoid exhaustive synthetic screening.

On Oct 31, a collaborative team of Chinese scientists published a groundbreaking study in Nature Chemistry, presenting an AI-assisted iterative experiment-learning paradigm that provides a compelling solution to this challenge. 

This work was the result of close cooperation between Professor Hexiang Deng (Wuhan University), Professor Jun Jiang & Linjiang Chen (the University of Science and Technology of China), and Professor Ben Zhong Tang (the Hong Kong University of Science and Technology), with Liang Zhang, Jiahui Du, and Zikai Xie serving as the co-first author.

This innovative approach, known as interactive experiment-learning evolution, enhances the synergy between experienced researchers and artificial intelligence (AI) models, allowing for continuous optimization and evolution.

The team applied this strategy to discover two-dimensional (2D) fluorescent COFs, synthesized via the imine condensation of 20 amines and 26 aldehydes.

A multilayer perceptron (MLP) was pre-trained using DFT calculations on orbital-level descriptors of 520 dimers and building blocks. Then, a Siamese neural network was employed to recommend new candidate COFs in successive learning cycles. The team identified a COF with a photoluminescence quantum yield (PLQY) of 41.3 percent after just four generations of experiment-model iteration.

A key innovation of this method is the incorporation of electronic structure information, such as frontier orbital energies and excited-state charge distributions, into the learning process. This integration enables more meaningful, chemically insightful predictions.

Analysis of the dataset revealed that fluorescence is influenced by HOMO–LUMO energy-level matching between building blocks and by the excited-state charge distribution within dimers.

These insights enabled the team to elucidate the structure-performance relationship of COFs and uncover underlying mechanisms, providing a transferable framework for material discovery that merges chemical knowledge with data-driven modeling.

The team's approach embeds chemical insights, including frontier orbital theory and electronic configurations, into the machine learning process. This allows the model to transcend mere statistical fitting and make predictions informed by chemical understanding, thereby enhancing interpretability.