A Practical Application of Machine Learning for the Development of Metallole-Based Fluorescent Materials.

We have built a prediction model of the fluorescence quantum yields of metalloles. Based on the suggestion by the prediction model, we synthesized 10 fluorescent molecules to confirm the prediction accuracy. By measuring the fluorescence quantum yields of the synthesized molecules, it was demonstrated that our prediction model reasonably classified the quantum yields with an accuracy of 0.7. In particular, the low quantum yields were perfectly predicted for the synthesized molecules, demonstrating the usefulness of our prediction model to screen out weakly fluorescent molecules from the candidates. On the other hand, the low precision of 0.5 was attributed to the bias in the training dataset containing many fluorine-containing molecules with high quantum yields. Our prediction model was then revised with the generator of candidate molecular structures for more efficient development of fluorescent materials with taking the applicability domain into account, and the improvement of the applicability was confirmed owing to the increment of the dataset.