Traditional material development and synthesis face some huge challenges. In particular, some materials with complex components or structures, such as organic molecules, macromolecules, and organic-inorganic hybrid materials, because of the huge tunability of their chemical spaces and synthetic pathways, often require huge trial and error costs. And in this process, very experienced and chemically intuitive experimenters are needed to design and complete the experiment. Therefore, the development of new tools can help to efficiently and quickly explore the synthesis reaction space. Recently, due to the rise of machine learning, it has become a new research hotspot. The synthesis of metal-organic nanocapsules is one of them. Because of its excellent performance in the fields of catalysis, gas adsorption and separation, sensors, etc., it has aroused great interest from researchers. However, as with the synthesis of other materials, the synthesis of metal-organic nanocapsules still relies on cumbersome, complex and inefficient trial and error paths.
Recently, the Lin Jian group of the Department of Mechanical Engineering at the University of Missouri and the group of Jerry L. Atwood, a professor of organic chemistry at the Department of Chemistry, proposed a machine learning algorithm to assist material synthesis. The algorithm uses the existing experimental data (including successful and failed experiments) to successfully predict the crystallization of metal organic nanocapsules under a given reaction condition (accuracy rate is greater than 90%), thereby greatly reducing the process of trial and error The investment in human and material resources generated in the process shortens the discovery cycle of new metal organic nanocapsules. The most important thing is that the algorithm can help to extract the hidden information of the material synthesis, thus helping to cultivate the experimenter's chemical intuition. Relevant research results were published in the Journal of the American Chemical Society, entitled "Machine Learning Assisted Synthesis of Metal-Organic Nanocapulses".
First, the researchers used the 486 experimental data compiled from the experimental notebook as the original data set, containing 193 records of the reaction product with single crystal (marked as 1) and 293 records without any reaction or precipitation (marked 0), see the picture above. Based on personal experience and literature reading, we identified 17 chemical characteristics that may affect the crystallization of metal organic nanocapsules, and divided the above original data set into a training set and a test set according to a 7/3 ratio.
The researchers compared nine different machine learning algorithms and found that the XGBoost algorithm showed the highest prediction accuracy rate of 91% and F1 test value of 87%, and also had the highest AUC value, recall rate and accuracy rate, see the above figure.
Secondly, through the characteristic importance function of XGBoost, the researchers found that in the preparation of metal-organic nanocapsules, reagents, organic ligands, modifiers and cations are the most important factors affecting the formation of single crystals, as shown in the figure above. In addition, the researchers also found that even if the number of chemical features was reduced from 17 to 6, the XGBoost algorithm still showed extremely high robustness.
Through a detailed study of XGBoost's decision-making process, the researchers summarized three possible paths for preparing single crystals of metal organic nanocapsules, as shown in the figure above. Researchers can work out appropriate reaction conditions based on the valence and radius of the metal cation.
Finally, the researchers designed a total of 20 experiments in three categories to verify the three chemical hypotheses. The results show that XGBoost has higher prediction accuracy than the researchers. At the same time, a new type of metal-organic nanocapsule single crystal SCP-4 was discovered, which was formed by connecting two different nanocapsule units to each other, as shown above.
The significance of this research is that for the first time, machine learning algorithms are used to synthesize metal-organic nanocapsules, which can not only reduce the number of synthetic reactions to reduce the investment of manpower and material resources, but also deeply analyze the chemical inspiration behind the reaction conditions to guide the next research direction . This method can also be extended to the synthesis and discovery of other organic and inorganic compounds by changing the chemical characteristics of the machine learning algorithm. At the same time, the machine learning algorithm and high-throughput synthesis will also bring unlimited possibilities for the discovery and development of compounds.
Article link: https://pubs.acs.org/doi/10.1021/jacs.9b11569