A lot of effort has been made over the past few years to develop various deep learning technologies in the field of materials screening and design. By the way, materials synthesis is another major barrier in materials development as it is still carried out laboriously by human researchers. Therefore, several groups are trying to automate the synthetic process by combining novel robots and artificial intelligences (AIs). And one of the essential AI technologies is the optimization of various reaction conditions, such as reagents, concentrations, temperature, and time, to obtain maximum yield with minimal resources. Traditionally, a number of active learning algorithms such as Bayesian optimization (BO) and random forest have been applied to solve this multi-armed bandit problem by adopting the strategies of exploitation and exploration. However, they relied only on real-time data generated from their experimental setups to train and update the active learning models. Generally, it takes a long time to accumulate enough data and eventually to reach the optimal condition. To overcome the issue, a hybrid-type dynamic optimization (HDO) algorithm that utilizes both previous literature and current experimental data is proposed. Unlike other optimization techniques, HDO algorithm uses two acquisition functions and recommends the most favorable experimental condition by comparing the relative advantages of exploitation and exploration. The first acquisition function is based on the deep neural network (DNN) to predict the reaction yield trained with abundant literature data. It has characteristics that emphasize the exploitation by selecting conditions with high yield but low uncertainty. The other acquisition function is based on conventional BO multiplied by user-defined weight and trained using current experimental results. The DNN extracts valuable information from rich existing data and applies it to current optimization problem when the reliability is guaranteed. And BO complements the weaknesses of the DNN model when previous data lack relevant knowledge and current experimental data can provide more powerful information. In the evaluation test, HDO showed optimization speeds 2.7 times and 1.3 times faster than random search and BO, respectively. We expect the proposed methodology to be a useful tool not only in autonomous synthesis systems but also in manual synthesis experiments by chemists. 

Hybrid-type dynamic optimization of reaction conditions for organic synthesis

Macromolecular Chemistry: The Second Century

[Author Benefits](https://scimeetings.acs.org/?utm_source=pubs_content_marketing&utm_medium=website&utm_campaign=0320_MCF_NPI_Launch_Spring_Homepage&ref=pubs_content_marketing)

[How to use SciMeetings](https://storage.googleapis.com/dl.morressier.com/acs/SciMeetings-ACSSpring2020-HowToGuide_M3020-Final.pdf)

[Frequently Asked Questions](https://storage.googleapis.com/dl.morressier.com/acs/SciMtng_ACSNMSp2020_FAQ_M3120_Final.pdf)

ACS Spring 2020 National Meeting & Expo

The official archive and repository for all ACS Meetings & Expositions abstracts and select posters or presentations.