This project aims to generate "Artificial Intelligence in Chemical Reaction Design Discovery" (AICReDD) that predicts "the whole picture of the behavior of atoms" in chemical reactions and suggests useful and unknown chemical reactions one after another. This will be done by integrating technologies in computational chemistry, information science, and materials informatics. The following is an overview of “Artificial Intelligence in Chemical Reaction Design and Discovery” (AICReDD) this project envisions.

Development of Reaction Path Database

The foundation of AICReDD is a reaction path database based on quantum chemical calculations on computer. The reaction path database can be developed using the AFIR method implemented in GRRM, the world's first versatile quantum chemical calculation-based automated reaction path search program. The AFIR method exhaustively predicts the possible isomers of the chemical composition, the decomposition and formation pathways of each isomer, and the isomerization pathways between isomers based on the information of potential energy obtained by quantum chemical calculations from the input chemical composition. Using this method, we will run an automated reaction path search by feeding various chemical compositions and make a database from the resulting reaction paths. Further, a web-based data analysis platform will be built, where the obtained database can be shared and used for various purposes.

Demonstration Experiments of Unknown Chemical Reactions Suggested by Reaction Path Database

We will also conduct experiments with chemical reactions suggested by the database. The chemical reaction database derived from quantum chemical calculations contains novel reactions that are beyond human knowledge; therefore, it is very valuable to actually experiment with and realize them. Our expectation is that some of these reactions will be innovative ones that will benefit humanity. Demonstration experiments also serve to combine real-life chemistry with the virtual database built solely on calculations.

Furthermore, a robotic system designed for organic synthesis will be introduced, which is expected to greatly increase the efficiency of experiments. In the second half of the project, we will also work on research to combine the AICReDD with the robotic system. In other words, a new system will be developed where the chemical reaction suggested by AICReDD is fed into the robotic system to conduct a series of experiments. As a sub-project, we will also attempt to use the robotic system to remotely conduct a part (or most) of our experimental chemical research. This will revolutionize the management style of experimental chemistry laboratories during the epidemic of COVID-19.

Building a System Which Selects Calculation Targets and Automatically Grows the Database

To attain the above goals, how to select the chemical composition for calculation will determine the value of the database. In the first half of the project, organic chemists with extensive knowledge of chemistry will select calculation targets. However, in order for the AICReDD to be truly practical, data diversity is the key. In other words, we need a system to systematically select calculation targets from a vast chemical space, which is too large to tackle with a simple approach. Our project will resolve this challenge by using combinatorial optimization technique. Ultimately, we will develop a system to grow the database by itself, where the next calculation targets are selected automatically. Building a System Which Makes Proposals of Social Needs Finally, the following question needs to be answered: what should we suggest from the database? The answer should be based on an overall assessment of various factors such as the inclusion of novel chemical transformations, easy availability of raw materials, and high value of products, etc. Therefore, we aim to build a system making proposals that meet social needs by using technologies like machine learning and referring to catalogs of known reactions and source molecules.

Into a New Realm of Prediction and Control

In addition to the above objectives, we will work on refining and generalizing the element technologies such as the automated reaction path search method and combinatorial optimization technique, which are the bases of this project.

We envision a future when "chemistry" will evolve from a study mainly relying on trial-and-error method into a new discipline that can afford prediction and control. Then, it will be possible to customize molecules that will be used as the building blocks of pharmaceuticals and devices. Our hope is that the AICReDD generated by this project will serve as the foundation for such a future.