Researchers from SDK, AIST, and ADMAT conducted AI-based searches for polymers with desired properties, aiming to demonstrate the effectiveness of AI technology in the process of polymer design. As a model case, they focused on glass transition temperature, an index of heat resistance.
Using 417 different types of structural data on polymers with known structures and glass transition temperatures, they conducted an AI-based search for a polymer with the highest glass transition temperature to see whether it is possible to shorten the development cycle.
First, randomly selected 10 sets of data were supplied as training data for AI. The Extended Connectivity Circular Fingerprints (ECFP) method was applied to the training data, digitizing structural features of polymers.
Then, using Bayesian optimization*, the researchers made repeated efforts to predict and verify a polymer with the highest glass transition temperature out of the remaining 407 sets of data. Thus, they checked the number of trials required until the discovery of a target polymer. To prevent the influence of the choice of data on the results, 500 examinations were conducted with different sets of initial training data, and the average number of trials was evaluated.
As a result, the researchers succeeded in discovering a target polymer with the highest glass transition temperature with an extremely small number of trials, namely, 4.6 times of trials on the average. This figure is about one-fortieth of the number of trials required under random selection of polymers, confirming the effectiveness of AI-based polymer design.
To build AI, it is necessary to convert the features of polymers into numerical values. By applying ECFP (a method for representing monomer structures) to this development work, the researchers found repeating units of molecules (such as functional groups) extracted automatically and structural features expressed appropriately in numerical vectors.
Using AI built on these data, the researchers realized high-precision property predictions at the speed of 0.25 second per polymer. Thus, it became possible to make comprehensive property predictions for a large number of candidate polymers within a limited amount of time.
Furthermore, by using Bayesian optimization in prediction, the researchers discovered a polymer with the highest glass transition temperature out of about 400 types of candidate polymers, based on 13.6 sets of training data.
Prior to this study, it was believed that AI-based predictions with a small number of training data would be less accurate, and that a large amount of training data would be required.
However, this development work suggests that AI would be effective in solving problems even in the area of developing most-advanced functional materials, where only a limited amount of training data seems to be available.
*Bayesian optimization: A method for selecting the next candidate material to be examined by considering estimated prediction error, in addition to predicted values.