(154g) High Strength Polymer Composites Design for Lightweight Vehicles Via Stacked Transfer Learning and Explainable Artificial Intelligence

According to the global low-carbon policy, the need to develop lightweight materials for the purpose of improving vehicle fuel efficiency is increasing. Among them, polymer composites are used for about 12% of the interior and vehicle frame of vehicles, and the application parts are gradually expanding based on excellent moldability and lightness. Accordingly, it is required to develop a high-strength polymer composites that compensates for the disadvantages of plastics that are weak in strength.

When designing materials, the mechanical properties (tensile strength, ductility) of the polymer composites are important metrics, and the tendency of mechanical properties depending on the composition can be confirmed. However, the complex correlation according to the type of polymer and reinforcement is not clear. Accordingly, material designers rely only on experience and intuition to design polymer composites, which consumes a lot of time and workforce. Most of the previous studies used machine learning to predict the mechanical properties of polymers or reinforcements individually. In the case of the mechanical property prediction model of the polymer composites developed in the previous researches, The first study predicted the mechanical properties of the polymer composites according to the microstructure, and the mechanical properties of the composites according to the specific bonding structure can be predicted, but the mechanical properties of the composites according to the type of the polymer and the reinforcement cannot be predicted. The second study is the most closely related to this study as it developed AI for predicting mechanical properties of polymer composites according to the individual properties and synthesis ratio of polymers and reinforcements. The two-dimensional molecular structure information and individual mechanical properties of polymer and reinforcement, as well as process variables of polymer composites, are required for the model. Therefore, when designing a polymer composite using this prediction model, there is a disadvantage in that it is not possible to make predictions about polymers and reinforcements without data. Like most material design studies using machine learning, the amount of trainable data for polymer composites design is absolutely insufficient, so the coverage of material design has been very limited. In addition, all of the developed predictive models are black-box models and do not provide any judgment basis to users regardless of their accuracy.

In this study, we proposed an AI model for predicting mechanical properties of a polymer composites using only the three-dimensional molecular structure, synthesis ratio, and process conditions to cover a wide range of chemical space without requiring experimental values of the constituent materials. The problem of data shortage for most machine learning studies was overcome through transfer learning. In addition, LIME and SHAP, the most popular methods of the Explainable AI (XAI), were used to present a persuasive judgment basis to experts and non-experts in the field of polymer composites design.

First, A total of 265 types of polymer composites were produced to generate train data, which were prepared by adding fiber-based and inorganic fillers in 5 different compositions to 14 types of polymers most used in vehicles. Next, mechanical properties such as tensile strength, tensile modulus, elongation, Poisson's ratio, flexural strength, and flexural modulus were obtained through tensile and flexural tests. The first part of the model is a parallel connected CNNs that learn the representation of the three-dimensional molecular structure of polymers and reinforcements. But training CNNs requires large amounts of data, so in order to efficiently extract features such as structural information and functional group from molecular structure, pretrained CNNs were trained by mechanical property values such as tensile modulus, Poisson’s ratio, flexural strength, and flexural modulus of polymer composites. After that, an FC layer was stacked to the pretrained CNNs and was trained to predict the tensile strength and ductility of the polymer composites through transfer learning with 265 types of polymer composites data. The performance of this predictive model based on 3D structures increased by approximately 7%, from 84% to 91%. Finally, we select 30 polymers and reinforcements suitable for vehicle lightweight materials and polymer composites composition was set between 90% and 50% as the search space. After that, we use the developed mechanical property prediction model and Bayesian optimization, high-strength and high-ductility polymer composites candidates were searched within the search space. The high-strength, high-ductility polymer composite resin design system was demonstrated five times, and the accuracy was about 87%, confirming the high accuracy. We also used an explainable A.I. (XAI) algorithm to reveal the relationship between three-dimensional molecular structure and mechanical property. Three-dimensional molecular structures that have a great influence on the model, such as CH3 and benzene ring, have been identified, and through this model analysis, a more reliable model can be provided to design experts and non-experts.