Machine Learning Driven Material Property Prediction Framework for Intelligent Self Healing Polymer Composites Design

Abstract

Self-healing polymer composites are a revolutionizing class of materials that have the potential for autonomous mechanical repair for improved usage and sustainability. However, it is a challenge to design composites with the best possible mechanical strength, thermal stability and healing efficiency considering the complex interactions of polymer chemistry, filler content and process parameters. This study proposes a hybrid physics-informed deep learning (DL) framework with active design optimization (HPIDLO) for the prediction and design optimization of the main properties of self-healing polymer composites. The framework incorporates physics-based feature embedding, multi-task deep neural network, active learning-based data augmentation and surrogate-aided optimization. Experimental and simulated data sets of 481 polymer formulations were used to train the model. Quantitative results demonstrate that HPIDLO shows an MAE of 2.98 MPa, mean square error of 4.56 MPa, R squared of 0.92 for tensile strength with healing efficiency of 92.4% and thermal stability of 179 degC, which outperforms 5 state-of-the-art methods. The framework offers an efficient, accurate and generalizable tool for intelligent composite design, which can guide the formulation of experiments, decrease costs and increase innovation.