Abstract
Accurately predicting the materials’ responses, such as strain and energy, under certain loading conditions is crucial for developing fundamental structure-property relationships and facilitating material design. However, this process can be computationally expensive and challenging, especially for heterogeneous material systems with a large design space, where physics-based repetitive numerical simulations may be required. Furthermore, conducting physical experiments over such a large design space can be both time-consuming and costly. To address these challenges, convolutional neural networks (CNNs) have become increasingly popular as a computationally feasible way to make high-fidelity predictions for various materials, based on simulation results or experimental data. CNNs are particularly useful for materials with complex microstructures that are difficult to characterize or quantify, especially when suitable descriptors are not available. However, these models often suffer from poor transferability and reduced robustness due to limited training data. To overcome this challenge, we propose using multi-task learning (MTL) to provide deep learning models with more knowledge of material behaviors. MTL can improve the performance of multiple related learning tasks by leveraging useful information among them. In this paper, we explore a novel MTL model to improve the fidelity of predictions for both displacement and strain energy. We verify the ability of our proposed model by using the Material MNIST dataset as a case study, where one task is to predict strain energy based on different material microstructures. We demonstrate that, through MTL, the CNN model can achieve a better understanding of materials’ responses and achieve more robust performance in predicting responses for previously unseen material structures.
