TY - GEN
T1 - HOW TO ENCODE MICROSTRUCTURE IN MACHINE LEARNING
T2 - ASME 2023 International Design Engineering Technical Conferences and Computers and Information in Engineering Conference, IDETC-CIE 2023
AU - Wu, Yulun
AU - Li, Yumeng
N1 - Publisher Copyright:
© 2023 American Society of Mechanical Engineers (ASME). All rights reserved.
PY - 2023
Y1 - 2023
N2 - Accurately predicting the response of materials under different loading conditions is crucial for designing and developing new materials with desired properties. However, this process can be computationally expensive and challenging, especially for heterogeneous materials with complex microstructures. Recently, machine learning has been widely used to address the challenge for developing predictive models for various material systems with reduced reliance on extensive experimental testings and repetitive expensive physics simulations. The microstructure of a material plays a critical role in determining its properties, making it a key factor that needs to be accounted for in predictive modeling. Heterogeneous materials, specifically, often have complex microstructures with numerous features like pores, inclusions, and grain boundaries, which need to be accurately captured but is hard to be quantified for developing machine learning based predictive models. Therefore, accurate encoding of microstructural features is essential for making reliable predictions. Nevertheless, how to effectively and efficiently capture the complex microstructural features in developing machine learning based predictive models largely remains an open question for researchers in the field of materials science. In this paper, we present a comparison study of different encoding methods for microstructures in machine learning models. Specifically, we investigate pre-defined encoding methods and automatic encoding methods for a synthetic heterogeneous material system. the performance of each machine learning model is evaluated by predicting material responses such as strain energy. Our results show that convolutional neural networks (CNNs) have the ability to auto-encode the microstructure information of material and make promising prediction, especially when good pre-defined descriptors are not available. Overall, this study provides valuable insights into the performance of different encoding methods for microstructures in machine learning models, and can inform the development of more accurate and efficient models for materials science applications.
AB - Accurately predicting the response of materials under different loading conditions is crucial for designing and developing new materials with desired properties. However, this process can be computationally expensive and challenging, especially for heterogeneous materials with complex microstructures. Recently, machine learning has been widely used to address the challenge for developing predictive models for various material systems with reduced reliance on extensive experimental testings and repetitive expensive physics simulations. The microstructure of a material plays a critical role in determining its properties, making it a key factor that needs to be accounted for in predictive modeling. Heterogeneous materials, specifically, often have complex microstructures with numerous features like pores, inclusions, and grain boundaries, which need to be accurately captured but is hard to be quantified for developing machine learning based predictive models. Therefore, accurate encoding of microstructural features is essential for making reliable predictions. Nevertheless, how to effectively and efficiently capture the complex microstructural features in developing machine learning based predictive models largely remains an open question for researchers in the field of materials science. In this paper, we present a comparison study of different encoding methods for microstructures in machine learning models. Specifically, we investigate pre-defined encoding methods and automatic encoding methods for a synthetic heterogeneous material system. the performance of each machine learning model is evaluated by predicting material responses such as strain energy. Our results show that convolutional neural networks (CNNs) have the ability to auto-encode the microstructure information of material and make promising prediction, especially when good pre-defined descriptors are not available. Overall, this study provides valuable insights into the performance of different encoding methods for microstructures in machine learning models, and can inform the development of more accurate and efficient models for materials science applications.
KW - convolution neural networks
KW - machine learning
KW - material response
KW - microstructure encoding
UR - http://www.scopus.com/inward/record.url?scp=85179131407&partnerID=8YFLogxK
UR - http://www.scopus.com/inward/citedby.url?scp=85179131407&partnerID=8YFLogxK
U2 - 10.1115/DETC2023-116704
DO - 10.1115/DETC2023-116704
M3 - Conference contribution
AN - SCOPUS:85179131407
T3 - Proceedings of the ASME Design Engineering Technical Conference
BT - 49th Design Automation Conference (DAC)
PB - American Society of Mechanical Engineers (ASME)
Y2 - 20 August 2023 through 23 August 2023
ER -