Symposium SB10: New E-Textile Materials and Devices for Wearable Electronics
Symposium EQ01: Progress in Thermoelectrics—From Traditional to Novel Materials

Symposium DS02: Integrating Machine Learning with Simulations for Accelerated Materials Modeling

Seunghwa Ryu, Korea Advanced Institute of Science & Technology

Machine Learning Based Design of Composite Structures

Written by Henry Quansah Afful

Composite materials have been explored for a myriad of applications owing to their superior properties over their individual constituents. These materials can be grouped, based on the available models, as particulate-reinforced, random microstructure, and periodic structures. There is a limitation in the frameworks available for designing materials systems with more predictable and superior properties. Machine learning (ML) is being employed to help predict and design much better composite materials based on available experimental and computational data. However, ML-based design faces challenges in extrapolation into unknown design spaces and limited datasets. To solve the extrapolation problem, Seunghwa Ryu proposed an active learning approach where the ML models were gradually updated with more data until the global optimum mechanical property was found. For this to be effective, results from the ML model need to be validated with experiments and simulations and new data fed into the neural network. Ryu demonstrated how the Bayesian optimization can be used to tackle the problem of limited datasets and low quality data by identifying regions with largest uncertainty and working to decrease the uncertainty. Ryu employed this approach to improve the toughness of a nacre-inspired composite.


The comments to this entry are closed.