Machine Learning-Based Design Methods Considering Data Characteristics and Design Space Size

Date: 2023/11/24 - 2023/11/24

Academic Seminar: Machine Learning-Based Design Methods Considering Data Characteristics and Design Space Size

Speaker: Seunghwa Ryu, Full Professor, Department of Mechanical Engineering, Korea Advanced Institute of Science and Technology

Time: 10:00 - 11:30, November 24, 2023 (Beijing Time)

Location: GIFT (East Wing of Bao Yugang Library), Room 200

Abstract

The landscape of material science and manufacturing has transformed significantly with the advent of machine learning (ML). This toolkit of data driven methods accelerated the discovery and production of new materials by accurately predicting the complicated physical processes and mechanisms that are not fully described by existing material theories. Yet, with an array of intricate ML models at our disposal, the pressing question remains: Which ML algorithm is best suited for our needs? In this presentation, we aim to provide insights to strategically select appropriate models aligned with specific design challenges. We further segment material design challenges into: 1) deep learning based interpolation problem: ample training data capturing design space trends. 2) deep learning based extrapolation problem: immense design space demanding more than just the initial training dataset. 3) limited data scenario: instances where only handful of dataset is available. 4) multi-fidelity datasets: a combination of concise, precise datasets and expansive, approximate ones. The most successful machine learning-based surrogate models and design approaches will be discussed for each case along with pertinent literature.

Biography

Seunghwa Ryu, a full professor of Mechanical Engineering at KAIST, embarked on his academic career with a BS degree from KAIST in 2004, followed by a PhD from Stanford in 2011. After conducting postdoctoral research at MIT in 2012, he returned to KAIST in 2013 to start his professional career. His research interests lie in predicting material properties through theory and computer simulations across multiple scales and using artificial intelligence (AI) algorithms to efficiently design next-generation materials and products. He has published over 130 papers in international journals and holds positions on the editorial boards of Scientific Reports and Frontiers in Materials.