본문 바로가기 대메뉴 바로가기
닫기

research

Deep Learning Framework to Enable Material Design in Unseen Domain​
View : 2602 Date : 2021-09-29 Writer : PR Office

Researchers propose a deep neural network-based forward design space exploration using active transfer learning and data augmentation

Figure 1: Schematic of deep learning framework for material design space exploration. Schematic of gradual expansion of reliable prediction domain of DNN based on the addition of data generated from the hyper-heuristic genetic algorithm and active transfer learning.

< Figure 1: Schematic of deep learning framework for material design space exploration. Schematic of gradual expansion of reliable prediction domain of DNN based on the addition of data generated from the hyper-heuristic genetic algorithm and active transfer learning. >

A new study proposed a deep neural network-based forward design approach that enables an efficient search for superior materials far beyond the domain of the initial training set. This approach compensates for the weak predictive power of neural networks on an unseen domain through gradual updates of the neural network with active transfer learning and data augmentation methods.

Professor Seungwha Ryu believes that this study will help address a variety of optimization problems that have an astronomical number of possible design configurations. For the grid composite optimization problem, the proposed framework was able to provide excellent designs close to the global optima, even with the addition of a very small dataset corresponding to less than 0.5% of the initial training data-set size. This study was reported in npj Computational Materials last month.

“We wanted to mitigate the limitation of the neural network, weak predictive power beyond the training set domain for the material or structure design,” said Professor Ryu from the Department of Mechanical Engineering. 

Neural network-based generative models have been actively investigated as an inverse design method for finding novel materials in a vast design space. However, the applicability of conventional generative models is limited because they cannot access data outside the range of training sets. Advanced generative models that were devised to overcome this limitation also suffer from weak predictive power for the unseen domain.

Professor Ryu’s team, in collaboration with researchers from Professor Grace Gu’s group at UC Berkeley, devised a design method that simultaneously expands the domain using the strong predictive power of a deep neural network and searches for the optimal design by repetitively performing three key steps. 

First, it searches for few candidates with improved properties located close to the training set via genetic algorithms, by mixing superior designs within the training set. Then, it checks to see if the candidates really have improved properties, and expands the training set by duplicating the validated designs via a data augmentation method. Finally, they can expand the reliable prediction domain by updating the neural network with the new superior designs via transfer learning. Because the expansion proceeds along relatively narrow but correct routes toward the optimal design (depicted in the schematic of Fig. 1), the framework enables an efficient search.

As a data-hungry method, a deep neural network model tends to have reliable predictive power only within and near the domain of the training set. When the optimal configuration of materials and structures lies far beyond the initial training set, which frequently is the case, neural network-based design methods suffer from weak predictive power and become inefficient. 

Researchers expect that the framework will be applicable for a wide range of optimization problems in other science and engineering disciplines with astronomically large design space, because it provides an efficient way of gradually expanding the reliable prediction domain toward the target design while avoiding the risk of being stuck in local minima. Especially, being a less-data-hungry method, design problems in which data generation is time-consuming and expensive will benefit most from this new framework.
The research team is currently applying the optimization framework for the design task of metamaterial structures, segmented thermoelectric generators, and optimal sensor distributions. “From these sets of on-going studies, we expect to better recognize the pros and cons, and the potential of the suggested algorithm. Ultimately, we want to devise more efficient machine learning-based design approaches,” explained Professor Ryu.This study was funded by the National Research Foundation of Korea and the KAIST Global Singularity Research Project.

-Publication
Yongtae Kim, Youngsoo, Charles Yang, Kundo Park, Grace X. Gu, and Seunghwa Ryu, “Deep learning framework for material design space exploration using active transfer learning and data augmentation,” npj Computational Materials (https://doi.org/10.1038/s41524-021-00609-2)

-Profile
Professor Seunghwa Ryu
Mechanics & Materials Modeling Lab
Department of Mechanical Engineering
KAIST

multi-scale mechanicsmaterial simulationsmachine learningmaterials designingcomputational materialsmetamaterial structuredeep learningGlobal Singularity Research ProjectGlobal Singularity Research Professor Seunghwa Ryu College of Engineering Department of Mechanical Engineering Mechanics & Materials Modeling Lab
List
Releated news