인공 신경망을 이용한 구오스테나이트 결정립계의 재구성 및 크기 예측 (Reconstruction and Size Prediction of Prior Austenite Grain Boundary (PAGB) using Artificial Neural Networks)

서지정보

· 발행기관 : 대한금속·재료학회
· 수록지 정보 : 대한금속·재료학회지 / 58권 / 12호 / 822 ~ 829페이지
· 저자명 : 김봉규, 구남훈, 이종혁, 한준현

초록

To automatically reconstruct the prior austenite grains from as-quenched martensitic structure, we applied a deep learning algorithm to recognize the prior austenite grains boundaries hidden in the martensitic matrix. The FC-DenseNet architecture based on FCN (fully convolutional networks) was used to train the martensite and ground truth label of the prior austenite grain boundaries. The original martensite structures and prior austenite grain boundaries were prepared using different chemical etching solutions. The initial PAGS detection rate was as low as 37.1%, which is not suitable for quantifying the basic properties of the microstructure such as grain size or grain boundary area. By changing the weight factor of the neural net loss function and increasing the size of the data set, the detection rate was improved up to 56.1%. However, even when the detection rate reached 50% or more, the quality of the reconstructed PAGS was not comparable to the analytically calculated results based on EBSD measurements and crystallographic orientation relationships. The prior austenite grain size data sets were obtained from martensite samples via the FCDenseNet method, and had a linear correlation with the mechanical properties measured in the same samples. In order to improve the accuracy of the detection rate using neural networks, it is necessary to increase the number of neural networks and data sets.

영어초록

To automatically reconstruct the prior austenite grains from as-quenched martensitic structure, we applied a deep learning algorithm to recognize the prior austenite grains boundaries hidden in the martensitic matrix. The FC-DenseNet architecture based on FCN (fully convolutional networks) was used to train the martensite and ground truth label of the prior austenite grain boundaries. The original martensite structures and prior austenite grain boundaries were prepared using different chemical etching solutions. The initial PAGS detection rate was as low as 37.1%, which is not suitable for quantifying the basic properties of the microstructure such as grain size or grain boundary area. By changing the weight factor of the neural net loss function and increasing the size of the data set, the detection rate was improved up to 56.1%. However, even when the detection rate reached 50% or more, the quality of the reconstructed PAGS was not comparable to the analytically calculated results based on EBSD measurements and crystallographic orientation relationships. The prior austenite grain size data sets were obtained from martensite samples via the FCDenseNet method, and had a linear correlation with the mechanical properties measured in the same samples. In order to improve the accuracy of the detection rate using neural networks, it is necessary to increase the number of neural networks and data sets.

참고자료

· 없음