Deep Neural Network-Evaluated Thermal Conductivity for Two-Phase WC-M (M = Ag, Co) Cemented Carbides
Shiyi Wen,
Xiaoguang Li,
Bo Wang,
Jing Tan,
Yuling Liu,
Jian Lv,
Zhuopeng Tan,
Lei Yin,
Yong Du
Affiliations
Shiyi Wen
School of Metallurgy and Environment, Central South University, Changsha 410083, China
Xiaoguang Li
Key Laboratory of Computing and Stochastic Mathematics (Ministry of Education), School of Mathematics and Statistics, Hunan Normal University, Changsha 410081, China
Bo Wang
Key Laboratory of Computing and Stochastic Mathematics (Ministry of Education), School of Mathematics and Statistics, Hunan Normal University, Changsha 410081, China
Jing Tan
State Key Laboratory of Powder Metallurgy, Central South University, Changsha 410083, China
Yuling Liu
State Key Laboratory of Powder Metallurgy, Central South University, Changsha 410083, China
Jian Lv
Institute of Engineering Research, Jiangxi University of Science and Technology, Ganzhou 341000, China
Zhuopeng Tan
Ganzhou Achteck Tool Technology Co., Ltd., Ganzhou 341000, China
Lei Yin
Ganzhou Achteck Tool Technology Co., Ltd., Ganzhou 341000, China
Yong Du
State Key Laboratory of Powder Metallurgy, Central South University, Changsha 410083, China
DNN (Deep Neural Network) is one kind of method for artificial intelligence, which has been applied in various fields including the exploration of material properties. In the present work, DNN, in combination with the 10-fold cross-validation, is applied to evaluate and predict the thermal conductivities for two-phase WC-M (M = Ag, Co) cemented carbides. Multi-layer DNNs were established by learning the measured thermal conductivities for the WC-Ag and WC-Co systems. It is observed that there are local-minimum regions for the loss functions during training and testing the DNNs, and the presently utilized Adam optimizer is valid for breaking the local-minimum regions. The good agreements between the DNN-evaluated thermal conductivities and the measured ones manifest that the DNNs were well trained and tested. Moreover, another 1000 input data points were randomly generated for the established DNNs to predict the thermal conductivities for WC-Ag and WC-Co systems, respectively. Compared with the thermal conductivities predicted by the previously developed physical model, the presently established DNNs show similarly robust predicting ability. Concerning the efficiency, it is demonstrated in the present work that machine learning is promising to explore the material properties, especially in the high-dimensional parameter space, more efficiently than previous models, and thus can considerably contribute to the corresponding material design with less time consumption and costs.
