Prediction Of Mechanical Properties Of Mg-RE Alloy Based On Convolutional Neural Network

Magnesium alloys,currently the most promising lightweight alloy materials,have broad application potential in fields such as transportation,aerospace,and telecommunications.However,they face numerous challenges in research and production processes.For instance,the massive and heterogeneous data of magnesium alloys impede the process of standardization and normalization;the majority of magnesium alloys,due to their highly dense hexagonal structure and limited slip systems at room temperature,have poor plastic formability;and the research on high-performance magnesium alloys is still in an exploratory stage.Thus,enhancing the mechanical properties of magnesium alloys has become the focal point of ongoing research.Yet,mechanical properties are influenced by many factors such as alloy composition,process parameters,and organizational structure,the interactions of which are complex.To date,no traditional mathematical model can accurately describe their quantitative relationships.To boost the efficiency of traditional experimental research,develop new research methods,and accelerate the research process of magnesium alloys,the adoption of interdisciplinary integration,computer-aided design,and the introduction of artificial intelligence technology have gradually become new directions in magnesium alloy research.In this paper,Mg-RE magnesium alloy is taken as the research object,and a series of tensile tests and microstructure observation are carried out,and the network database of artificial neural network is established.Convolutional neural network is built in Python language,with metallographic diagram of rare-earth magnesium alloy as input variables,ultimate tensile strength,yield strength and elongation as output variables,and 800 times of training and verification processes are carried out in cycles to generate a mechanical properties prediction model,which realizes the establishment of the mapping relationship between microand macro-properties of materials.The prediction model is actually verified,and the internal relationship between its microstructure(grain size,morphology of precipitated phase and texture analysis)and mechanical properties is analyzed,and the main contributions of grain boundary strengthening and dislocation strengthening to yield strength are mastered.The specific research contents and results are as follows:(1)The mechanical performance prediction model of rare earth magnesium alloy studied in this project was developed based on demand analysis.The overall task of the development is to connect the microstructure and mechanical properties of rare earth magnesium alloy.By introducing artificial neural network technology into the study of rare earth magnesium alloys,a relationship model between the microstructure and properties of rare earth magnesium alloys can be established.Python language was selected for developing the convolutional neural network model used for performance prediction because of its rich libraries and frameworks for machine learning and deep learning,which can simplify the development and training process of neural networks,and Py Side2 was chosen for developing the human-machine interactive interface due to its strong support for object-oriented programming and massive basic class library.(2)An artificial neural network model based on CNN was established to predict the mechanical properties of Mg-9Gd-Y-Zn-Zr alloy.The model uses a network architecture of convolutional layer plus pooling layer plus two fully connected layers.Py Torch framework was used with a convolution kernel size of 3x3,4 convolution layers,a stride of 2,and Relu function as activation function.Maximum pooling was used,and the training and validation were repeated 800 times to generate the mechanical performance prediction model.After optimizing the parameters and structure of the model,the variance of the mechanical performance parameters output by the model was significantly reduced,and the overfitting phenomenon was weakened.The mean square errors of tensile strength and yield strength were 3.25% and 4.77%,respectively,and the accuracy of the model reached 95%.(3)The mechanical performance prediction model was verified in practice,and the errors of the tensile strength of the three samples were 3.53%,0.45%,and 4.36%,respectively,while the errors of the yield strength were 4.97%,3.24%,and 4.04%,respectively,and the errors of elongation were 4.8%,5%,and 2.9% respectively,which were within the error range of the model.The sample with the best mechanical performance is mainly due to the grain refinement caused by static recrystallization,which plays a role in strengthening the fine grains.The contribution of grain refinement to its mechanical properties enhancement was 98.1 MPa,while the contribution of dislocation strengthening was 14.91 MPa.