Study On The Microscopic Phase-field Of Ni-Al-V Alloy Precipitation Based On Artificial Neural Network

Nickel based alloys have excellent strength,corrosion resistance,and oxidation resistance.Therefore,Ni alloys are widely used in marine,environmental protection,energy,petrochemical,and food fields.The microstructure of a material determines its macroscopic properties.Studying the precipitation mechanism of Ni alloys helps to accurately analyze the microstructure of the alloy,thereby improving its properties.The micro diffusion phase field focuses on the changes in atomic scale information,so it can describe the microstructure evolution of atoms during alloy precipitation.In this paper,a micro diffusion phase field model is used to simulate the precipitation process of ternary Ni₇₅Al_2.5V_22.5 alloy at temperatures of 1050K and 1350K,respectively.The effect of temperature on the precipitation mechanism and atomic occupancy probability of Ni₇₅Al_2.5V_22.5alloy is analyzed,as well as the change trend of corresponding order parameters and ordered precipitation phase volume fraction;Then,at a temperature of 1050K,and at Al concentrations of 0.025,0.05,and 0.1,respectively,the precipitation process of the simulated alloy was calculated,and the effect of Al concentration on the precipitation mechanism and atomic occupancy probability of the Ni-Al-V alloy was analyzed,as well as the change trend of corresponding order parameters and ordered precipitation phase volume fraction.When the temperature of the alloy increases,the undercooling of the system decreases,resulting in a decrease in the driving force of alloy precipitation.As a result,the precipitation speed slows down,extending the incubation period of the two ordered phases,DO₂₂ and L1₂.The precipitation mechanism of DO₂₂ phase changes from ordering and unstable decomposition to unstable decomposition.With the increase of Al concentration,the nucleation incubation period of L1₂ phase gradually decreases,and the incubation period of DO₂₂ phase gradually extends.Finally,L1₂ phase will precipitate before DO₂₂ phase,and the number of L1₂ phase gradually increases.However,once the initial conditions of alloy precipitation have changed,it is necessary to re execute the computational simulation program to obtain the corresponding precipitation results,which takes a long time.In order to shorten the simulation calculation time and improve the calculation efficiency,this paper proposes to establish an artificial neural network model instead of a micro diffusion phase field model to directly predict the impact of different alloy temperatures or Al concentrations on alloy precipitation results.The main research results include:（1）A micro diffusion phase field model based on artificial networks is established.BP neural network is a feedforward supervised network model with strong generalization ability and nonlinear mapping ability.Its input includes alloy temperature and Al concentration in the alloy,and its output includes atomic occupancy probability at a certain lattice point,average long program parameters,and volume fraction of ordered phases.Generate a large amount of training data through a micro diffusion phase field model.The single simulation time of the BP neural network is only 0.6 seconds,and the mean square error is less than 5%.Compared with the calculation time of the micro diffusion phase field,the efficiency is improved by about 970times.（2）Improve the prediction accuracy of the artificial neural network model.The fundamental method to improve the accuracy of model prediction is the selection of algorithms and the adjustment of network structure.There are three typical algorithms in BP neural network systems,namely,Levenberg-Marquardt algorithm,Bayesian Regulation algorithm,and Scaled Converge Gradient algorithm.When building models for different data,this article selects the most suitable algorithm from the BP neural network system as the training method for the model based on the type and trend of the data,constantly adjusting the number of hidden layers and the number of neurons therein,to avoid problems such as low learning efficiency,slow convergence speed,and susceptibility to minima in the model.It is found that the BP network model with a single hidden layer and about 30 neurons is the network structure with the highest prediction efficiency and accuracy.（3）Analyze the prediction error of the artificial neural network model.The mean square error of all BP neural network models established in this paper is less than 5%,and the average goodness of fit index R value of each model is above 0.8,meeting the design requirements.After each model is established,10 sets of new data are used to test the model,and the correlation coefficient R² is found to be above 0.86,indicating that all models have strong applicability.