Research On Prediction Method Of Composite Material Constitutive Model Based On Deep Material Network

The automotive industry has proposed higher requests for lightweight technologies for automobiles against the background of excessive energy consumption and continuous environmental destruction.Due to the excellent material properties of composite materials,using composite materials to replace the metal materials of the car can effectively reduce the weight of the car,increase the cruising range,reduce exhaust emissions,and effectively achieve energy saving.The performance of the composite material is closely related to the content,performance,distribution and interface characteristics of each component.The experimental method is difficult to carry out multi-variable control.The calculation method of finite element simulation is limited by the performance of the computer.The analytical method based on the multi-scale theoretical analysis is poor in versatility.All these reasons cause the high cost and long period of time to obtain the mechanical properties of composite materials.In this paper,based on the tool of neural network,a multi-scale constitutive model was proposed that can accurately predict the macroscopic performance of composite materials according to the micro information.The mechanical performance parameters of composite material can be accurately inferred,and the derivation time can be greatly reduced.The model proposed in this paper is a supervised learning process based on big data.First,the sampling experiment design of material properties based on the COSMOS material library was completed.According to the sampling process of material properties through the acceptance-rejection sampling method based on the Monte Carlo sampling method,we get the data sample collection of 500 sets of material parameters.Based on the secondary development of ABAQUS and Python,the composite material RVE models was established,and the high-fidelity composite material mechanical property data samples were obtained through finite element simulation calculation.A total of 500 groups of long fiber composite material samples,500 groups of reinforced particle composite material samples and 100 samples groups of random short fiber composite material were obtained.On the basis of the above work,this paper proposes an improved binary-tree deep material network structure,which combines the physical calculation layer and the weight pruning algorithm to obtain a deep learning-based composite material mechanical performance prediction model with physical significance.Our dataset is divided by the training set-test set division method in scikit-learn.The accuracy of the network structure increases as the height of the binary tree increased,and the training time also increased with the training process based on the dataset.When the height of the binary tree reached 7,the error rate of inference can be reduced to less than 5%.The training time is about 90 minutes(The calculation time for the predicted modulus of the deep material network is about 2.5seconds).For the random short fiber model,we trained the small batch of high-latitude data through migration learning pre-training-fine-tuning method.Compared the results with the range of analytical solution by Voigt and Reuss methods and finite element simulation,this paper analyzed and improved the models with low accuracy,between the calculated value of the improved model and the upper limit of the analytical solution range is-4.16%,the relative deviation from the lower limit of the analytical solution range is 3.72%,and the relative error with the finite element simulation is 2.72%,of which the finite element simulation calculation time is 623 s,the calculation time of deep material network is 2.516 s.It is concluded that the calculation results of this article are within the expected range.The calculation capacity of the model reaches the expected.And the calculation time can be effectively reduced.