Research On Performance Prediction And Topology Optimization Of Carbon Fiber Nonwoven Composites Based On Artificial Intelligence

With the rapid development of science and technology and advanced manufacturing technology in modern industry,modern industry has higher and higher requirements for lightweight products.The ways to reduce weight mainly include structural optimization design,application of advanced technology and application of lightweight materials.Carbon fiber nonwoven composites(CFNWC)are considered to be a promising lightweight material due to their low cost,short manufacturing cycle and good lightweight effect.At the same time,using the parallel topology optimization method to perform macro and micro topology optimization design can further improve the lightweight effect of the lightweight material.In addition,considering the complex working conditions in engineering practice,topology optimization design with stress constraints is also a lightweight method that fits the actual working conditions.However,due to the unique microstructure of CFNWC,it is difficult to accurately predict their properties using homogenization methods,and the computational cost is very high due to the large number of design parameters and the need for multiple iterations to obtain the final topology in parallel topology optimization design and stress constrained topology optimization.Therefore,a data-driven method is proposed to predict the mechanical behavior of carbon fiber nonwoven composites,and at the same time,a deep learning-based topology optimization design model is established to accelerate the generation of topology-optimized structures.The research contents of this paper are as follows.(1)A data-driven model combining principal component analysis and Bayesian neural network is developed to predict the mechanical behavior of carbon fiber nonwoven composites.First,PCA is used to reduce the dimension of the stress-strain curve.Secondly,Bayesian neural networks are established to predict the stress-strain curve of the carbon fiber non-woven composite material,and artificial neural networks are established to predict the stress-strain curve of the carbon fiber non-woven composite material and compare with the results of the Bayesian neural network.Finally,the accuracy of the proposed method is verified by a new testing dataset,and the prediction model is used to select the microscopic parameters of carbon fiber nonwoven composites.(2)A parallel topology optimization model based on deep learning is developed to achieve fast generation of cross-scale topology optimization structures.First,a crossscale topology-optimized structure for neural network training is generated by means of a bi-directional evolutionary structural optimization method based on carbon fiber nonwoven composites.Then,a coupled neural network model combining residual neural network,U-net architecture and SE neural network is established.Finally,the accuracy and rapidity of the coupled neural network in generating cross-scale topologies is verified by testing the neural network with a new set of data.(3)A deep learning-based stress-constrained topology optimization model is developed to achieve fast generation of stress-constrained topologies.First,the dataset for neural network training,validation and testing are generated by a bi-directional evolutionary structural optimization method with stress constraints based on carbon fiber nonwoven composites at different volume fractions of carbon fibers.Then,based on transfer learning,the coupled neural network model used to generate the cross-scale structure is transferred to this training task to realize the rapid generation of the topology optimization structure with stress constraints.Finally,a new set of data is used to illustrate the effectiveness and accuracy of the coupled neural network model to rapidly generate topologically optimized structures with stress constraints.