In this paper,a novel mixed wavelet-learning method is developed for predicting macroscopic effective heat transfer conductivities of braided composite materials with heterogeneous thermal conductivity.This innovative...In this paper,a novel mixed wavelet-learning method is developed for predicting macroscopic effective heat transfer conductivities of braided composite materials with heterogeneous thermal conductivity.This innovative methodology integrates respective superiorities of multi-scale modeling,wavelet transform and neural networks together.By the aid of asymptotic homogenization method(AHM),off-line multi-scalemodeling is accomplished for establishing thematerial databasewith highdimensional and highly-complexmappings.Themulti-scalematerial database and the wavelet-learning strategy ease the on-line training of neural networks,and enable us to efficiently build relatively simple networks that have an essentially increasing capacity and resisting noise for approximating the high-complexity mappings.Moreover,it should be emphasized that the wavelet-learning strategy can not only extract essential data characteristics from the material database,but also achieve a tremendous reduction in input data of neural networks.The numerical experiments performed using multiple 3D braided composite models verify the excellent performance of the presentedmixed approach.The numerical results demonstrate that themixedwaveletlearningmethodology is a robustmethod for computing themacroscopic effective heat transfer conductivities with distinct heterogeneity patterns.The presentedmethod can enormously decrease the computational time,and can be further expanded into estimating macroscopic effective mechanical properties of braided composites.展开更多
A deep learning based homogenization framework is proposed to link the microstructures of porous nickel/yttriastabilized zirconia anodes in solid oxide fuel cells(SOFCs)to their effective macroscopic properties.A vari...A deep learning based homogenization framework is proposed to link the microstructures of porous nickel/yttriastabilized zirconia anodes in solid oxide fuel cells(SOFCs)to their effective macroscopic properties.A variety of microstructures are generated by the discrete element method and the meso‑scale kinetic Monte Carlo method.Then,the finite element method and the homogenization theory are used to calculate the effective elastic modulus(E),Poisson’s ratio(υ),shear modulus(G)and coefficient of thermal expansion(CTE)of representative volume elements.In addition,the triple-phase boundary length density(LTPB)is also calculated.The convolutional neural network(CNN)based deep learning model is trained to find the potential relationship between the microstructures and the five effective macroscopic properties.The comparison between the ground truth and the predicted values of the new samples proves that the CNN model has an excellent predictive performance.This indicates that the CNN model could be used as an effective alternative to numerical simulations and homogenization because of its accurate and rapid prediction performance.Hence the deep learning-based homogenization framework could potentially accelerate the continuum modeling of SOFCs for microstructure optimization.展开更多
基金supported by the National Natural Science Foundation of China(No.12001414)the Fundamental Research Funds for the Central Universities(No.JB210702)+4 种基金the open foundation of Hubei Key Laboratory of Theory and Application of Advanced Materials Mechanics(Wuhan University of Technology)(No.WUTTAM202104)the China Postdoctoral Science Foundation(No.2018M643573)the Natural Science Basic Research Program of Shaanxi Province(No.2019JQ-048)the National Natural Science Foundation of China(Nos.51739007 and 61971328)supported by the Center for high performance computing of Xidian University.
文摘In this paper,a novel mixed wavelet-learning method is developed for predicting macroscopic effective heat transfer conductivities of braided composite materials with heterogeneous thermal conductivity.This innovative methodology integrates respective superiorities of multi-scale modeling,wavelet transform and neural networks together.By the aid of asymptotic homogenization method(AHM),off-line multi-scalemodeling is accomplished for establishing thematerial databasewith highdimensional and highly-complexmappings.Themulti-scalematerial database and the wavelet-learning strategy ease the on-line training of neural networks,and enable us to efficiently build relatively simple networks that have an essentially increasing capacity and resisting noise for approximating the high-complexity mappings.Moreover,it should be emphasized that the wavelet-learning strategy can not only extract essential data characteristics from the material database,but also achieve a tremendous reduction in input data of neural networks.The numerical experiments performed using multiple 3D braided composite models verify the excellent performance of the presentedmixed approach.The numerical results demonstrate that themixedwaveletlearningmethodology is a robustmethod for computing themacroscopic effective heat transfer conductivities with distinct heterogeneity patterns.The presentedmethod can enormously decrease the computational time,and can be further expanded into estimating macroscopic effective mechanical properties of braided composites.
基金This work was supported by the National Natural Science Foundation of China(Nos.11932005,12172104)the National Key R&D Program of China(No.2018YFB1502602)Shenzhen Science and Technology Innovation Commission(JCYJ20200109113439837).
文摘A deep learning based homogenization framework is proposed to link the microstructures of porous nickel/yttriastabilized zirconia anodes in solid oxide fuel cells(SOFCs)to their effective macroscopic properties.A variety of microstructures are generated by the discrete element method and the meso‑scale kinetic Monte Carlo method.Then,the finite element method and the homogenization theory are used to calculate the effective elastic modulus(E),Poisson’s ratio(υ),shear modulus(G)and coefficient of thermal expansion(CTE)of representative volume elements.In addition,the triple-phase boundary length density(LTPB)is also calculated.The convolutional neural network(CNN)based deep learning model is trained to find the potential relationship between the microstructures and the five effective macroscopic properties.The comparison between the ground truth and the predicted values of the new samples proves that the CNN model has an excellent predictive performance.This indicates that the CNN model could be used as an effective alternative to numerical simulations and homogenization because of its accurate and rapid prediction performance.Hence the deep learning-based homogenization framework could potentially accelerate the continuum modeling of SOFCs for microstructure optimization.
