Study On Macro-meso Mechanical Behaviors Of Concrete Based On Multiple Deep Learning Models

Concrete material,that has the advantage of high compressive strength,is the most widely used building material at present.The research of its mechanical behavior is one of the most challenging tasks in the field of solid mechanics,which has aroused extensive attention and research of a large number of scholars.At the meso-scale,concrete is considered as a kind of composite material composed of coarse aggregate,mortar matrix and pores.It is widely accepted that the nonlinear mechanical properties of concrete at macro-scale are closely related to the morphology and distribution features of each meso-scopic component.But on the one hand,it is extremely challenging to establish a clear expression of the relationship between meso-scopic characteristics and macro-scopic mechanical properties.On the other hand,meso-scale mechanics studies usually require a large number of real meso-structure samples,although CT technology can extract the meso-structural features of concrete,repeated high-cost CT scanning is time-consuming and labor-intensive,which cannot provide large-scale samples for the study of meso-numerical simulation.Deep learning is expert in excavate the internal characteristic information of data and has been widely applied in many fields.In recent years,deep learning has achieved pioneering research results in the fields of macro-scopic mechanical properties prediction of heterogeneous materials and reconstruction design of heterogeneous materials.However,the related theory,application scenarios,and specific research methods are still in the preliminary exploration stage,especially the research results on concrete,a typically heterogeneous composite material,are very limited.Based on this,according to the advantages and characteristics of different algorithms in deep learning,this paper takes concrete materials as the research object,and study the macro-meso mechanical behavior of concrete.On the one hand,Goog Le Net is utilized to predict the macro-scopic stress-strain curve of concrete meso-model,on the other hand,with the help of generated adversarial network,exploratory research work has been carried out on the reconstruction of concrete meso-model.The main research contents are as follows:(1)Prediction of stress-strain curve of concrete meso-model.Firstly,a two-dimensional random aggregate model with different aggregate content and porosity was established.The stress-strain curves under static uniaxial compression were obtained by using ABAQUS.Data preprocessing was performed on the finite element model and the stress-strain curves to establish a data set for deep learning.Secondly,the data-set is trained by the Keras-based Goog Le Net model and the improved loss function.Finally,the prediction results of the stress-strain curve,the macro-scopic mechanical properties extracted from the stress-strain curve are compared and analyzed with the numerical results of the finite element.The test results of Goog Le Net model and other classical CNN models are also compared and analyzed.The results show that Goog Le Net model could excavate the complex mapping between concrete meso-model and stress-strain curve,and achieve rapid and accurate end-to-end prediction.(2)Research on reconstruction and mechanical behavior of concrete meso-model.Firstly,the CT cross-sectional images of the concrete specimens are collected,and image data-set of the 2D concrete meso-scopic model including aggregate and mortar matrix is obtained by means of a series of image preprocessing methods.Secondly,the concrete meso-model is reconstructed by loading well-trained model parameters based on DCGAN training.Thirdly,DCGAN is trained based on data-set,and concrete meso-scopic model is reconstructed by loading model parameters.Finally,the consistency and differences in mechanical responses under uniaxial compressive loading between generative and original samples at the same aggregate content are analyzed.The results show that the DCGAN model could achieve rapid and efficient reconstruction of concrete meso-model,and could provide reliable reconstruction models for meso-scopic numerical simulation experiments.