| With the rapid development of computer in the 21st century,the application of machine learning in geotechnical engineering,especially in the determination of geotechnical parameters,is more and more extensive.At present,the determination methods of rock and soil mechanics parameters mainly include experimental method,reverse analysis method and empirical method,but all of them have certain limitations.Based on this,a machine learning method was proposed in this paper to establish the relationship between the basic physical parameters of soil and the constitutive model parameters of soil,and then the constitutive model parameters of soil were obtained by forward analysis combined with the experimental method.Using the trained model,the Duncan-Chang model parameters and modified Cambridge model parameters of soil can be obtained only through simple laboratory tests.The main contents of this paper are summarized as follows:（1）The triaxial experiments of sand consolidation and drainage were collected,and a sand model with initial water content,asymmetrical coefficient,coefficient of curvature,the average particle size,dry density and other basic parameters as input parameters and Duncan-Chang model parameters as output parameters was established,and the correlation between input parameters and output parameters was analyzed.The dry density has the greatest influence on the parameters of Duncan-Chang model.Artificial neural network and support vector machine（SVM）are used to predict the parameters of Duncan-Chang model of sand,and the prediction error of neural network is about 20%,and the prediction error of support vector machine is less than 10%.At the same time,the influence of kernel function on the prediction model and the influence of average particle size d50on the parameters of Duncan-Chang model are discussed.（2）the consolidated drained triaxial experiments of fine-grained soils on the collection,the correlation of the input parameters and output parameters are analyzed,and established the fine-grained soils in the soil type,liquid limit,plastic limit,moisture content,density of input parameters,with Duncan model parameters for the output parameters of support vector machine（SVM）model.The prediction accuracy of support vector machine models with different kernel functions is compared,and the RBF kernel function is the best,and the relative error of the final model is less than 10%.（3）The data of modified Cambridge model for clay were collected,and a support vector machine model with liquid limit,plastic limit,water content and density as input parameters and modified Cambridge model parameters as output parameters was established.The importance of input parameters and output parameters was correlated.The results showed that:Liquid limit,plastic limit,water content and density all have great influence on the modified Cambridge model of clay,and the final prediction error of the model is about10%.Finally,the modified Cambridge model parameters of MIT embankment are predicted by using the trained modified Cambridge model,and the settlement and pore pressure of embankment are calculated by using the predicted parameters,and the results are good.（4）The design of graphical user interface（GUI）of SVM is implemented,so that researchers who do not know machine learning can quickly use SVM to train and predict data and apply it to engineering design. |
