Study On Constitutive Model Of Frozen Silty Clay Based On Machine Learning

With the increasing demand for engineering construction in seasonally frozen regions,the study of frozen soil constitutive relationship is of great significance for the design,safety,and stability of underground projects.The development of constitutive models requires a large amount of experimental data to support,but the acquisition of experimental data often requires a lot of social resources,which limits the development and application of constitutive models.In addition,due to the highly nonlinear and strong coupling relationship between initial moisture content,freeze-thaw cycles,confining pressure and deviatoric stress,it is difficult to accurately predict the constitutive relationship.Based on this,this paper constructs a large database through experimental data to realize the theory-driven frozen silty clay constitutive model prediction.This study provides a new and effective method for constitutive prediction of frozen soil.In this paper,75 groups of triaxial compression tests of frozen silty clay at low temperature were carried out.Considering the effects of freeze-thaw cycles,initial moisture content and confining pressure on its strength,cohesion and internal friction angle,the mechanical properties of frozen silty clay were revealed.The results showed that under the same confining pressure condition,the strength of frozen silty clay will first increased and then decreased with the decrease of initial moisture content.With the increase of the number of freeze-thaw cycles,the strength of frozen silty clay gradually decreased.After 9freeze-thaw cycles,the strength tended to stabilize,and the confining pressure did not change the strain characteristics.Additionally,there was a correlation between initial moisture content,freeze-thaw cycles,confining pressure,and the microstructure,interparticle bonding strength,and ice lenses of frozen clay.In this paper,the constitutive relationship of frozen soil was described by Shen Zhujiang,binary medium and arc tangent model.Based on the statistical damage theory,a statistical damage constitutive model was proposed.It was found that the statistical damage constitutive model can well describe the strain softening characteristics of frozen silty clay.The determination coefficient(R~2)of the statistical damage-antitangent model and the statistical damage-Shen Zhujiang model was higher,and the root mean square error(RMSE)was lower.The model prediction results were consistent with the experimental results.In this paper,75 sets of experimental machine learning were conducted,60 sets of data were selected for training,and 15 sets of data were tested.The application of machine learning in predicting the deviatoric stress of frozen silty clay was studied.It was found that four machine learning ensemble regression algorithms,such as XGBoost,Light GBM,Random forest and Extra＿tree,had high accuracy in prediction.The frozen soil constitutive model based on the theory-driven long short-term memory deep neural network(TR-PSO-LSTM),which combined fitting data and time series data,can better predict the deviatoric stress of frozen silty clay after 18 freeze-thaw cycles.The minimum R~2of the model was 98.15%,which provided a useful reference for multi-source data fusion in the accurate prediction of frozen soil constitutive in seasonal frozen area.