Research On Migration Of Scale Effect Of Coarse-Grained Soil With Machine Learning

Coarse-grained soil is a key construction material in geotechnical engineering,and its mechanical properties play an important role in engineering applications.However,due to limitations in size,time,and fund,triaxial apparatus used to measure the mechanical properties of coarse-grained soil are often mismatched with the particle size of field soil samples,making it difficult to directly measure the properties of largeparticle-size soil samples.Traditional methods typically rely on mathematical models based on test data of scaled soil samples to estimate the properties of original soil samples.However,due to the scale effect,the estimated results may have significant errors.In addition,simulation formulas are mostly based on empirical data,including many parameters that are difficult to determine their physical meaning and have limitations.In recent years,researchers have applied machine learning to reproduce the stress-strain behavior of coarse-grained soil with the maximum particle size less than 60mm and demonstrated the potential of machine learning.However,current research has not yet addressed the study of the mechanical properties of large coarse-grained soil.This paper aims to use the correlation between the mechanical properties of coarse-grained soil with different maximum particle sizes for transfer learning,and predict the deviatoric stress of coarse-grained soil with the maximum particle size more than 60mm and even up to 200mm based on machine learning methods.Besides,the model compression is introduced to reduce the model complexity and improve its generalization.The main innovation points and contributions of this paper are as follows:(1)To address the lack of available coarse-grained soil triaxial test data sets,4 data set have been established in this paper,containing 25 108 data points from 568 triaxial tests conducted over the past 50 years,both domestically and internationally.The data sets are diverse and the maximum particle size can reach 200mm.In addition,4 data sets are divided and preprocessed based on the maximum particle size of the coarsegrained soil.(2)Based on the similarity and distribution characteristics of the data sets,an incremental regression transfer learning model iReTL is proposed to learn the correlation between different maximum particle size soil samples from data sets,and to explore the deviatoric stress curve of large-grained coarse-grained soil.Experimental results show that the proposed iReTL exhibits good performance in predicting the deviatoric stress of coarse-grained soil.(3)To address the potential overfitting problem caused by the small amount of data of large coarse-grained soil,iReTL learns information from data sets of soils with different maximum particle size for indirect data augmentation.On the other hand,the model compression and regularization terms are introduced to reduce model complexity and improve generalization.Experiments show that model compression simplifies the model structure,and reduces complexity while not sacrificing or even improving model performance.