Application of Machine Learning To Predict Cyclic Resistance of Silty Sands

Abstract

Machine Learning, an effective tool, is being utilized in geotechnical engineering owing to the amount of data generated in the past. This work utilizes a machine learning based framework to identify a suitable density measure for the cyclic resistance of silty sands and subsequently tests the ability of several ML algorithms in producing the cyclic resistance curve. For this purpose, the published literature is considered a potential source for collecting the results of cyclic triaxial tests conducted at confining pressure of 100KPa on moist-tamped samples of silty sands. The compiled data includes several influencing parameters like the Number of cycles to cause initial liquefaction (N), grain size (D50), coefficient of uniformity (Cu), void ratio, relative density, equivalent void ratio, relative compaction, emin, emax, erange, and fines content. Performing feature selection indicates relative compaction and erange as the most influential parameters. Relative compaction even proved a better parameter than relative density in terms of normalizing the effect of fines on cyclic resistance. On comparing the accuracy of the several models created, Gaussian Process Regression (GPR) emerged as the best-performing algorithm for this problem. The GPR model is then validated on the unseen data. It is noticed that the model predicts the variation of the CSR with the Number of cycles (N) quite satisfactorily but over-predicts with an error up to 25%. This error can be attributed to the combined effect of other minor factors that are not included as input parameters. It can be summarized that relative compaction and erange are capable of predicting the curve with reasonable accuracy.