Abstract: Under dynamic loading, the deformation and dynamic modulus of weak soil, such as soft clay, can change. This study aims to investigate the effect of long-term dynamic loads on soft clay by collecting dynamic parameters through dynamic triaxial testing. The analysis focuses on exploring the relationship between the softening index, cumulative plastic strain and cyclic stress ratio (CSR). The data exhibits a certain concentration, prompting the utilization of machine learning methods. Using a logistic regression classifier to determine soil stability, it is found that soil is less likely to collapse with lower CSR, higher loading frequency and increased softening index. The effect of frequency is observed to be less significant than that of CSR. To establish a connection between dynamic parameters and the static behavior of soil samples, a random forest model is used to predict the softening index, which is challenging to obtain directly. The prediction results demonstrate a good fit of the model and pass the significance test. The model achieves a precision of 0.93 for the training set and 0.79 for the test set. It is revealed that cumulative plastic strain plays a crucial role in predicting the softening index, outweighing other factors such as undisturbed or remolded soil, overconsolidation ratio (OCR), loading waveform, frequency, CSR and soil state type (stability/failure) based on deformation curves. This suggests promising prospects for applying the trained model and machine learning techniques.

Key words: dynamic triaxial test, cumulative plastic strain, softening index, logistic regression, random forest