Evaluation of the collapse susceptibility of loess using machine learning

Evaluating collapse susceptibility of loess is essential for the construction of transportation lines in loess regions as it provides guidance for ground treatment. For the existing methods, a large number of boreholes need to be drilled along transportation lines to collect intact samples for laboratory tests, which make them very time and cost-consuming. In this study, loess’s collapse susceptibility is evaluated using Multi Expression Programming (MEP) and Back-Propagation Neural Network (BPNN). According to analysis of wetting-induced loess collapse, the gravimetric water content at the initial state ($w_0$), net vertical stress ($\sigma -u_a$), void ratio at the initial state ($e_0$), void ratio at the liquid limit state ($e_L$), and plastic index ($I_p$) are chosen as input variables. A comprehensive database incorporating 200 oedometrically soaking tests is established to train and test the two algorithms. The collapse potentials of loess are well predicted using MEP and BPNN, as demonstrated by high values of coefficient of determination ($R^2>0.88$) and small values of mean absolute error $(MAE<0.008)$ and root mean squared error $(RMSE<0.012)$. Following ASTM D5333-03 [5], the degree of loess collapse is classified with accuracies of 98 % and 90 % for MEP and BNPP respectively. Furthermore, sensitivity analysis shows the contribution of each variable to the prediction of collapse potential follows the order of $e_0>e_L>I_p>\sigma -u_a>w_0$. The machine learning is expected to assist the code of practice ASTM D5333-03 [5] in achieving an efficient site-investigation of collapsible loess for the construction of transportation lines.