Experimental and Modeling of Residual Deformation of Soil–Rock Mixture under Freeze–Thaw Cycles

Projects in seasonal frozen soil areas are often faced with frost heaving and thawing subsidence failure, and the foundation fill of most projects is a mixture of soil and rock. Therefore, taking soil–rock mixture with different rock contents as research objects, the residual deformation of soil–rock mixture under multiple freezing–thawing cycles is studied. In addition, the deep learning method based on the artificial neural network was pioneered combined with the freezing–thawing test of the soil–rock mixture, and the Long short-term memory (LSTM) model was established to predict the results of the freezing–thawing test. The LSTM model has been verified to be feasible in the exploration of the freeze–thaw cycle law of a soil–rock mixture, which can not only greatly reduce the period of the freeze–thaw test, but also maintain a high prediction accuracy to a certain extent. The study found that the soil–rock mixture will repeatedly produce frost heave and thaw subsidence under the action of freeze–thaw cycles, and the initial frost heave and thaw subsidence changes hugely. With the increase of the number of freeze–thaw cycles, the residual deformation decreases and then becomes steady. Under the condition that the content of block rock in the soil–rock mixture is not more than 80%, with the increase of block rock content, the residual deformation caused by the freeze–thaw cycle will gradually decrease due to the skeleton function of block rock, while the block rock content’s further increase will increase the residual deformation. Furthermore, the LSTM model based on an artificial neural network can effectively predict the freezing and thawing changes of soil–rock mixture in the short term, which can greatly shorten the time required for the freezing and thawing test and improve the efficiency of the freezing and thawing test to a certain extent.