Abstract: In order to investigate the changes of the additional force in the back of the shaft lining wall, which is caused by the thawing effect of the frozen wall in the construction of the special sinking shaft with freezing method, indoor freeze-thaw cycle tests have been carried out under different moisture contents, unidirectional temperature gradients and external loads, taking Xuzhou clay as the object of study. The results show that: when the test system was under water replenishment, the clay under 1.4 ℃ /cm unidirectional freezing gradient has 0.98 mm settlement, which is 60.6% larger than the clay with the same dry density but under 2.0 ℃/cm freezing gradient (with 0.61 mm in settlement). Under the same unidirectional freezing temperature gradient, the thawing settlement increases with the increment of load, of which the increasing trend of is similar while the increasing magnitude is not consistent. Based on the fact that the thawing character of freezing soil is effected by multiple factors, a relational database about thaw settlement coefficient with multi-sample and multi-factor has been built by utilizing the improved artificial neural network. The error analysis certifies that the improved prediction algorithm has greater accuracy.

Key words: artificial frozen soil, temperature gradient, thaw settlement coefficient, neural network