Abstract: Agricultural development in the Hetao Irrigation District has long been constrained by water scarcity and soil salinization. Late autumn irrigation, an innovative practice conducted during the soil freezing period, integrates salt leaching with moisture conservation, demonstrating significant water-saving potential and promising application prospects. Frozen soil infiltration is a critical physical process during late autumn irrigation, directly controlling the distribution of irrigation water within the soil profile and the efficiency of salt leaching. However, a systematic understanding of the intrinsic mechanisms governing frozen soil infiltration and salt redistribution under the coupled effects of multiple factors—freeze-thaw cycles, initial water content, and tillage practices—remains lacking. To realistically simulate the late autumn irrigation scenario in the irrigation district, this study conducted indoor soil column freezing-infiltration experiments. The experimental design was fully based on local field observation data and agricultural practices. Treatments included different initial soil water contents (20%, 25%, and 30%), with or without tillage, and with or without freeze-thaw cycles (10 cycles). The dynamics of soil temperature, moisture, and salt content profiles were systematically monitored, and the variations in infiltration rate and desalination rate were analyzed. This aimed to reveal the influence of key late autumn irrigation conditions on the frozen soil infiltration process and its underlying water-heat-salt coupling mechanisms. Furthermore, by comparing the predictive performance of three traditional infiltration models (Horton, Philip, Green-Ampt) and two deep learning models (LSTM, LSTM-Attention), this study provides a basis for model selection in simulating frozen soil infiltration under complex conditions.The results indicate that freeze-thaw cycles (FTCs), tillage practices, and initial soil water content jointly regulate frozen soil infiltration and desalination processes through coupled water-heat-salt dynamics. The frozen soil infiltration process sequentially undergoes three distinct stages: initial rapid infiltration stage, phase-transition stage, and stable infiltration stage. Compared to unfrozen soil, pore ice blockage reduces the average infiltration rate. However, FTCs significantly enhance the final infiltration capacity of the soil under identical moisture and tillage conditions by improving soil structure and increasing pore volume and connectivity. The initial water content primarily governs the duration of the phase-transition stage; higher moisture levels prolong the ice-water phase transition process, thereby reducing the average infiltration rate during this stage. Tillage accelerates initial infiltration by increasing surface porosity and enhances the infiltration-promoting effect of FTCs by intensifying thermohydraulic exchange processes, consequently effectively shortening the duration of the phase-transition stage.FTCs significantly increased the overall desalination efficiency by 25.63～46.04% compared to unfrozen conditions, promoted more uniform salt leaching across soil layers (desalination rates reached 69.49～89.26%), and significantly reduced the spatial heterogeneity of salinity. A clear coupling effect was observed between initial water content and tillage on desalination efficiency: under high water content conditions (W3), tillage decreased desalination efficiency regardless of FTCs, primarily due to facilitating surface crust formation. In contrast, the peak desalination efficiency was achieved with the combination of tillage and moderate water content (W2). Regarding the prediction of frozen soil infiltration, the accuracy of traditional models decreased significantly: the Philip model performed poorly (R²= 0.140), and the Green-Ampt model parameters exhibited physically ambiguous abnormal shifts (the model parameter for effective pore volume, \theta _s-\theta _l decreased from 0.7574 to 0.503). Although the Horton model maintained relatively high accuracy (R²= 0.886), its relative root mean square error increased to 16.33%, indicating limited applicability in frozen environments. The LSTM-Attention model, leveraging its powerful capability for nonlinear fitting and temporal feature capture, outperformed the traditional models and the standard LSTM, achieving excellent performance with R²= 0.999 and RRMSE= 1.24%. Future research should focus on integrating physical mechanisms into deep learning models to overcome their "black box" limitations and extend their application to regional scales through distributed modeling approaches, thereby providing reliable tools for precise water and salt regulation in the irrigation district.