Abstract: The complex geomorphic units and active geological structures in Tibet provide a good breeding environment for debris flow in the region, but also pose a great threat to human life and property. The evaluation of debris flow susceptibility can identify key areas for disaster prevention and reduction in the region. Taking Bhumi County and Medog County of Tibet Autonomous Region as the study area, 12 factors with high influence on debris flow, including elevation, slope, stratigraphic lithology and rainfall, were selected by Pearson Chi-square test algorithm as evaluation indexes, and 282 debris flow points and non-debris flow points in the study area were taken as sample database. Based on ArcGIS platform, four susceptibility evaluation models were established by using information method and machine learning method, and ROC curve and AUC index were introduced to evaluate the susceptibility accuracy of debris flow. The results show that: (1) Considering the different types of debris flow in different dimensions and the different controlling factors, the normalization coefficient of latitude and air temperature is used as the evaluation index of debris flow susceptibility, which can eliminate the excessive response of debris flow to temperature in low altitude areas to a certain extent. (2) Air temperature, distance from water system, distance from road, formation lithology and elevation are the main factors of debris flow occurrence in the study area; Factors such as vegetation coverage, terrain humidity, and slope also play an important role. (3) Considering the relationship between the disaster point of debris flow and the classification attributes of the impact factors, the classification attributes of the impact factors were assigned scores and trained as input features. The prediction effect of the machine learning model was good, and the average AUC was 0.980, which was better than the traditional information model on the whole. (4) The AUC of SVM model is as high as 0.987, the FR value of the highly prone region is 41.13, and the prediction area of highly prone regions takes up the smallest proportion, so it has the ability to perform high-precision prediction in large-scale regions.