Abstract: Hilly and mountainous regions are extensively distributed along the southeast coast in China, where various geological disasters occur frequently under the influence of factors such as typhoons and heavy rains. Debris flow disasters triggered by heavy rains in this area are characterized by suddenness, cluster occurrence, and certain destructiveness, and they may impose an impact on hydraulic structures and even cause serious hazards such as flow blockage if occurring near river channels or valleys. Because of the high uncertainty in their process, effective dynamic parameter inversion methods are of great significance for analyzing the characteristics of their evolution and formulating effective disaster mitigation measures. This paper develops a new parameter inversion method based on the multi-output support vector regression (M-SVR) sub-model, taking the "5·8 Taining Debris Flow" event as a study case. First, based on the debris flow dynamics calculation model Geoflow_SPH, we construct a parallel calling framework to numerically simulate the triggering feature parameter combinations－including internal friction angle, unit weight, and average source thickness－and generate initial training samples that contain input parameters and motion characteristics. And the sample set (1000 groups) is divided into training sets and test sets in proportion, and the M-SVR sub-model is trained using grid search technology. Then, we use the sub-model to predict and calculate the subdivided inversion calculation sample set (8000 groups), taking as the benchmark the debris flow velocity recorded in the geological survey report at three control sections, and selecting the parameter combination with the minimum mean square error (MSE) as the final inversion result through calculating the predictions. Finally, we examine the role of the drainage screen structure in obstructing the movement of debris flow and controlling the range of impact. The research results help clarify the evolution of rainstorm-type debris flow disasters and lay a theoretical basis for application of various disaster mitigation measures.

Key words: debris flow, multi-output support vector regression, parameter inversion, drainage screen