[M-IS15] Geophysical particulate gravity current

[MIS15-P04] Estimation of rheological properties of mass movements using convolutional neural network

Komoriya Masaya¹, *Hajime Naruse¹ (1.Department of Geology and Mineralogy, Graduate School of Science, Kyoto University)

Keywords:debris flow, landslide, machine learning, shallow water equation

An inverse model was constructed to estimate the initial flow thickness, concentration, yield stress, and viscosity at the onset of the flow. A total of 3220 datasets of initial conditions and sediment distributions were created, and they were used to train the CNN., The CNN model was then tested by 20 unknown datasets not used for training. The result indicated that the CNN model can precisely estimate the flow rheological parameters and the initial depth from mass transport deposits. At the same time, assessing the sediment concentration in the flow is difficult. The inverse model developed in this study was applied to the landslide deposit of the 2018 Eastern Iburi Earthquake and the debris flow deposit of the slope failure in Ashikita-machi, Kumamoto Prefecture, Japan, associated with the July 2020 torrential rainfall event. The results indicated that the yield stress and viscosity coefficients of mass transport in each event were 8850 Pa and 7766 Pa S, 10008 Pa, and 9201 Pa-S, respectively.
These values are close to the rheology of the gravel-rich flow estimated in the laboratory experiments of existing studies. The obtained estimates were given to the forward model to calculate the behavior of the process. As a result, the distributions of the two event deposits were well reproduced, suggesting the validity of the model estimates. The inverse model is expected to be a practical method for assessing the risk of future slope disasters.