This study aims to establish an inverse model for reconstructing properties of turbidity currents from their deposits. Although several attempts of in-situ measurements have been succeeded, flow properties of large-scale turbidity currents that build submarine fans remain unclear. Recurrence interval of turbidites in natural levees or lobes of submarine fans is generally 100s years, while in-situ observation in modern submarine canyons reported yearly to monthly occurrence of turbidity currents. That is, turbidity currents that substantially form submarine fans may be extremely low frequency and large-scale than currently observed. To understand the behavior of fan-building flows occurring with sub-millennial scale recurrence interval, the quantitative reconstruction of flow conditions from geologic records is required. However, inversion of flow properties in natural scale has been impractical because the heavy computational load of 2D or 3D models of turbidity currents has prevented to perform inversion requiring iteration of calculations. To this end, we propose a method using 2D shallow-water model with the deep-learning neural network (DNN). In this method, a horizontal 2D shallow-water model of turbidity currents was employed as the forward model. Numerical simulation was repeated to obtain horizontal distribution of thickness of turbidites under various initial conditions, and then this synthetic data set was used for supervised training of DNN. After the training phase finished, DNN properly estimated initial conditions of turbidity currents (e.g. initial flow height, velocity, etc.) from artificial test data set that was also produced from the forward model. In previous methods, the computational cost of 2D model was too high to be employed as the forward model for turbidite inversion, while our methodology can reconstruct initial conditions of turbidity currents instantaneously. Although production of the training data set requires 1000s times repetition of calculation, all calculation of forward model can be perfectly parallelized, so that the problem of calculation efficiency can be easily solved by using PC cluster. Future application of our method to actual deposits will contribute to understand the development processes of submarine channels or lobes in the long term.