In the past, typical approaches simulating the groundwater flow are based on continuous trials to approach to the targeted measurement accuracy. In this study, we propose a new neural network based on feedback observer technique to estimate the hydro-geological structure and hydraulic parameters of a large-scale alluvial fan in Taiwan. We develop an under-ground water level observer (UGW-LO) based on feedback control theory to simulate the dynamics of groundwater levels and estimate water levels of wells in the large area. In the proposed observer system, a large-scale back-propagation neural network (BPNN) is proposed to simulate water levels dynamics of multiple wells. The simulation results are fed back as a reference for BPNN to approach to refine estimation. Based on that model, a groundwater flow is simulated correctly by software MODFLOW. Experimental results indicate that the innovative method works better than conventional regression estimations. The learning ability of BPNN also contributes to overcome the gap between legacy dynamics UGW equations and real UGW dynamics. The applicability and precision are verified in a large scale experiment that is beneficial to the management of underground waters and reduce the risk of ground-sink.