A Parameter Inversion Framework For Computationally Expensive, Physically-Based Hydrologic Models Using Deep-Learning Methods
Presenter: Amanda Triplett P241
Co-Author(s): -
Advisor(s): Laura Condon
1Hydrology & Atmospheric Sciences
Inverse modeling has long been used in the field of hydrology to estimate causal parameters that are difficult to observe from their easier to measure effects. In this study, we aim to use a convolutional neural network (CNN) to predict the causal parameter hydraulic conductivity from the resulting water table depth. To achieve this, we create training images by running an ensemble which varies hydraulic conductivity to establish its relationship with water table depth. We use the physical groundwater-surface water model ParFlow to represent a domain across the majority of the contiguous United States. Next, a CNN uses water table depth, along with additional variables such as elevation, slope and long term recharge, to predict hydraulic conductivity. Finally, the CNN generated hydraulic conductivity fields are run through ParFlow and the resulting water table depth fields are compared to the originals. Our results show that we can reproduce hydraulic conductivity fields for various out of sample cases, as well as get excellent matches between the resulting water table depth fields. Additionally, the CNN accurately reflects physical relationships by its shifting of the hydraulic conductivity prediction according to increases and decreases in long term recharge. In this study we also use the model performance to address the strengths and limitations of this approach and when it should and should not be used. This framework outlines a powerful approach for parameter inversion which is especially impactful for computationally expensive physical hydrologic models like ParFlow, for which it is difficult to use traditional methods.