Identifying subsurface contamination is challenging as the sources are not directly perceivable. Aquifer contamination only gets noticed when it is measured in one of the observation wells. As remediation of the contaminated sites is expensive and time-consuming, it is essential to locate sources of contamination for efficient remediation design and water resources management. Simulation-optimization approach is popularly used for identification of contaminant sources. However, the numerous runs required for the simulation model by utilizing the optimization algorithm makes this approach computationally expensive. Alternatively, the simulation model can be substituted by a surrogate model which can significantly reduce the computational cost. In this study, a deep neural network (DNN) based surrogate model is proposed for simulating the transport of a reactive contaminant Tritium in a hypothetical aquifer. The DNN is trained by considering injection rates at possible source locations as inputs and concentration at observation wells simulated by meshfree Radial Point Collocation Method (RPCM) as output. RPCM efficiently handles instabilities associated with advection and reaction dominant problems in comparison to grid/mesh-based methods. The backpropagation approach is used to optimize the weights and biases of the DNN using adaptive moment estimation (ADAM) as an optimizer. The performance evaluation of the surrogate model yields Mean Squared Error (MSE) close to zero and correlation coefficient (R2) of 0.99. An inverse model is developed by linking the DNN surrogate model and Particle Swarm Optimizer (PSO). The application of the inverse model show that the DNN-PSO model can predict the injection rates at possible source locations accurately.

The accuracy of streamflow forecasts is important for efficient monitoring and mitigation of flood events. Unfortunately, the uncertainty in the model control variable which includes model parameters, initial and boundary conditions, propagates through the model, resulting in the degradation of streamflow forecast. Various studies in the past have shown the potential of soil moisture assimilation in hydrological models resulting in the improved forecast. Further, the efficiency of assimilation is based on the number and the distribution of observations used. This study proposes a new approach called Forward sensitivity method (FSM), which operates in two phases. By running the model and forecast sensitivity dynamics forward in time, the first phase places the observations at or near where the square of the forecast sensitivity with respect to the control takes maximum values. Then using only this subset of observations, the second phase estimates the unknown elements of the control by solving a resulting weighted least squares problem. The power of this approach is demonstrated by assimilating ASCAT soil moisture observations into a conceptual Two Parameter Model in a medium sized watershed lying in the Krishna River Basin, India. The model run extends for four monsoon years from June 2007 to June 2011 and two assimilation scenarios were tested. The first scenario uses all the observations, whereas, the second uses only sensitive observations during assimilation and the results were then compared against open loop simulation (model run without assimilation). Sensitivity results indicate that observations during monsoon time alone are sufficient for assimilation purpose, which accounts for only 37.42 percent of total observations. Also, the estimation and forecast results show improved streamflow performance when using only sensitive observations. From the results, it is concluded that FSM based assimilation can help in reducing the computation time greatly. Further, this study will be critically helpful in the places where data availability remains a major problem.