Groundwater contaminant source characterization through artificial neural networks

Water plays a crucial role in human life and in all its activities. For this reason, all water resources and in
particular groundwater should be managed in a sustainable way in order to satisfy current needs and without
causing environmental consequences. Unfortunately, economies based on intensive agriculture and industrial
production lead to unsustainable use of water, the effect of which also includes the contamination of aquifers.
In this context, the identification of the location of the contaminant source with its release history has attracted
great attention within the scientific community called upon to provide theoretical methods to limit the spread
of the contaminant. To identify remediation strategies immediately is essential to have a tool that can provide
accurate results in real time. With this aim, surrogate models can become the conceptual models of primary
choice being able to study forward and inverse transport problem using a number of observations, which is
not much greater than the unknown parameters to be calculated, reducing in this way the computational cost
compared with other more complex models. Data-driven surrogate models lead to the field of Artificial
Intelligence where neural networks, trained on a finite dataset, are able to estimate the desired output by
means of a learning process emulating the behavior of the human brain.
In this work, a feedforward artificial neural network (FFWD-ANN) has been developed to analyze different
cases as surrogate model. The investigated domain has been selected from a literature study (Ayvaz, 2010)
and the training dataset has been randomly developed by means of the Latin Hypercube Sampling in order to
reduce the number of forward simulations. Initially, the network has been trained to solve forward transport
problem. In the proposed approach, the ANN well estimates the pollutant concentrations in 7 monitoring
wells, at different times, by using as input data the release history at two contaminant sources with known
locations. Then, the surrogate model has been trained to deal with inverse transport problem related to
different application cases: 1. estimation of the release history at one contaminant source with known location;
2. simultaneous estimation of the release history and location of one contaminant source; 3. estimation of the
release history at two contaminant sources with known location; 4. simultaneous estimation of the release
history at two contaminant sources with known location and error on observations.
The results have been compared with literature data (Ayvaz, 2010; Jamshidi et al. 2020). Artificial Neural
Network seems to be well suited to dealing with this type of forward and inverse problems, preserving the
reliability of the results and reducing the computational burden of numerical models.