Deep learning emulators for groundwater contaminant transport modelling

Abstract Groundwater flow and transport models are routinely applied for contamination risk assessments and remediation plan design. The computational burden of such models has limited their application when a large number of model runs are required to conduct history-matching, sensitivity analysis, uncertainty analysis and optimisation. Deep learning has shown great potential in relaxing this limitation by emulating the process-based models for computational-intensive application. The current study investigates the feasibility of implementing deep learning emulators for a generalised flow and contaminant transport model based on the hydrogeology of an aquifer located in South Australia where contamination risk from on-shore gas development needs to be assessed. Three types of representative predictions are emulated: (i) contaminant concentration at a specific location and time (point variable), (ii) objective functions used for model calibration and uncertainty analysis (lumped variable), and (iii) contaminant breakthrough curves (sequence variable). Our study demonstrates that accurate, efficient and scalable deep learning emulators can be achieved with only a few thousand training samples for a simple transport model. Equipped with the fast-running DNN emulators, contaminant transport models can be applied in more sophisticated environmental modelling to support risk-based decision making. The study sheds light on hybrid modelling that combines the strength of process-based environmental models and data-driven machine learning models.