Postgraduate Profiles

David Schafer

David Schafer profile photo

Thesis: Use of surrogate models for effective reactive transport model calibration and predictive uncertainty

In order to maintain the health of groundwater systems and to remediate those systems that have been contaminated, numerical models need to be developed to address the feasibility of potential management scenarios. However, these groundwater systems can be complex both in their hydrogeological as well as their geochemical structures. Therefore, making accurate predictions for these systems requires complex groundwater flow and reactive transport models whose parameterisations may be complex and model run times may be quite long. Due to these factors, addressing the calibration and predictive uncertainty of complex reactive transport models is often infeasible and often neglected.

A potentially feasible methodology to undertake uncertainty analysis in a computationally more efficient way is to undertake surrogate modelling, where the complex reactive transport model is partially replaced with a simplified, computationally more efficient model. However, simplified surrogate models will produce some undesired bias that must be considered when addressing the predictive uncertainty of the complex-model. Therefore, in theory, the best simplified-model is one in which both the run times are feasible, and the introduced bias is minimised. However, this may not be achievable with a single surrogate model and a sufficiently minimised bias may only be possible if multiple simplified models, each derived through a different method of simplification, were used in concert. The primary objective of this PhD research project is to develop and test a robust multi-model approach that has general applicability to complex reactive transport models.

To address the research questions the project will initially define a high-resolution complex synthetic reactive transport model. This “virtual aquifer” model will be used to investigate:

1. Methods for optimal simplified/surrogate model selection (with minimized bias) when representing a complex reactive transport model. Is the “optimal” surrogate effective enough? Is any surrogate effective when it comes to reactive transport?

2. Can an effective surrogate model prediction actually consist of an “average” prediction among several surrogate models? Could methods like Bayesian Model Averaging be used to accumulate the predictions of several simplified models into one surrogate prediction whose bias is minimized?

Following the development of a new approach the methodology will be tested for a case study involving the large-scale implementation of the Perth groundwater replenishment scheme. This case will be studied to investigate:

3. Can the new algorithms/methods be used alongside methods of model discrimination to effectively evaluate the predictive uncertainty for a real world transport model that not only contains parameter error but also structural error (i.e., many conceptual models)?

Why my research is important

Groundwater supplies 43% of the total consumptive irrigation water worldwide (approximately 545 000 GL/annum) and is the source of drinking water for three quarters of the people in the countries of the European Union and about 70% of the population of China (Siebert et al., 2010). In Western Australia, over 55% of the distributed public water supply (approximately 173 GL/annum) water is sourced from groundwater (Pink, 2012). Groundwater use worldwide increased spectacularly from around 100 000 GL/annum to around 1 000 000 GL/annum in the second half of the twentieth century (Burke and Villholth, 2007).

The ‘intensive use of groundwater’ mainly from the twentieth century onwards has led to many concerns and management issues (Findikakis and Sato, 2011). A listing of ‘undesired effects’ of intensive groundwater is given in Bear and Cheng (2010) and includes concerns about rises and falls in water levels, maintaining groundwater discharge base-flow to rivers and springs, maintaining water quality, movement of the seawater wedge in coastal aquifers and ensuring residence times for managed aquifer recharge schemes. How best to protect, control, prevent and remediate such ‘undesired effects’ forms the basis of groundwater management practice (Findikakis and Sato, 2011).

In addition to the undesired effects due to the intensive use of groundwater, groundwater contamination remains a major groundwater management concern in many countries and a still increasing concern in rapidly industrialised countries such as China. Contaminants may include pathogenic organisms (e.g. bacteria, viruses), inorganic contaminants (e.g. salts, heavy metals), and organic contaminants (e.g. chlorinated hydrocarbons, aromatic hydrocarbons) (Bear and Cheng, 2010, Appelo and Postma, 2005). The contamination may be mobilised in-situ (e.g. arsenic, fluoride, salt, acid mine drainage) or may be due to discharge to the aquifer (e.g. nitrate from agricultural practices, sewage, industrial waste). A detailed classification of the different sources of groundwater contamination is contained in the US Congress Office of Technology Assessment report (OTA, 1984). Contaminant sources may be broadly described in terms of whether they are point sources (e.g. septic tanks, storage tanks, transportation spills etc) or diffuse sources (e.g. agricultural application of fertilizers and pesticides across broad areas). They may also be described according to their temporal rate of release i.e., a one-time spill or a continuous spill (Bear and Cheng, 2010). Some contaminants become dissolved and move with the groundwater and/or interact with the aquifer matrix (solute transport) while other contaminants may occur as separate non-aqueous phase liquids (NAPLs) (Zheng and Bennett, 1995).


Jun 2013

Jun 2017