<div>Understanding groundwater flow dynamics within the Great Artesian Basin (GAB) is critical for responsible management of groundwater from an environmental, economic and cultural perspective. Numerical groundwater flow modelling involves generating a simplified representation of a groundwater system and using Darcy’s Law to simulate groundwater flow rates and the distribution of hydraulic heads throughout the system. This is a pilot study aimed at developing a workflow for groundwater flow modelling of the Great Artesian Basin using Bayesian methods. In this report, we present our initial results from building and running a steady-state groundwater flow model of the entire GAB. We demonstrate a Bayesian inference framework to generate an ensemble of groundwater flow models allowing an assessment of the uncertainty of model parameters and flow velocities.&nbsp;</div><div>Several models have been built to simulate groundwater flow across various areas and layers of the GAB. Most of these models aimed to predict the likely impacts on the groundwater system of some future scenario, generally climate change or groundwater extraction relating to mining activities. While these models are well-suited to their purpose, their focus on particular regions or aquifers makes them unsuitable for investigating large-scale groundwater flow throughout the GAB. In contrast, the model built as part of this study captures the entire GAB and aims to simulate large-scale flow. Although not in scope for this pilot study, the questions a model at this scale is capable of addressing include characterising 3D flow within hydrogeological layers, computing groundwater flux between aquifers and between sub-basins, inferring hydraulic properties and identifying poor quality data. As this model is steady-state and uses hydraulic head data from before the year 2000, it provides a baseline estimate of groundwater flow without considering recent anthropogenic forcing or transient system stresses.&nbsp;</div><div>The GAB is represented as a 14 hydrogeological layer model including basement, Permo-Carboniferous basins, Mesozoic sedimentary aquifers and aquitards and Cenozoic layers. This includes updated hydrogeological surfaces from the GAB project. The input data consisted of 8,065 hydraulic head measurements and 6,151 estimates of recharge rate while the model parameters were a single hydraulic conductivity value for each of the 14 layers. The modelling domain was discretised using 10 x 10 km cells in the horizontal plane and the mesh was deformed vertically to fit between the topography and the basement surface, with the resulting mesh having a vertical discretisation of no coarser than 50 metres. The top boundary condition was a constant head boundary that was a smoothed version of topography. The sides and bottom of the model have no flux boundary conditions and a buffer zone around the GAB was included to minimise boundary effects.&nbsp;&nbsp;</div><div>In total 2500 groundwater flow simulations were run using a Bayesian inversion framework. The inversion sampled various combinations of input parameters to find models with a relatively low misfit, which was calculated by squaring the difference between the observed and simulated values of hydraulic head and recharge. Rather than searching for a global minima, the Metropolis Hastings Markov Chain Monte Carlo sampling algorithm was used to explore a range of possible models and estimate the posterior distribution of each layer’s hydraulic conductivity.&nbsp;</div><div>The model performed adequately and the model parameters were generally consistent with the prior probability distributions based on previous modelling studies. However, the posterior distribution of model parameters were very broad indicating the model was not particularly informative in its current form.&nbsp;&nbsp;</div><div>Groundwater flow velocity vectors from the maximum likelihood model were used to investigate groundwater trends within the Cadna-owie-Hooray aquifer. Uncertainty of model predictions were investigated by calculating the groundwater flow velocity variance across the ensemble. This study demonstrates that it is technically feasible to use Bayesian inference to probabilistically mode groundwater flow across the entire GAB. However, for this approach to yield useful results, more work is required to understand the impacts of simplifying assumptions about layer properties, the quality of the input data and model structure on the resulting flow model.&nbsp;</div><div><br></div>

<div>Surface magnetic resonance (SMR) techniques image subsurface water using the electromagnetic response of resonant hydrogen nuclei in water. Here we introduce the SMRPInv (Surface Magnetic Resonance Probabilistic Inversion) package, which couples a high-performance forward modeller for SMR data, and a Gaussian process based non-linear Bayesian inversion. Both the forward and inverse codes are part of the freely available, open source HiQGA (High Quality Geophysical Analysis) codebase written entirely in Julia. We summarise the relevant forward physics, the necessary data processing of free induction decay at an SMR sounding, followed by the estimation of subsurface water content with a non-linear parameterisation. Results are presented for synthetic inversions as well as field data from Western Davenport (Northern Territory). Comparisons are made against downhole logging data, together with results from a deterministic inversion of the same SMR soundings. Through this, we demonstrate that a probabilistic approach is key to conceptualising variability of subsurface water content.&nbsp;</div>