The trouble with groundwater models
In the modern world where a development, let’s say a coal mine, extends down below the water-table, specialists almost universally use modeling software to try to guess what will happen when it does.
Hydrogeologists nod their heads and make canapé jokes about how much these models can be manipulated to say whatever their makers want them to say, yet they are virtually all that is used to predict groundwater impacts. The only exceptions are smaller developments where a hydrogeologist, engineer or plumber depending on the budget and how much the regulators are caring, use their expert knowledge, simpler analytical models and tea runes to make their guesses of impact.
The much more sophisticated, modern numerical models are all based on a brilliant program originally developed by the United States Geological Survey, called ModFlow. These models represent the ground and groundwater conditions as little cells which transmit simulated water volumes depending on their programmed nature and what water their neighboring cells are transmitting to it. The cells are subdivided into groupings representing strata, landscapes, surface water and subsurface elements, in a way that simplistically represents what the modeler understands of what happens in the earth. Then they run various scenarios through the model to simulate the impact of the mine or whatever it is, and from this predictions are made. There’s another important aspect about how models are “calibrated”, but I’ll get back to that.
In unbiased and clever hands, these models are wonderful tools, which along with expert knowledge help us accurately predict the impacts of the proposed development. In this lovely world, the predictions would be presented to the regulators (on behalf of the government and the people they serve) truthfully and fairly, in a way that specifically addresses all the key risks and reports the uncertainties of the predictions. This is not, sadly, our world.
In our modern world, the government agency responsible for assessing the groundwater bits of a proposed development will require that a model be prepared in order to predict the impacts. In most jurisdictions there are standards set for how models are to be developed for this purpose. A model will then be “built” by specialist groundwater consultancies, hopefully with the benefit of having completed some investigations to allow subsurface conditions to be approximated. The model will invariably be paid for by the developer/proponent, and will usually be presented in a report as an attachment to the Environmental Impact Statement.
In most states, somewhere in an airless basement of one of the agencies will sit a government hydrogeologist, either singly or in a small herd. They will be given the model report and other information if they’re lucky, and asked to decide (usually within days) if its predictions are true, or at least plausible, and whether, in the agency’s eyes and according to the planning laws of the land, the proposal impacts are acceptable.
Dear reader, I am one of these privileged and underlit groundwater regulators, and I am here to tell you that these models are currently dangerous, and are being used in ways that vary from offering a pretty good prediction to committing environmental fraud.
In my possibly jaundiced view, there are three things badly wrong with modern numerical groundwater models:
The models themselves are almost completely inaccessible to the regulators. They are normally presented as reports, describing the key features of the model and what it predicted. Even if the model itself is demanded by the agency, it’s very unlikely that the regulator will have the skill (I don’t) or time to look under the bonnet; to question how the model has been built, what assumptions it really makes and whether prediction scenarios could be better formulated.
The model is not designed to answer the question(s) that the regulator or community really wants to know. Groundwater modellers focus their craft on getting the model to represent ground and groundwater conditions in a way that doesn’t defy one or more laws of nature, and to make their modeled world consistent with a selection of field observations, the process of “history matching” we call model calibration.As the guru of model calibration Dr John Doherty has said in many forums, a model should not be designed just to represent the world and an impact scenario. Each model should be very specifically designed to predict the likelihood of a negative consequence (what the regulator or community is really worried about, such as pollution reaching a water supply or groundwater levels falling below a farmer’s well), and to do it in a way that there is only a very small risk that the prediction will turn out to be wrong.
The way that modeled uncertainties are typically presented is untruthful and frequently misleading. It is not possible to know the earth’s subsurface completely as we must interpret it from boreholes, outcrops and the collective knowledge of what has been studied elsewhere. Even if we did have perfect knowledge of ground and groundwater conditions, models must make simplifying assumptions if they are to be able to make predictions within the hours or days allotted to them. These and other issues mean that the models and modellers must make a very large number of judgement calls and assumptions about how they represent the groundwater regime.Despite all of this being known, most modeling reports refer only to simplistic measures of uncertainty and leave the important remainder unsaid. In truth, the real uncertainty involved in a model prediction cannot normally be measured as incorrect conceptualisations, measurement and structural model noise and the way that the model attempts to internally “fix” parameters which are preventing it from reaching “convergence” are rarely or never known. Nevertheless, much more clarity could be provided about the uncertainty of a prediction, but it is simply not in the modeller’s or developer’s interest to disclose it.
There are simple solutions to the above issues, but these solutions will take more effort by modelers and won’t happen until the regulators insist that they do. I would be happy to share my thoughts on how this could be done with anyone, and welcome any thoughts you might have as well.