2019 Darcy Lecture

John Doherty, Ph.D.

John Doherty, Ph.D., is the author of PEST, a software package that is widely used for groundwater model calibration and uncertainty analysis. He has worked for more than 35 years in the water industry, first as an exploration geophysicist and then as a modeler. Doherty been employed by both government and industry, and has also worked at numerous universities where he undertook research and supervised postgraduate students. Currently he works for his own company, Watermark Numerical Computing, doing consulting, research, programming, and education, mainly on issues related to model deployment in support of environmental management and impact assessment.

Doherty Will Offer a Choice of the Following Two Lectures at Participating Venues in 2019:

“Dancing with Models - The Importance of Model Partner Software”

Numerical simulators of groundwater flow and transport cannot fulfill their decision-support potential on their own. Instead, they must be used in partnership with equally sophisticated software that links models to data acquired at sites that they simulate - and that links models to the decisions that they are intended to support. As they “dance” with a model, these software packages can accomplish tasks such as history-matching, uncertainty analysis, predictive hypothesis-testing, sensitivity analysis, management optimization, and optimization under uncertainty.

The importance of model partner software is not nearly as widely appreciated in the groundwater industry as it should be. Model developers often design input/output protocols that make linkage to partner software difficult or impossible. Model users often build models that are unnecessarily complex, take too long to run, and have questionable numerical health. Education in model-value-adding numerical algorithms is rarely offered to modelers by universities. Graphical user interfaces do not provide comprehensive support for the wide range of ancillary tasks that decision-support modeling requires.

This lecture will explore how models can best serve the decision-making process. In doing so, it demonstrates the indispensable role that model-value-adding software should play in this process. It also addresses some currently available packages, as well as an easy-to-use, public domain, parallel model run manager with a nonintrusive model interface that allows rapid development of model partner software by any programmer.

"Starting from the Problem and Working Backwards"

Many groundwater models are commissioned and built under the premise that real world systems can be accurately simulated on a computer - especially if the simulator has been “calibrated” against historical behavior of that system. This premise ignores the fact that natural processes are complex at every level, and that the properties of systems that host them are heterogeneous at every scale. Models are, in fact, defective simulators of natural processes. Furthermore, the information content of datasets against which they are calibrated is generally low.

The laws of uncertainty tell us that a model cannot tell us what will happen in the future. It can only tell us what will NOT happen in the future. The ability of a model to accomplish even this task is compromised by a myriad of imperfections that accompany all attempts to simulate natural systems, regardless of the superficial complexity with which a model is endowed. This does not preclude the use of groundwater models in decision-support. However it does require smarter use of models than that which prevails at the present time.

It is argued that, as an industry, we need to lift our game as far as decision-support modeling is concerned. We must learn to consider models as receptacles for environmental information rather than as simulators of environmental systems. At the same time, we must acknowledge the defective nature of models as simulators of natural processes, and refrain from deploying them in a way that assumes simulation integrity. We must foster the development of modelling strategies that encapsulate prediction-specific complexity supported by complexity-enabling simplicity. Lastly, modelers must be educated in the mathematics and practice of inversion, uncertainty analysis, data processing, management optimization, and other numerical methodologies so that they can design and implement modeling strategies that process environmental data in the service of optimal environmental management.


Powered by Firespring