Data Assimilation and Uncertainty Analysis in Decision-Support Modelling

  • Course dates: October 30th - November 3rd, 2023
  • Speaker: Dr. John Doherty
  • Course location: UNINE Unimail building, rue Émile-Argand 11, Neuchâtel, Switzerland. (Mon: room C001, Tues-Fri: room A317)
  • Information on PEST: pesthomepage.org
  • Cost: 
    • PhD student members of the Water-Earth Systems PhD School: no fee
    • External PhD students, other academic members, and those in government and industry: please contact School.Earth-Water@unine.ch
  • Registration: please fill in the REGISTRATION FORM
  • Questions: please contact School.Earth-Water@unine.ch 
  • Detailed programme

Course Description

The course is conducted by John Doherty, author of PEST and 2019 Darcy Lecturer. It will show you how to use and understand PEST and PEST++. However, it will consider much more than this. You can benefit from attendance of this course even if you never use PEST again, and even if you are not a modeller.

The course examines the theory and practice of model calibration and uncertainty analysis. In doing so, it examines the role of history-matching in environmental modelling (especially groundwater modelling). It goes on to examine what decision-support groundwater modelling can and cannot achieve. This has repercussions for how models are used in decision-making. The course explores this as well.

Environmental models (including groundwater models) do not have magical predictive powers. They may be complicated, and they may be expensive. However they are not crystal balls. Nor are they necessarily good simulators of subsurface processes. Fortunately, they do not need to be.

The accuracy of any prediction of future environmental behaviour is set by the information content of available data. Decision-support modelling implements location-specific, environmental data-processing. History-matching is essential to this. It allows a model to turn data into information, and transmit this information to a prediction. This allows the uncertainties of decision-critical predictions to be quantified and reduced. This is all that decision-support modelling can achieve. Furthermore, this is what it MUST achieve. Decisions can then be made with full knowledge of the risks that are associated with different courses of management action.

A model that serves the imperatives of decision support does not always need to be “realistic” nor complex in order to be useful. For a start, it must be readily useable with software such as PEST and PEST++. This is because history-matching is essential to data assimilation, and data assimilation is essential to uncertainty reduction. Secondly, a decision-support model must represent that which is only vaguely known about the subsurface in ways that are adjustable rather than fixed. This is fundamental to exploration and reduction of uncertainty. It often requires prediction-specific abstraction rather than picture-perfect digital representation of the subsurface.

All of these issues are discussed in the course. We also discuss the mathematics and practicalities of PEST, PEST++ and similar model-value-adding software. These include:

  • linear algebra
  • inversion
  • model parameterisation using pilot points
  • accommodation of nonuniqueness through regularisation
  • accommodation of nonuniqueness through Bayesian methods
  • direct predictive hypothesis-testing

We will also have time to discuss how the principles of decision-support modelling can be used to address specific issues that are important to you.