Model Calibration and Quantification of Predictive Uncertainty Using PEST


Dr. John Doherty (author of PEST)

Dr. Philippe Renard, Dr. Philip Brunner

October 7-11, 2013 - Neuchâtel, Switzerland

Centre of Hydrogeology and Geothermics, University of Neuchâtel

About PEST

PEST (Parameter ESTimation) is a general-purpose, model-independent, parameter estimation and model predictive uncertainty analysis package developed by Dr. John Doherty. PEST is the most advanced software readily available for calibration and predictive uncertainty analysis of groundwater, reservoir, surface water, and other models. Using PEST you can:

  • apply advanced and efficient regularization techniques in calibrating your models – thereby extracting maximum information content from your data on the one hand and expert knowledge on the other hand
  • undertake linear and nonlinear predictive error and uncertainty analysis of model outputs
  • simultaneously parameterize and calibrate several models using multiple datasets
  • accommodate heterogeneity using advanced spatial parameterization techniques
  • combine PEST with stochastic field generation to efficiently explore pre- and post-calibration parameter and predictive uncertainty
  • conduct parallel model optimization runs across PC or UNIX networks
  • compare the worth of different proposed data acquisition strategies in reducing model predictive uncertainty, thereby optimizing resources allocated to such tasks
  • quantify the contributions to model predictive uncertainty made by different parameter types
  • establish the irreducible uncertainty of a model prior to calibrating that model
  • quantify the reduction in predictive uncertainty accrued through model calibration
  • use a model for quantitative hypothesis-testing
  • provide risk analysis input to the decision-support process

Traditional methods of parameter estimation and uncertainty analysis based on a handful of zones of assumed piecewise parameter constancy do not provide the flexibility needed to extract vital information from expensive data, or to quantify the benefits and risks associated with different environmental management options.

 Some of the advances made over the last few years in PEST to overcome these problems include:

  • Combining regularized inversion with the use of pilot points as a spatial parameterization device
  • Use of the unique and extremely efficient “SVD-assist” regularized inversion methodology which combines subspace and Tikhonov methods. Inversion can be carried out with model-run efficiencies comparable with traditional approaches – even though the model may use hundreds of parameters
  • Combining regularized inversion with stochastic field generation to produce many different calibration-constrained parameter fields with high levels of model run efficiency. All of these can then be used in making any model prediction in order to explore the uncertainty of that prediction
  • Use of Pareto methods to compare the merits of different environmental management strategies and implement model-based hypothesis testing
  • Deployment of linear methods to rapidly assess the relative worth of different strategies for future data acquisition

PEST is accompanied by a large number of utility programs which support its use with commonly used models (such as MODFLOW, FEFLOW, MODFLOW-USG, MT3D, SEAWAT, TOUGH and others). This software automates PEST file set-up for complex parameter estimation and uncertainty analysis problems involving one or multiple models.

Other members of the PEST suite implement the global SCE and CMA optimisation algorithms which can be employed where high levels of model nonlinearity and/or poor model numerical behaviour make use of gradient based inversion methods difficult.

This page has been updated the 6/1/2013