Fundamental Problem Setup

 The fundamental setup prepares the model to be processed by PEST. Its steps are common to all basic and advanced PEST methods. Before you start with the fundamental setup, we recommend you to take a look on the section Is the FEFLOW model ready? to check some preliminary steps.

  • Creating a new FePEST problem
    Open a new project and choose the related FEFLOW model

  • Parameters
    A list of adjustable parameters including their locations (zones of parameter constancy or continuous parameter distributions interpolated from pilot points) defines which parameters are to be estimated.

  • Observations
    Historical data that act as the calibration targets is supplied. This choice will constitute the measurement objective function, whose reduction is one of the primary targets of model calibration.

  • Prior Knowledge
    A calibration solution based on the minimization of the objective function alone is non-unique in almost all cases of environmental modelling. Expert knowledge (Prior Knowledge) of the expected geological situations lowers the degree of non-uniqueness by preferring parameter distributions that are considered to be plausible by geological expertise.

  • Regularization
    Even after the inclusion of rich observation data and constraints of expert knowledge, a common calibration problem can still show significant non-uniqueness. This can introduce numerical instability to the inversion process. Tikhonov regularization, subspace regularization and regularization by Super Parameters (SVD-Assist) are techniques quite useful in these situations. Especially subspace methods and regularization by Super Parameters can separate parameter space into combinations of parameters which are estimable on the basis of the current calibration dataset and combinations of parameters which are inestimable, and hence define the null space. Estimation of only estimable parameter combinations ensures numerical stability of the fitting process. It also allows calculation of some informative statistics related to parameter identifiability and parameter/predictive uncertainty. In addition, if using SVD-Assist, significant reductions of optimization run-time can be achieved as PEST only needs to undertake as many model runs per iteration as the number of uniquely estimable combinations of parameters.

  • Parallel Computing
    Distributing the computational load of the undertaking many model runs to support the inversion process over a number of different machines reduces the overall processing time. Parallelization is a requirement for almost any highly-parameterized model to meet usual project time frames.

Table of Contents