Methods and guidelines for effective model calibration

CAL3
-2.50
-1.25
0.00
1.25
2.50
-25
0
25
50
Weighted sim ulated value, dim ensionless
We
ighted residua
l, 
dimens
ionless
Heads
Flow s
Prior
CAL0
(Out of range flow : -102, -1.1)
-2.50
-1.25
0.00
1.25
2.50
-25
0
25
50
Weighted sim ulated value, dim ensionless
We
ighted residua
l, 
dimens
ionless
Heads
Flow s

---

55
The third attribute of better models is that optimum parameter values will tend to be more 
reasonable, both in terms of the estimated values and their values relative to one another. Unreal-
istic optimized parameter values often are disconcerting to users, as mentioned in the section "Lack 
of Limits on Estimated Parameter Values" and under Guideline 9. 
In some cases the model evaluation may indicate that the data are insufficient to identify a 
best model from several possible alternatives, in which case any predictions of interest need to be 
simulated using all reasonable models.
Poeter and McKenna (1995) present an innovative method of using indicator kriging to 
generate possible models that differed in the zonation used for the hydraulic-conductivity field. 
They then estimated hydraulic conductivities using MODFLOWP. The synthetic test case used al-
lowed them to show that the additional analysis provided by nonlinear regression tended to produce 
more accurate transport predictions than could be attained without the use of regression. The addi-
tional analysis included determining the best-fit parameters for each model through regression, and 
then omitting models for which at least one of the following conditions occurred: (1) the best-fit 
parameter values were unrealistic in that obviously coarser deposits had lower hydraulic conduc-
tivities than finer grained deposits, (2) the best-fit parameter values were substantially different 
than expected, (3) the model fit was significantly worse than for other models, or (4) the regression 
did not converge. The dramatic improvement in the predictions produced by models screened using 
these criteria indicated that their application is likely to be useful for identifying more accurate 
models.