HydroForum - Groundwater, Water Resources, Hydrology - "Inverse Modeling: Is it too good to be true?"

HydroForum - Groundwater, Water Resources, Hydrology

"Inverse Modeling: Is it too good to be true?"

Email this topic to a friend
Printer-friendly version of this topic
Previous Topic | Next Topic

Home Conferences Groundwater, Surface Water, Watershed and Geochemical Modeling (Protected)

Original message

jose_vaughan Mar-10-00, 10:18 AM (GMT)

"Inverse Modeling: Is it too good to be true?"

I have received concerns from some of the list members regarding my recent posting on calibrating transient groundwater models with automatic calibration tools. They pointed out that automatic calibration problems, involving transient groundwater models are normally ill posed, as such unrealistic model parameters < exceptionally high specific yields, excessive parameter variation in adjoining zones, negative hydraulic conductivity, etc.> are often computed with the automatic calibration tools. The failure to compute reasonable model parameters by these tools results from one or more of the following problems:
1. Small random errors in observed/measured data result in larger errors in the computed derivatives, and these errors cause excessive spatial variation in the computed parameters. Sometimes, the computed values are even sensitive to the precision of the derivatives.
2. Subjective input data is used to simulate model output. And, this output is compared with the observed data by the automatic calibration tools in order to compute model parameters.
3. Conceptualization, numerical, truncation and modeling errors are always introduced during the development of a groundwater model.
4. There are many different ways available to formulate objective function (calibration criterion) to represent the error of calibration and the computed calibration parameters are largely influenced by how the objective function is formed. As such, model parameters which are computed by these tools, are based on the spurious data rather than by the useful data.
5. The non-linear behaviour on the groundwater flow equation makes it difficult to discriminate between parameters to which the objective function is relatively sensitive, as the sensitivity of the parameter changes with its value.
6. Although this technology is more than 30 years old, no known method of solution has been found to find the global optimum of a non-linear optimization. This is because, the behavior of the solution procedure is the same when approaching a local minimum as when approaching the global minimum.
I have my doubts with the recent comparison of automated calibration with driving a car. This is because, the calibration domain is filled with many different types of errors and parameters, complicated non-linear parameter behaviours, many local minimums and subjective input data. As such, the calibration domain could be compared to a thick forest and driving is not possible in this domain even with eyes wide opened.
Within the surface water modeling community who is more homogeneous group than the groundwater modeling community, the inverse models are not that popular. The inverse models are useful to estimate parameters which control the physical processes in a small scale experiment and for certain extent to analyse the pumping test data. And, my intuition tells me that relying completely on inverse models is bit of a waste of time. Groundwater modeling is multi-disciplinary task and only reliably good model can be developed with group of people who understand maths, physics and hydrogeology.
I would like to know the purpose of MODFLOW-2000? I believe that MODFLOW-2000 is written in such a way so that it runs within an inverse program. But, our experience show that inverse models do not provide satisfactory results for natural systems. And, in my opinion, the use of MODFLOW-2000 will definitely create another group of modelers who don’t have any knowledge on how the natural systems behave and only know how to manipulate the numbers for their answers. There is already a group of modelers campaigning that the groundwater models can be developed in many different ways in order to provide many different outcomes.
BTW, a current modeling dispute has been brought to my attention. In that dispute, the client believes that the consultant’s model neither meets their expectation nor honours the local hydrogeology. On the other hand, the consultant argues that the model has been developed using industry accepted software products, as such their model conforms to the acceptable modeling practices.
José Vaughan
E3 Consultancy
Earth, Environmental and Economics Consultants
we work for sustainable future

Alert Edit | Reply | Reply With Quote | Top

Table of contents

RE: Inverse Modeling: Is it too goo..., dugmason, 03:03 PM, Mar-10-00, (1)

Email this topic to a friend Printer-friendly version of this topic	Previous Topic \| Next Topic
Home Conferences Groundwater, Surface Water, Watershed and Geochemical Modeling (Protected)
Original message

Lobby | Topics | Previous Topic | Next Topic

Messages in this topic

dugmason Mar-10-00, 03:03 PM (GMT)

1. "RE: Inverse Modeling: Is it too good to be true?"

Although I am no expert on inverse modeling, I have been doing a fair bit of research into the subject since I have been evaluating several different groundwater modeling software packages.
It seems to me the problems you have mentioned are a common concern among anyone using parameter estimation to calibrate their model. However, in speaking with one of the developers, it seems that this problem can be easily avoided with PEST2000 (or WinPEST or Visual PEST or whatever flavor of PEST you use) by simply putting upper and lower bounds on the parameter values which you are optimizing. You can also maintain relationships (ratios) between parameters values by linking them.
I have seen the program at work and it is quite impressive. It seems like it will save me alot of time instead of manually adjusting the parameter values and trying to 'guess' which way I need to asjust them to arrive at a calibrated solution.
Naturally, I won't blindly submit my modeling results without thoroghly reviewing the values for each parameter and ensuring the reasonableness of the results. However, if someone does submit results without checking them and justifying them, you cannot blame the programs used to arrive at the results, nor should you assume the technology is not useful.
It's like saying the internet is not a good technology because it is used to propogate pornography. The internet technology is not at fault, the people who use it for this purpose are at fault.
Excuse me while I get down off of my soap box.
Just my two cents.
Douglas Mason
JUST GROUNDWATER, INC.

Remove | Alert Edit | Reply | Reply With Quote | Top

Lock | Archive | Remove

Lobby | Topics | Previous Topic | Next Topic

Powered by DCF2000 ©1997-2000 by DCScripts. All rights reserved.

jose_vaughan	Mar-10-00, 10:18 AM (GMT)
"Inverse Modeling: Is it too good to be true?"
I have received concerns from some of the list members regarding my recent posting on calibrating transient groundwater models with automatic calibration tools. They pointed out that automatic calibration problems, involving transient groundwater models are normally ill posed, as such unrealistic model parameters < exceptionally high specific yields, excessive parameter variation in adjoining zones, negative hydraulic conductivity, etc.> are often computed with the automatic calibration tools. The failure to compute reasonable model parameters by these tools results from one or more of the following problems: 1. Small random errors in observed/measured data result in larger errors in the computed derivatives, and these errors cause excessive spatial variation in the computed parameters. Sometimes, the computed values are even sensitive to the precision of the derivatives. 2. Subjective input data is used to simulate model output. And, this output is compared with the observed data by the automatic calibration tools in order to compute model parameters. 3. Conceptualization, numerical, truncation and modeling errors are always introduced during the development of a groundwater model. 4. There are many different ways available to formulate objective function (calibration criterion) to represent the error of calibration and the computed calibration parameters are largely influenced by how the objective function is formed. As such, model parameters which are computed by these tools, are based on the spurious data rather than by the useful data. 5. The non-linear behaviour on the groundwater flow equation makes it difficult to discriminate between parameters to which the objective function is relatively sensitive, as the sensitivity of the parameter changes with its value. 6. Although this technology is more than 30 years old, no known method of solution has been found to find the global optimum of a non-linear optimization. This is because, the behavior of the solution procedure is the same when approaching a local minimum as when approaching the global minimum. I have my doubts with the recent comparison of automated calibration with driving a car. This is because, the calibration domain is filled with many different types of errors and parameters, complicated non-linear parameter behaviours, many local minimums and subjective input data. As such, the calibration domain could be compared to a thick forest and driving is not possible in this domain even with eyes wide opened. Within the surface water modeling community who is more homogeneous group than the groundwater modeling community, the inverse models are not that popular. The inverse models are useful to estimate parameters which control the physical processes in a small scale experiment and for certain extent to analyse the pumping test data. And, my intuition tells me that relying completely on inverse models is bit of a waste of time. Groundwater modeling is multi-disciplinary task and only reliably good model can be developed with group of people who understand maths, physics and hydrogeology. I would like to know the purpose of MODFLOW-2000? I believe that MODFLOW-2000 is written in such a way so that it runs within an inverse program. But, our experience show that inverse models do not provide satisfactory results for natural systems. And, in my opinion, the use of MODFLOW-2000 will definitely create another group of modelers who don’t have any knowledge on how the natural systems behave and only know how to manipulate the numbers for their answers. There is already a group of modelers campaigning that the groundwater models can be developed in many different ways in order to provide many different outcomes. BTW, a current modeling dispute has been brought to my attention. In that dispute, the client believes that the consultant’s model neither meets their expectation nor honours the local hydrogeology. On the other hand, the consultant argues that the model has been developed using industry accepted software products, as such their model conforms to the acceptable modeling practices. José Vaughan E3 Consultancy Earth, Environmental and Economics Consultants we work for sustainable future
Alert	Edit \| Reply \| Reply With Quote \| Top

dugmason	Mar-10-00, 03:03 PM (GMT)
1. "RE: Inverse Modeling: Is it too good to be true?"
Although I am no expert on inverse modeling, I have been doing a fair bit of research into the subject since I have been evaluating several different groundwater modeling software packages. It seems to me the problems you have mentioned are a common concern among anyone using parameter estimation to calibrate their model. However, in speaking with one of the developers, it seems that this problem can be easily avoided with PEST2000 (or WinPEST or Visual PEST or whatever flavor of PEST you use) by simply putting upper and lower bounds on the parameter values which you are optimizing. You can also maintain relationships (ratios) between parameters values by linking them. I have seen the program at work and it is quite impressive. It seems like it will save me alot of time instead of manually adjusting the parameter values and trying to 'guess' which way I need to asjust them to arrive at a calibrated solution. Naturally, I won't blindly submit my modeling results without thoroghly reviewing the values for each parameter and ensuring the reasonableness of the results. However, if someone does submit results without checking them and justifying them, you cannot blame the programs used to arrive at the results, nor should you assume the technology is not useful. It's like saying the internet is not a good technology because it is used to propogate pornography. The internet technology is not at fault, the people who use it for this purpose are at fault. Excuse me while I get down off of my soap box. Just my two cents. Douglas Mason JUST GROUNDWATER, INC.
Remove \| Alert	Edit \| Reply \| Reply With Quote \| Top