check lithologies and try shallow data

Today, I check lithologies.

Summary:

There are ten different lithologies contained in total, from 2 to 11. After trial and error, the best is to use lithologies from 3 to 11. I also deleted the last 15 depths for better data.

The results are as follows. The first is 4 conductivity prediction. Four corresponding R2 of testing data are 0.76, 0.77, 0.69, 0.77. They are a little better than yesterday's results.

The second is 4 permittivity prediction. Four corresponding R2 of testing data are 0.05, 0.28, 0.35, 0.61. They are a little better than yesterday's results.

In conclusion, the lithologies have some kind of effect on prediction performance but not much. For now, lithologies from 3 to 11 together can predict dielectric data the best.

In addition, I also checked shallow data (what we use before are deep data). But the result shows that they are worse than deep data. I checked the all of them in IP and found that shallow data are even more in disorder than deep data. That may be the reason why shallow data have worse prediction performance.

In conclusion, we should still use deep data.

 Next week, I will check the following things which could be useful for improving prediction performance:
1. change the training function and change the model;
2. use predicted con with 11 inputs to predict per together;
3. produce magnitude, tan and predict them;
4. predict with NMR.

Follow the instructions

Today, I follow the instructions as the meeting.

Summary:

First, I predict 4 real con and 4 imaginary con together with 19 inputs. The results of 4 imaginary con are still not good, which are converted from real permittivity.
The reason may be that the conversion is just multiplying some parameters. So there will be little help to improve the prediction.
The first figure is 4 real con prediction performance. The R2 of the testing data are 0.76, 0.83, 0.63, 0.77, which are as good as before.

 The second figure is 4 imaginary con prediction performance. The R2 of the testing data are 0.14, 0.17, 0.29, 0.54, which are as bad as before.

In conclusion, the conversion from real permittivity to imaginary conductivity may be little help for improving the prediction performance.

Second is to predict them with just 11 inputs.
The first figure is 4 real con prediction performance. The R2 of the testing data are 0.75, 0.74, 0.73, 0.74.

  The second figure is 4 imaginary con prediction performance. The R2 of the testing data are 0.05, 0.16, 0.42, 0.52.

In conclusion, the deletion of 8 inputs make the prediction of 4 real con more stable and increase the prediction of 4 imaginary con a little. Generally, 11 inputs are similar to 19 inputs in prediction performance. Actually, more inputs will not decrease the prediction performance. However, we can use just 11 inputs to predict 8 con since 19 inputs do not perform much better.

Third, I check different lithologies step by step. For now, there are 10 different lithologies before division, category from 2 to 11.

1. I check the performance by selecting category from 4 to 10.
The first  is 4 real con prediction performance. The R2 of the testing data are 0.68, 0.58, 0.62, 0.66.
The second  is 4 imaginary con prediction performance. The R2 of the testing data are 0.13, 0.15, 0.32, 0.39.
So the performance is worse.

Tomorrow, I will continue to check lithologies. After that I will continue the instructions with magnitude and tan and NMR. Also, the change of training function may be a choice.

predict Sw

Today, I continue to read some papers and start to predict Sw.

Summary:

Some properties which may affect dielectric logging data:
the content of water, oil and rocks;
different rock minerals;
different lithologies;
mud cake;
water salinity;
temperature, so maybe depth;
in low frequency, conductivity dominates, in high frequency, dielectric constant dominates

Since we use permittivity (both dielectric constant and conductivity) to calculate Sw with some mixing models as I read from papers, why not predict Sw directly?
I found Sw in dielectric data and I predict it.

The training R2 is 0.75 however the testing R2 is 0.07, which is very bad.
I think I can first try to find methods to predict Sw of dielectric data accurately. Then maybe I can also predict per and con more accurately.

Tomorrow, I will try to improve the performance of prediction for Sw.

Finish reading about basic knowledge of permittivity and conductivity

Today, I read three more papers and know about dielectric data more.

Summary:

The task of quantifying petrophysical parameters is divided into two steps.

1.      Invert for the electromagnetic properties of the formation (permittivity and conductivity) from magnitude and phase of electromagnetic wave recorded at the receiver. These electromagnetic properties are obtained for each frequency of operation to obtain their variation with respect to frequency (dispersion).

2.      Valid mixing models are required to relate petrophysical parameters of formation and dispersive electromagnetic properties.

It is shown that the presence of noise in the data can lead to ill-posed inverse problems where multiple answers can be present.

Monte Carlo method is used to study the effect of the noise on the inverted petrophysical properties of the formation.

With the increase of salinity in water, the permittivity of water decreases a little while the conductivity of water increases much. In addition, water salinity will not affect the properties of rocks much.

There are many methods developed for dielectric data. Some of them are for inverting for permittivity and conductivity with different frequencies in order to obtain petrophysical properties. Some of them are for defining quality control criteria for dielectric logging.

Tomorrow, I will start to look for methods to improve the prediction performance of dielectric data. Also, I will try to find papers about how to predict dielectric data accurately.

there is not a single relationship between permittivity and frequency

Today, I read two more papers and find something new.

Summary:

Tomorrow, I will try to finish reading and start coding. However, I really want to spend a little more time on reading so that I can know more about dielectric dispersion.

Basic knowledge of permittivity

Today, I read some papers to better understand permittivity and conductivity.

Summary:

On Tuesday and Wednesday, I met some friends in SPWLA and learn a lot from them. One of them also does research on dielectric data and he sent me some papers for me to better understand dielectric data.

Next week, I will continue to learn about the basic knowledge of dielectric data and try to find solutions for predicting them accurately.

delete some depths but get similar results

Today, I delete some depths but get similar results.

Summary:

Today, I change my model and predict 8 dielectric data together.
I delete some depths according to lithology step by step. But the results are not improved.

I plan to attend SPWLA tomorrow and come back with Hao together.

do some trials

Today, I did some trials.

Summary:

I review all my codes and delete more outliers for outputs, but there is little improvement.

I compared the prediction results of predicting conductivity and permittivity separately and together. For the individual one, they are almost the same, so for now, it shows that whether to predict them separately or together will not affect the results' accuracy so much.

Also, because of the results shown yesterday, I recommend to choose depths as training and testing data randomly.

After discussing with Yifu, he tells me that f0=22MHz, f1=100MHz, f2=350MHz, f3=960MHz, which are from Hess.

Some trials may be done next week:
1. add 64 bin T2 distribution as inputs
2. see if delete some depths will improve the prediction result

change and compare the model

Today, I first delete 4 inputs and change the model. Second, I compare the random and non-random results of prediction.

Summary:

All work is shown in PPT.

Tomorrow, I will try to add NMR T2 data into inputs and find ways to improve the model.

add more inputs to get a better performance of the model

Today, I add more inputs to predict dielectric data. The performance is a little better than before.

Summary:

I add QElan data and water saturation data because I think they are related to permittivity. Also, I delete some outliers so that all data are more concentrated. But there is still a gap of accuracy between conductivity and permittivity data.
Although there is an improvement in prediction performance of the model, it is not much compared with previous', just about 2-3 percent.

The following are the prediction results of 4 conductivity data.

The following are the prediction results of 4 permittivity data.

Tomorrow, I will continue to think if I can improve the prediction accuracy of dielectric data, especially permittivity data.

separate conductivity and permittivity

Today, I tried to improve the performance of the model.

Summary:

Since conductivity and permittivity is different. I build two similar models for them separately using the same inputs.

The following is the predicting result of conductivity. It shows an improvement now for conductivity.

The following is the predicting result of permittivity. It shows bad performance though.

I also discuss with Yifu and search online for the basic knowledge of conductivity and permittivity. Tomorrow, I will try to find more logging data which are more related to them. It may be helpful for improving performance of the model.

predict dielectric data

Today, I predicted 8 dielectric data.

Summary:

8 dielectric data are predicted at the same time.
 After data-preprocessing, the predicted results are as follows.

The first figure is the total performance.

The second figure is the separated comparisons.

The total R2 is about 0.8, which is not very good. The separated comparisons can also be improved as I think.

Tomorrow, I will try to improve the performance of the ANN model. Also, I hope you can have time so that we can have an individual meeting tomorrow.

prepare for predicting dielectric data

Today, I prepare the PPT for the meeting and select data for predicting dielectric data.

Summary:

The inputs I select are as follows.
GR
DPHZ
NPOR
PEFZ
RHOZ
VCL_HILT
AT10
AT90
DTCO
DTSM
VPVS

Since Yifu is not in the office. I cannot make sure which data to be predicted.

Next week, I will discuss with Yifu and decide which data to predict. Also, I will make a list of all logging data available in abernathy well at NMR T2 distribution's depth range.

change the model for abernathy data

Today, I change the model for abernathy data.

Summary:

I use the following data as inputs to change my model.
GR
DCAL
DTCO
DTSM
DPHZ
NPOR
PEFZ
RHOZ
VCL_HILT
UOIL

The following is the comparison between original T2 distribution (blue) and predicted T2 distribution (red) of the changed model.

In the changed model, R2 of the training data is 0.6166 and NRMSE of the training data is 0.1983. The following is the R2 and NRMSE distribution of the training data.

In the changed model, R2 of the testing data is 0.5526 and NRMSE of the testing data is 0.2059. The following is the R2 and NRMSE distribution of the testing data.

In conclusion, there are not enough inputs in abernathy data, so it is not very accurate to use this ANN model to predict NMR T2 distribution.

Tomorrow, I want to read books and papers about algorithms.

import BHP abernathy data and check well 2 data again

Today, I talked with Pratiksha and I import BHP abernathy data and check well 2 data again.

Summary:

After importing BHP abernathy data again, I can find more data now.
The following data are those that can be used.
GR
DCAL
DTCO
DTSM
DPHZ
NPOR
PEFZ
RHOZ
VCL_HILT
UOIL(it is the only one abernathy has, which is related to QElan data)

Now I am sure that these data cannot be obtained in abernathy.
VPVS
AT10
AT90
All QElan data

Also, I do the  same thing again for well 2 in Hess data and make the same conclusion as two days ago, which is that the QElan data and useful T2 distribution data do not overlap.

Today, I also find some papers about Levernberg-Marquardt and conjugate gradient algorithms. I start to read about them. I want to know why CG is more suitable for large number of neurons in ANN models than LM from basic mathematics. It is difficult to understand all the mathematics of them, but I want to have a try.

Tomorrow, I will apply BHP data into my ANN model and continue to read papers about algorithms.

deal with Abernathy data

Today, I dealt with Abernathy data.

Summary:

I tried to look for every file that may be related to Abernathy data, and import all of them into IP. But there are so many data which cannot be used for inputs. The results I conclude are shown as follows.

There are many inputs absent in Abernathy data:
VPVS
AT10
AT90
VCL_HILT

All QElan data


In addition, some inputs in Abernathy data have some problems.
DCAL: all data are equal to 0.
DTCO: the depth range of it is totally different from the others such as GR. So it is difficult to include it into inputs.
DTSM: the depth range of it is totally different from the others.

I checked the email and found that Pratiksha is also dealing with BHP data. Tomorrow, I plan to discuss with her and check if there is something missing in my files or if there are some technique problems to prevent me from obtaining better inputs.

apply ANN model into well 2 data

Today, I applied the ANN model into well 2 data.

Summary:

In well 2, not all data have the same range. The following data 's depth range is from 10316.5 ft to 10850 ft. They are all QElan data.

DTCO_REC_QE

NPHU_EC_REC_QE

ILLITE_QE

CHLORITE_QE

BOUND WATER_QE QUARTZ_QE K-FELDSPAR_QE CALCITE_QE DOLOMITE_QE ANHYDRITE_QE UWATER_QE UOIL_QE

However, the useful T2 distribution 's depth range is from 8997 ft to 9109 ft. All other T2 data are the same for the same bin. So all QElan data in well 2 can not be applied into predicting T2 distribution. As a result, all we have for the inputs in well 2 are shown as follows.

GR_EDTC

DCAL AT10 AT90 DPHZ PEFZ RHOZ

There are just 7 inputs in well 2. I have build an model with these inputs before and they perform not well.

In addition, the depth range of my model is from 10687 ft to 10985.5 ft. So the depth range of my model and well 2 data are not overlapped. In conclusion, well 2 may not be able to be applied into my model if I do not obtain additional data.

Tomorrow, I want to try BHP data.

change the model

Today, I change the model.

Summary:

Since there is difference between inputs of 2 wells. I should change my model in well 1 so that it can be applied to well 2 without training. After changing the model, I use data in well 1 to test the accuracy of model. The result is not bad and is almost as good as the former one.

The following is the comparison between original T2 distribution (blue) and predicted T2 distribution (red) of the changed model.

In the changed model, R2 of the training data is 0.7213 and NRMSE of the training data is 0.1610. The following is the R2 and NRMSE distribution of the training data.

In the changed model, R2 of the testing data is 0.6854 and NRMSE of the testing data is 0.1625. The following is the R2 and NRMSE distribution of the testing data.

Next week, I will apply this model into Hess well 2.

deal with the data

Today, I dealt with the data in well 2.

Summary:

I find most of the data except from the following ones:

VPVS (maybe no need)

DTCO (QElan reconstructed curve)

DTSM (did not find)

NPOR (find similar)

VCL_HILT (illite chlorite mon)

For VPVS, I did not find it. Maybe there is no need to use it because it is the ratio of DTSM to DTCO.

For DTCO, I find QElan reconstructed curve instead of original one.

For DTSM, I did not find it. But since DTCO and DTSM have a strong relationship, maybe there is no need to use both of them.

For NPOR, I find similar one NPHU, and it is also QElan reconstructed curve.

For VCL_HILT, I did not find it. But Yifu said it is the sum of ILLITE, CHLORITE and MONTMORILLONITE.

Tomorrow, I will verify them. If there will be difference between two wells' inputs, I will first change the model in well 1 and apply it in well 2.