I was based in Cairo during this time with Reservoir Seismic Services division of [WG-SLB](https://www.slb.com/). I worked as Senior Reservoir Geophysicists, Business Development Manager, and Business manager during this time. But mostly job remained technical. This center was responsible to provide its services to Australia, Asia, ME, and Africa. The project variety and their geologies varied so much that it was never boring and less challenging. 

Here I got opportunity to apply Artificial Neural Network Technology extensively and widely. I will discuss few of them below:

# RockCell, Geoframe #
Rockcell used radial basis function (RBF) to improve the accuracy of prediction. One can read [How radial basis functions work](http://desktop.arcgis.com/en/arcmap/latest/extensions/geostatistical-analyst/how-radial-basis-functions-work.htm)  for more information about RBFs.

I published a paper in an Internal Forum of SLB about my work using this technology. The title of the paper was --->
###### Identification of laminated pay through innovative integration of AVO inversion & borehole electrical images  - KG Basin ###### 
The abstract of the paper is given below:

The laminations cannot be resolved through normal wireline logs because the thickness of laminations is of the order of few mm to few cm. Therefore, identifying laminations as a package instead of trying to resolve them through any seismic measurements is worthwhile. To identify laminations and other litho-classes in the well borehole electrical image data and petrophysical data were combined through artificial neural network based approach. The measured elastic properties of all the lithofacies in the borehole are used to derive probability distribution functions (PDF) in a Baysian framework, which successfully predicted all the litho-facies including laminated pay in well domain. The error analysis of inverted seismic attributes with respect to wireline logs suggests that the total error is litho-class dependent and was maximum for gas sand. The PDF’s generated by combining the Backus average data for gas sand and without Backus average data for other litho-classes produced better litho-classification which matches more closely with well.  This work also suggests that to further improve the accuracy of lithofacies prediction, a suitable geological attribute needs to be derived from seismic data.

# Neural Networks in Petrel #
[Petrel](https://www.software.slb.com/products/petrel/petrel-geophysics/classification-and-estimation), a flagship petrotechnical analysis platform of SLB, has very comprehensive set of artificial neural network modules and they are tightly integrated in such a way that same set of modules can be used for single well to multiple well log based analysis as well as to extend them to the geological models which are populated by all kinds of petrophysical properties.

### Well Log editing or Missing Log Generation ###
All projects required that well data quality is checked and improved where there are problem including generating logs wherever they are not available. Before the availability of Neural Network based tools it was often done on adhoc basis which obviously can not be defended during peer reviews as everybody can have a different opinion. Therefore, when ANN technology became available I immediately used it for well log editing. 

Methodology is simple and described below:
1. Create a wellbore quality flag  i.e. if Caliper - BS >Threshold then 0 else 1
2. Select input logs plus the flag and output log. Output log is one that is to be edited or generated
3. Train the neural network
4. Now change the Flag to 1 for entire section/ and or boreholes
5. Predict the log
6. Replace original log with predicted log where it is bad guided by flag in step 1 or if missing then use entire predicted log

I used and promoted its use in all the projects done in RSS, Cairo. It was also submitted as best practice method in InTouch - a knowledge base of Schlumberger.

### Property Modeling ###
Property modeling is last step for geological modeling in any reservoir characterization project. It is here we integrate well log data with seismic data to come up with more accurate property modeling. Various attributes of seismic data e.g. amplitude, frequency, curvature, coherence, relative amplitude, gradient, absolute acoustic impedance/PR/Density etc can be used to guestimate and populate geological model with reservoir properties e.g. Lithoclass,  VCL, Porosity, Saturation, permeability, geomechanical properties etc. 

It is desirable that if seismic derived attributes can be used, calibrated and validated at well locations, then much more accurate and reliable static model can be produced.

Large number of attributes can be derived from surface seismic and there can not be any obvious relationship derived between seismic attributes and reservoir properties. Therefore, the use of neural network becomes essential to exploit the benefits of seismic measurement for property modeling.

A simple approach is as following:
1. Identify logs and attributes to be used as input
2. Identify reservoir properties to be populated in Geological model as output
3. Upscale and extract these input and output logs at well locations
4. Train the neural network
5. Now use trained neural network to predict the reservoir property away from the well
6. QC and validation


I used above methodology, with required variations as needed, in many projects. One important project I did was for a client in Sudan. The work was published in Internal forum in Schlumberger. Its title was -->
 ###### Enhancing Overall Performance for Well Placement through Seismic Inversion ######
Below is its abstract:

The objective of well placement is to land the well at right depth and azimuth and ensure to penetrate specific intervals in the reservoir unit. To successfully achieve this objectives, pre-job model of earth sub-surface is used in conjunction with Geosteering technology (example Periscope)  to capture both vertical & lateral heterogeneities so that the target zone as well as landing point can be defined within +/-5m radius. An error of more than that in pre-job model might lead to unsuccessful well placement.

Conventional approach for building 2-D or 3-D pre-job earth model is based upon krigging of well data controlled by structural interpretation. This approach is suitable for a homogeneous environment but not for heterogeneous environment e.g. fluvial, deltaic, turbidites etc. 

Seismic inversion is an established geophysical technique whereby the elastic rock properties are estimated directly from seismic data at each common mid point - CMP location (generally 12.5m by 12.5m). The inverted elastic properties can be transformed to build 3-D model of reservoir parameters or log responses of the subsurface using suitable empirical transforms or properly trained neural network. 

A Company has drilled few horizontal wells in its field using conventional pre-job modeling approach with varying degree of success. Reservoir Seismic Services, possibly first time in Schlumberger, introduced new concept to Company for pre-job modeling for well placement guided by seismic inversion. To test the concept, Company awarded a pilot project for well placement of well XX. 

The proposed workflow comprises of the following eight elements: 
1)	Well Log Conditioning
2)	Seismic Data Conditioning and Well-Seismic Calibration
3)	Low Frequency Modeling
4)	ISIS Global Seismic Inversion
5)	Velocity Model Building and Time-Depth Conversion
6)	Petro/Log-Acoustic analysis
7)	Hi-Resolution Stochastic Simulation for Acoustic Impedance
8)	Property/Log Response Modeling using Neural Network

![](https://cdn.steemitimages.com/DQmYzs3TDCBHewJBbtD4MiBSK8axsjW8z3j7U6o4b1L5SJb/image.png)

A more comprehensive, seismic and well constrain pre-job 3-D earth model of Gamma Ray response was produced using the GR logs from 4 wells, seismic inversion over 10 km2, seismic velocity, and structural interpretation. The seismically driven 3-D earth model captures much better the heterogeneity and complexity of the target area particularly in the inter-well space than any traditional pre-job earth model. 

The new model further, successfully explained the varying successes and failures of previously drilled wells which had improved the confidence in the new model. Based on the analysis of previous wells against the new model – the planned trajectory of well XX was modified so that the well would penetrate much cleaner sands (more porous) than it would have other-wise. 

The well was drilled and successfully geo-steered using Periscope. The conclusions from post-drill analysis are as following:

Log/Property modeling guided by seismic inversion matches with the drilling results.
1. The NTG of the horizontal section increases to 78% versus 54%.
2. The per day oil production rate also increased.
3. Reduced unnecessary changes in well trajectory during drilling.
4. No missing of target zones in drilled section. 
5. Improved overall well placement performance 

There was a depth uncertainty of <3m between model and actual depth for the top of formation. Although <5m uncertainty in depth is within the tolerance of Periscope still the depth uncertainty could be identified and reduced if time-depth is continuously updated in near real time using well-seismic tie utilizing LWD logs. 

Success of the pilot project had proven the concept and generated interest in wider application of the workflow on future well placement projects.


Keep Steeming.

## Previous Posts on this story ##

[My work in The field of Artificial Intelligence - Episode 1 (1990-1995)](https://steemit.com/ai/@prameshtyagi/my-work-in-the-field-of-artificial-intelligence-episode-1-1990-1995)
[My work in The field of Artificial Intelligence - Episode 2 (1998-2004)](https://steemit.com/ai/@prameshtyagi/my-work-in-the-field-of-artificial-intelligence-episode-2-1998-2004)

### Note ### 
 Any relevant comment will be upvoted.