Project Title:

Abstract
Synthetic Open Hole logs derived from Neural Network Model and Three-Detector Pulsed Neutron Tool.
The usage of combined pulsed neutron cased hole log and neural network model to derive a suite of synthetic open hole logs is shown as a valid alternative in formation evaluation when borehole stabil ity and conditions prevented normal open hole logging. This technique uses data from three-detector
pulsed neutron logging for feeding a neural network in which curve emulation models for resistivity, neutron, and density are generated. The three-detector pulsed neutron tool provides greater investigation depth into the formation and lower sensitivity to changes in the borehole condition than conventional A Case Study from Ecuadorian Oriente Basin
two-detector instrument and providing higher accuracy of emulated open hole data. Emulated open hole logs are used effectively to evaluate formations, estimate reserves and put wells into production in an efficient way, reducing risks and costs.
Neural network models imitate the function of biological neurons in the human brain; much like the human brain would solve the same problem. Like people, a neural network learns from examples, which is why it must be trained. For this purpose, at least a training well with both open-hole and pulsed SPE-WVS-703
neutron data is necessary. When these logs are associated and matched successfully with each other, the model can be applied to a nearby application well where only pulsed neutron log data is recorded.
The examples shown in this paper demonstrates the usefulness of the innovative method to emulate triple-combo logs; several wells from Ecuadorian Oriente basin are used in this document to exhibit applicability of this novel technology and the consistent results. The success of this technique is based on
pre-job planning required to characterize formation complexity and wellbore environment (i.e. caliper, completion goods, cement quality) for both the training well and the application well.
Project Investigator:
Sam Jorge Felix, Elizabeth Vicente

Introduction
Formation evaluation through electrical logs is a fundamental phase in determining well production potential, many parameters like shal e volume, lithology, porosity and fluids saturation are crucial to characterize the reservoirs and predict primary production and further hydrocarbon recoveries; however, in
certain areas or circumstances, basic open hole well logs are not always recorded due to tough hole conditions, operational constrain or unplanned events, circumventing necessary information to provide basic reservoir evaluation. By the other hand, newer and innovative cased hole logging tools and Figure 2
programs are being developed to fill this gap but faces huge challenges since processed results might have varying endings du e to errors introduced by reduced dynamic range of the measurements or by using conventional processing techniques bas ed on regular polynomial equations. The approach of
these algorithms is not always reliable as the uncertainties in calculations can be very large under certain conditions and not reflecting the complex characteristics of rock parameters and their fluids. Under this scenario, a methodology is presented that can resolve the problem mentioned earlier by
generating emulated basic open hole logs, these curves are obtained by processing three-detectors pulsed neutron curves recorded in cased hole and a neural network model.

Theory and Definitions

3D-Pulsed neutron downhole tool
The technology used for this case study is a pulsed neutron tool of three detectors. The principle of measurement is similar to a conventional two detectors tool where 14 MeV fast neutrons are generated from a periodic pulsed electronic source interacting with borehole and formation and are slowed down Figure 5
after a series of collisions with atoms of this environment; during the collisions the neutrons losses energy transferring it to the atoms and in turn these emits gamma rays of varying energy levels depending on their atomic number and elapsed time since generation. The first tens of mic roseconds after
neutrons generation high energy inelastic collisions take place; this period is important in detecting elements concentration like Carbon, Oxygen, Silicon, Calcium among others. After this period, to 1000 microseconds or more, neutrons are being slowed down to much lower energy (thermal neutrons) and Figure 4
are captured by surrounding atoms, emitting characteristic gamma rays depending on the element involved. The capture rate is mainly contributed by Hydrogen and Chlorine. Finally, unstable atoms generated by the neutrons collisions might take several seconds, minutes or more to go back to normal
stage. The scintillation detectors of pulsed neutron tools senses these gamma rays as an energy spectrum from which information are gathered. Depending on the acquisition mode, inelastic or capture, measurements include curves like formation macroscopic capture section Σ (sigma), Carbon to Oxygen
ratio C/O, inelastic to capture ratio, Oxygen activation, etc. However, a three detector tool, with an extra longer spacing away from the source, has a more precise measurement of the reservoir since the investigated volume is larger and beyond from the borehole.

Figure 1 Figure 3
Neural network model
Artificial neural networks are mathematic models that replicate the central nervous system functions. Several neural network models have been studied since decades to emulate biological functions. More recently, there are many applications in electrical engineering, computer science, robotics, medicine,
economy and petroleum engineering. Neural networks models are prone to be superior to other methods under following circumstances:
 Self-organizing and adaptability: utilizes adaptive learning algorithm and self-organization, offering better chance of robust process.
 Naturally inclined to acquire knowledge through experience, which are stored, as in human brain, in neuron’s relative weights of the interconnections.
 Non-linear processing: increases network capacity to approximate functions, discriminate patterns and greater immunity to noise.
 Parallel processing: usually uses a large number of processing nodes with high level of interconnectivity.
A neural network model is usually organized under a sequence of layers. The amount of layers is related to the parts a partic ular problem is wanted to be splitted, the amount of neurons or nodes lo be used at input or output level are also dependent on the requirements of the problem. In neural networks
with more than two layers, the input nodes are connected to intermediate layers and these in turn are connected to the output nodes. Intermediate layers are usually referred to as “hidden layers” since are not visible for input and output nodes; these “hidden layers” can also be called “main learning
detectors” since the activity pattern built in is a coding of what the components are considered key inputs; the algorithm used is also executed here. Additi onally, a network of connecting weight factors exists between layers of neurons indicating the relative input’s prominence over the linked output with
numerical values. Figure 1 shows a general artificial neural network diagram.

Methodology
Training well
In order to apply open hole curves emulation technique a “training well” is needed initially since curves learning is fundame ntal in the plasticity of the neural network model and is essentially the process in which each weighting factor or value of a neuron, as well as its correspondent threshold value, are
attuned to a function of learning algorithm with the purpose of approximate the generated information from the neural network to the expected results as close as possible.
The analyst should have at least the following information at the training phase:
 Open hole and Cased hole data of the training well; required open hole data is composed of formation density, neutron porosity, formation deep resistivity and gamma ray; cased hole data is acquired with three-detector pulsed neutron tool under capture mode.
 Caliper, this information is helpful to assess washout effects on pulsed neutron readings.
 Cement bond log to assess validity of pulsed neutron readings due to poor bonding or voids behind casing.
 Completion mechanical diagram with tally information and formation tops since the model is trained with certain well and reservoir characteristics.
The proposed neural network model to emulate open hole logs is a supervised learning network type, based on back propagation algorithm. The model considers 18 recorded cased hole curves as input, 36 hidden nodes as processing neurons and one output node with the desired emulated open hole
curve. Figure 2 shows the proposed neural network architecture. The analyst, however, can discriminate any input node if it is inappropriate for training or add other inputs if it reinforces the neural network.
All cased hole input data curves embrace characteristics of reservoir rocks and fluids. Table 1 shows the measurements and application of several input curves. These curves, which function as stimulus, are propagated forward with a weighting factor to generate an output. Figure 7
Output signal is compared to respective desired open hole curve and an error signal is calculated; error signal is back propagated to update weighting factors. This process works back and forth iteratively to get a minimum error signal between output signal and desired output for neutron porosity, formation
density and formation deep resistivity curves.

Figure 9
Application well
Once the neural network model is trained, this can be used in nearby wells.
Application well would have pulsed neutron data normalized to training well in consideration to environmental and calibration differences. Conversely, training well and application well should have the same borehole conditions, like bit size, tubular and drilling mud type.

Application examples
Application cases are presented here in oilfield M and N from Ecuadorian Oriente basin, where main reservoirs are from Cretaceous sandstones “Hollín” from Hollín Formation, “U” and “T” from Napo Formation, and marginally, “Basal Tena” from Tena Formation. Lithology are composed of sand-shale
sequences, sands with calcareous cementing agent, some with glauconite, porosities ranges from 10% to 18%. Due to reservoir depletion and complex borehole geometry, open hole logs are sometimes cancelled; emulated open hole logs using neural network model and three-detector pulsed neutron tool
has filled this gap appropriately.

Training well, oilfield M

Wells A and B were used as training wells to feed a neural network model to replicate application well C. Both wells A and B have the same stratigraphic column as shown in their correlation (Figure 3). Formation thickness has similar footage and development, they are located at the same area of the Figure 6
reservoir and most formations have comparable rock and fluid properties.
Figure 8
 Interval used as training section from well A in our neural network model is from 8760’ to 9980’ MD; a composite log with input and emulated curves from training well A is presented in Figure 5.
 Emulated resistivity curve from well A exhibit good results and features that resembled lithologic changes in the emulated interval. However, a small spike was noted in black circle showing lower resistivity readings as compared to open hole recorded data; this zone could be affected by poor cement
bonding as seen in cement map in Figure 4. Nevertheless, a crossplot between emulated resistivity and open hole resistivity, good correlation is evident with a R2 coefficient of 0.965 (Figure 5). Table 1
 Emulated neutron porosity curve also shows good results with a correlation coefficient of R2 = 0.965 with open hole curve for the entire trained section of 1220’ (Figure 5). In fact, neutron pororsity curve emulation shows that this curve is less sensitive to washout and pooor cement bonding effects.
 Formation density curve emulation shows good results in general, but in Figure 4, several red circled zones are highlighted where curve matching is less than desirable, this usually happens in washout sections as verifi ed with Caliper curve; fortunately, washout areas are in front of shaly sections
which are not reservoir sands. A crossplot was done to verify similarity between emulated and open hole recorded density curves, presenting a correlation coefficient of R2 = 0.795 for the whole emulated interval (Figure 6), which is a good value considering washout zones are present.
 Next training well B was used as input to our neural network model, covering from 9300’ to 10512’; a total footage of 1212’ w ere emulated in well B as illustrated in a composite log showing input and output curves for this training well.
 Emulated resistivity and neutron curves showed good matching with their respective open hole curves as seen on the crossplots in Figure 7; resulting correlation coefficient is 0.936 for the resistivity curve and 0.916 for the neutron curve. Emulated density curve is also affected by washout but in non-
reservoir zones, so formation evaluation can be adequately done; correlation coefficient in this case is R2 = 0.795 and is considered a good value by taking into account signs of washout zones.

Conclusions
Application well, oilfield M
 The presented technique proved to be useful as an alternative to provide basic formation evaluation curves when usual basic open hole logs cannot be acquired due to borehole stability, reservoir depletion or operational problems.
Well C was selected as application well since an incomplete operation during open hole logging arose: recorded resistivity, density and neutron in open hole from 9584’ to 10440’ only, and left unrecorded the interval from 10440’ to 11052’.  In order to have adequate training well for neural network learning we must consider formation lateral heterogeneities, well location, formation tops and mechanical configurations son emulated curves are generated under same
 The neural network trained previously with wells A and B was required to emulate resistivity, density and neutron curves, nec essary for a basic reservoir formation evaluation from 9670’ to 11052’. considerations.
 All emulated curves not only presented good response to existing lithology but also a good match with overlapping open hole s ection in the upper interval, from 9584’ to 10440’. See Figure 8.  Training and application wells should have the data acquired before first completion/production since reservoir and borehole conditions changes after a period of being produced.
 These emulated curves were used to process a basic petrophysical analysis, identifying two prospecting intervals in “U” sandstone, as secondary target, with So average of 70% and others two intervals in “T” sandstone with So average around 60%, as primary target.  This technique of logging in cased hole can reduce risk and save money in complex wells where open hole logging is compromi sed due to any operational issue, so it acts like an insurance or back-up as an alternative to gather
 From the petrophysical analysis, intervals 10846’-10875’ and 10894’-10907’ in “T” sandstone were completed; a stabilized production was recorded of Qf: 890 BFPD, Qo: 516 BOPD, Qw: 374 BWPD and 42% BSW (Figure 9) information when other sources are either too risky or cost-restrictive.
 Since the emulated curves are comparable to near-by open hole basic data, quality control and formation evaluation will be straight-forward, avoiding further complex techniques, processing programs and unusual tools in the area
with no previous knowledge of responses, limiting the confidence on the results.