As experience in a particular area is gained, a spreadsheet may be used to “adjust" the exponent to match local rock behavior : Figure 2

The Moveable Hydrocarbon Index is slightly greater than the carbonate 0.60 limit and the Bulk Volume Water is considerably larger than common carbonate values ( Figure 6 ) for either vuggy or intergranular / intercrystalline pore systems. Note that were BVW to be calculated with Sw(RR), it would be yet larger.

This is an interview with  Dr. Hansruedi Frueh, the founder of the company

Neuronics AG (www.neuronics.ch) in Zurich and  instructor for the course:

“ The Logic of Neural Networks for the Petrophysical, Seismic and Facies Estimation “ 

at GeoNeurale in Munich.

 

1)    GeoNeurale –

           What is the definition of "Artificial Intelligence" and which
           are its most direct technological applications ?


Hansruedi Frueh - 

Artificial Intelligence (AI) has many definitions, depending on the
 viewpoint. There is a consense, however, that solutions which include
learning, reasoning based on optimisation algorithms, data mining,
language understanding and deduction under uncertainty are located in
the domain of AI. Soft computing technologies, such as neural networks,
evolutionary algorithms, and fuzzy logic belong to it as well as
Bayesian networks (based on probability models) and hard computing
learing principles (e.g. reinforcement learning, support vector
machines, hidden Markov models etc.).
Technical applications which include AI are in use everyday everywhere.
Whether it is the coffee machine, the car, or the search engine in the
internet. And of course in robotics, a field which I work on since more
than a decade.
In the deep subsurface exploration, it also will cover an interesting potential
for the interpretation and spatial distribution analysis of petrophysical and
seismic attributes.


2)   GeoNeurale –


     To what extent do your robots take advantage of Neural Networks ?

Hansruedi Frueh - 

We are using neural networks preferrably when sensor values from
different modalities have to be integrated and related to a desired
behavior. Such a "learning machine" which performs sensor data fusion
and adaptive categorization of the input is exactly what is needed for
many applications, especially in mobile robotics. But we also use
Kohonen networks (a non supervised type of neural networks) to improve
the precision of the robot arm when it has to pick up objects from a
grid. In that case, not every point of the grid is teached in
separately, but only the corners. After an approximative calculation of
all points based on direct and inversed kinematics, there remain small
errors which have different causes (gravity, mechanical deviations,
unperfect models etc.). The integrated Kohonen network allows to learn
how to correct all remaining points from a few examples.



3)  GeoNeurale –

You are a pioneer in the field of applied Artificial Intelligence in Switzerland,
for applications in Robotics and Mechatronics.
The robot measures features of the real world  through its sensors. And performs
a topological representation based on the measurements.
In this worklflow of processing and interpretation, Neural Networks
play a very important role.

The Geoscientist measures the subsurface parameters through indirect and direct
methods, trying to represent the reality of an external world, the subsurface,
which can be measured only with spatial discontinuity and physical approximation.
3D Seismics and Petrophysics are still difficult to be correlated, petrophysical
properties are difficult to be realistically upscaled.
The Geoscientist is therefore continuously confronted with interpolation
issues at the search of the right analyitical methods for the topological laws
to be applied on the property space and on the geological / structural space.
Do you see a parallelism between the Robot´s Logic and the Geosciences
in their measurements and interpretation workflow ?


Hansruedi Frueh - 

I see the parallelism in the sensorics which is important for both
fields (although quite different sensors) as well as in the uncertainty
under which many decisions have to be made. Both need intelligent
algorithms in order to optimize behavior of the mobile robot or
learnings, respectively.


4) GeoNeurale –

To what extent do you think  Neural Networks applications can
help in the property estimation and classification in the Geosciences
Applications referred to the Subsurface Analysis ?
Could you give some examples ?


Hansruedi Frueh –

Neural networks can help to learn nonlinear data efficiently, if the
appropriate network type as well as a fine-tuned learning procedure is
applied. As mentioned above, neural networks are a good choice for tasks
which include sensor data fusion and categorization. Also, neural
networks can reduce the dimensions of data and provide models which can
be used, e.g., for pattern completion.

A typical example is parameter estimation for controllers or for errors
(as in the precision improvement example mentioned above).
Classifications are used in nearly all fields. Neural networks are often
combined with other methods, e.g. fuzzy logic or adaptive probability
(Bayesian) networks which can learn in terms of structures as well as by
modification of the events' probablities.


5) GeoNeurale –

We think now about specific petrophysical properties and log analyses.
A deterministic approach could  involve the application of the Archie´s
Equation or a modified version of it. If this doesn´t seem to be fully
applicable (as the measured logs show a deterministic incongruence but
at the same time we observe a statistical repeating pattern),
what are the main (AI) applications that we could use to perform such a
log analysis ?
 
Hansruedi Frueh –

In my view, questions as you put them are a matter of an
interdisciplinary approach. Neural networks can serve as "learning
machines" to relate input vectors to the most appropriate category which
may be a member of a model created by the network itself or by another
methodology. Alternatively, the correct approach may be a combination of
two different neural network types (e.g., Backpropagation and Kohonen).
It is for sure a very interesting topic, and the potential gain from a
solution based on Artificial Intelligence seems to be high.