This webinar features a 45-minute presentation by Mr. Rocky Roden (CV below), an industry thought leader, and Senior Consulting Geophysicist for Geophysical Insights. An interactive Q&A with Mr. Roden will follow his presentation.
Title: What Interpreters Should Know About Machine Learning | Presenter: Rocky Roden | Date: Tuesday, 5 May 2020
Topics and questions to be addressed:
- Why Machine Learning now?
- Address terminology confusion
- Types of Machine Learning
- Case studies
- Machine Learning and the “Black Box” connotation
- Machine Learning and compute power
- Future trends
Our lives are intertwined with applications, services, orders, products, research, and objects that are incorporated, produced, or effected in some way by Artificial Intelligence and Machine Learning. Buzz words like Deep Learning, Big Data, Supervised and Unsupervised Learning are employed routinely to describe Machine Learning, but how do these applications relate to geoscience interpretation and finding oil and gas. More importantly, do these Machine Learning methods produce better results than conventional interpretation approaches? This webinar will initially wade through the vernacular of Machine Learning and Data Science as it relates to the geoscientist. The presentation will review how these methods are employed, along with interpretation case studies of different machine learning applications. An overview of computer power and machine learning will be described. Machine Learning is a disruptive technology that holds great promise, and this webinar is an interpreter’s perspective, not a data scientist. This course will provide an understanding of how Machine Learning for interpretation is being utilized today and provide insights on future directions and trends.
Please mention the specific software doing these jobs.
Examples are from the Paradise AI workbench by Geophysical Insights.
What is the minimum and maximum window optimal for the analysis?
Usually a window from 50-500 ms around an interval of interest is a good range. However, it depends on the size of the geologic feature to be identified. For example, the edge of a salt diaper may require a much larger window. Remember if 64 neurons are used it will classify 64 clusters in a 100 ms window or over a 6 second window. The results will be very different.
Would you agree that the SOM analysis is somehow overcoming the limitations of "low resolution" seismic allowing to look a bit more in detail than what is conventionally possible, or the quality of the output is still very much dependent on the input seismic resolution/quality?
Yes. Good quality seismic going in is always better, but SOM analysis results are on a sample basis. Each sample in SOM classification is a combination of the attributes used, not a limitation of the frequency and wavelength from conventional amplitude data.
Semi-supervised makes use of newly labeled data added to the original labelled data set, but can % confidence of the new labelled data be included?
We cannot speak to other software products, but Paradise generates a probability volume with two different types of machine learning applications in the workbench.
Could you clarify the difference between PCA and SOM?
PCA is a linear algorithm that looks at the variance of different types of data to see which stand out. SOM is a nonlinear neural network approach that classifies numerous types of data into natural patterns and clusters.
Different programming needed for CPU and GPU or can a CPU usual program run on a GPU?
Programming is usually not the issue, but a machine learning approach may need to be transposed from a CPU to a GPU.
Is it possible to run SOM ML/DL in the Cloud? And if so, public and/or private?
We routinely run in the cloud on public cloud services. Therefore, it can run on any cloud – public or private.
Which is the preferred programming language that supports Machine Learning for Geoscientist?
Python is the most popular, but R and a few others are also employed.
What are the main pre-requisites for today interpreter to have so they can get into ML quickly?
Understanding which ML programs solve specific problems.
Do you think that these approaches, machine learning, supervised and or unsupervised can be applied with success to other reservoir data, at different scales such as wellbore images or core images (CT scan and so on)?
Absolutely. Cluster analysis has been used for years in well log curves. The correlation of lithofacies from logs tied to the seismic for areal distribution is promising today.
Slide 46 (Class 3 AVO example): What were the seismic stacks, gathers, velocities, etc. used as input to generate the SOM?
Based only on the stacked amplitude volume.
In what ways can ML be helpful in Seismic data processing, given that each data is kind of unique?
We have seen where ML can identify noise in the data, and then the class corresponding to noise can simply be turned off.
Given that Machine Learning is heavily dependent on data How does ML handle data conditioning?
ML has been applied in various facets of data processing and will be used more in the future.
How do we quantify the uncertainty associated with the Machine learning outputs and is there like a threshold values specific to results?
Depending on the ML approach the uncertainly can be quantified differently. An important issue is even though a statistical uncertainty can be calculated, that does not necessarily mean it has a high geological significance.
What interpreters do in real industry related to machine learning? Should they understand coding/programming? or they just run a software that has "tools"?
Today many geoscientists are involved in coding because there are not many tools available. It certainly helps if a geoscientist knows coding, but I think in the next 5-10 years there will be numerous applications on the market for geoscientists to choose from.
Unsupervised learning giving disruptive insights...how would you explain this to managers who want to see proves?
If managers don’t want to hear the technical story, then showing them examples and situations where it did work can go a long way to helping convince them.
How does machine learning handle multi-resolution (seismic) data?
If you mean spectral decomposition volumes machine learning works fine as input. It is just important to know that specific small frequency bands may be telling you something different than the full frequency volume.
Can you obtain resolution below Shannon's limit and when why? Thanks to instantaneous attributes?
The SOM process in Paradise can identify stratigraphy below conventional seismic tuning due to the analysis of multiple attribute volumes simultaneous at single sample scale – not wavelet based.
How can a neural network is made to do risk amplitude anamoly? How to identify lithology from seismic through these approach?
Within Paradise, there is a technique we call the proablity, which identifies anomalies. This method can often detect the presence of hydrocarbons when amplitude alone could not.