Artificial Intelligence (AI) is transforming the oil and gas industry at an increasingly rapid rate. No stranger to digital innovation, since the 1980’s the industry has adopted digital technologies to drive greater efficiencies, but today is different.


The biggest technology revolution we’ve seen, you’ve probably heard, “AI is the future”, but what can it do? How does this apply to the seismic interpretation workflow? And what are the implications for your people, your business and the energy industry as a whole?

AI or artificial intelligence isn’t really a technology. Designed to do the sort of things a mind can do, the key to AI is intelligence. Just like human intelligence isn’t one single dimension, AI is equally varied and structured in the same way. AI is many advanced technologies brought together to enable a machine to act with human-like levels of intelligence; providing context and meaning to the information it learns, imitate it and act accordingly.

To generalise, deep learning is a subfield of machine learning, and machine learning is a subfield of artificial intelligence. Not a definitive explanation, the best way to get a true understanding of AI is to recognise what technologies underpin it and how they can be applied to seismic interpretation. We’ll focus on the principal areas which are relevant to enhancing subsurface understanding and how they can be used to advance the seismic interpretation workflow we know today.


“Geoteric’s AI-Fault interpretation had an immediate impact on our subsurface understanding, enabling an improvement to our well planning process.”

Wellesley Petroleum |

AI’s ability to learn is essential. The ability to learn from experience without being explicitly programmed is arguably what makes machine learning (ML) intelligent and differentiates AI from other forms of computer automation.

Whilst traditional computers cannot fix problems on their own, ML can. Not only can ML learn from past data results. It can arguably make a better decision and do so at an incredible speed. This is often illustrated in popular culture examples, when AI bots take on the professionals.

However, ML is a large topic. It has several different learning styles which are important to define.



Learning is critical


Much like we learn and understand ourselves, ML must be trained too. In subsurface understanding, getting it right in our domain is more than just data science, it takes a blend of everything. From maths, physics, geology, geophysics, crucially user experience with current domain knowledge all must work together. Without that, these capabilities on their own are not enough.

Supervised learning takes labelled data which has been organised and described. It deduces the relevant features which characterises each label and learns to recognise those features in unseen data.


For example, you may show the algorithm a large number of labelled faults in seismic data. It would learn how to recognise a fault and spot it on any number of unseen seismic cubes and classify it.


Unsupervised learning doesn’t require pre-defined labels. Instead it takes an unlabelled data set, find similarities and variances within the data set, and categorises them into its own grouping.

Self-organising maps (SOM) is possibly the most common type of unsupervised learning in subsurface understanding. Classifying the data into clusters to identify important geological features during exploration, production and development.


Unsupervised Learning





Reinforcement learning is more of a trial and error feedback loop. Much like our popular culture example , the algorithm creates a “gaming environment”. After each action, it is told if it has “won” or “lost” so it can build a picture of actions which results in success.

As described below, interpreter led changes to the results can be captured within deep learning algorithm which runs on top of the seismic data. This will be used for quality control (QC).


The AI fault interpretations are proving to be very useful inputs for well planning. They are full-field from the shallow overburden to the reservoir, our most comprehensive fault interpretations to date. This has improved the rigor of our work so that our decisions are based on a better understanding of where faults may introduce drilling risks, and so that we can better understand lateral changes in our reservoir. They have also make it easy to compare quite different fault patterns defined using different seismic volumes. That this type of information can be provided in 2 weeks’ time is a real game changer.

Lead Geophysicist, Valhall |

A technique to enable the application of ML is deep learning (DL). Practised in image recognition or speech, in subsurface understanding we use the imaging side.

For DL to learn significant features within the seismic, it applies a layer (sometimes referred to as a convolution) to gain an understanding of all the features within that layer. The clue is in the word deep, the output of each filter provides the input for the next. As the processes continues to repeat; detecting both low and high-level features, the number of layers increases and as a result can be built up of countless layers.

To achieve this learning, it is powered by a neural network (most commonly CNN). A conceptual model which is based on the brain’s biology and mimics the network of neurons. As a result for an interpreter, removing the considerable amount of manual effort and time intensive feature extraction and classification.