As the world gets to grips with AI capabilities, new use cases are emerging. University College London is using AI to investigate healthcare, Transport for London is mapping tracks for maintenance – and the British Geological Survey (BGS) is considering how new analysis of rock data may deliver the energy transition.

The use of AI is not new for the energy industry. There are, though, new optimisations with enhanced abilities that are emerging to characterise geological samples.

AI and its cousin, deep learning, appear promising for scanning substantial subsurface data to support carbon capture and storage (CCS) and renewable energy pilot projects.

The value of subsurface data

CCS is crucial for decarbonisation. But CCS can only occur through examination of subsurface data to identify optimal projects.

High-quality subsurface data across Europe will support the pre-permit phase of energy transition projects, including CCS. For CCS sites, it is crucial to consider pore space, permeability, seal integrity, geochemical interactions, fluid flow and storage capacities.

“By providing precise geological data early in the planning process, we can reduce uncertainties and risks and expedite policy and investment decision-making,” according to EuroGeoSurveys.

Specific geologies

CCS tools must adapt to local geological conditions revealed from subsurface information, according to CGS Europe. The industry has a long way to go on considering subsurface for CO2 storage.

One area for improvement – and where AI could help – is by gathering important details within petrographic thin sections.

Subsurface information in petrographic samples can guide exploration and extraction efforts for pilot projects. Geologists use the petrographic microscope to examine rock properties and textures that reveal reservoir potential, storage capacity and geological histories.

This may involve deformed minerals impacting pores and storage volumes, to fluid interactions posing environmental, economic, and management risks for drilling sites.

Gathering this data manually is time-consuming, meticulous and prone to inaccuracy. And the larger the sample collection, the greater the problem.

AI explores existing samples

Drilling new boreholes to test geology for CCS is costly. Often, it is a better choice to use existing geological samples. The British Geological Survey (BGS) holds up to 500,000 thin sections. They were gathered with a view to resource extraction, but they can be used for other subsurface projects.

Analysing these manually would be repetitive and labour intensive. As such, the BGS is using AI to automate the analysis of samples.

BGS is hosting a Data Study Group (DSG) in September to test AI’s potential. This will explore what opportunities AI offers for analysis of samples.

Jules Manser, the applied skills programme manager at the Alan Turing Institute, explained how the DSGs work. They are “about bringing together interdisciplinary researchers with real world challenges, supporting the government missions of ensuring AI benefits all sectors and regions as well as supporting the investment in the AI ecosystem”.

The DSG will explore how to capture data from thin section images on grain (clast) mineralogy, porosity and permeability.

Paul Cleverley, visiting professor at Robert Gordon University (RGU) in Aberdeen, will host the event.

“We can use AI to generate new knowledge from vast amounts of historical archives. For example, repurposing old oil and gas reports and datasets to generate opportunities for CCS, geothermal and natural hydrogen,” he said.

Gathering and counting

AI tools can specialise in analysing visual information, like thin section analysis, using a microscope and camera. This was demonstrated at the SPE Middle East Oil and Gas Show and Conference 2019, where a fully automated workflow was created by mounting a digital camera on an optical microscope.

An algorithm to identify individual grains allowed the whole thin section to be scanned automatically. Results showed it was objective and consistent, saving geologists’ time. It went from two hours for the manual point counting exercise, to three minutes with automation.

If these AI tools work, they can significantly free up a geologist from manual data collection, reducing human error. But that does not mean AI automations are error-free.

Training the model

Deep learning algorithms can be trained to identify and learn from mistakes to enhance this visual capability. This is necessary to tackle complex geological features like specific textures with broad variations, such as mineral boundaries in limestone vs dolomite.

A 2023 study showed what happened when researchers introduced deep learning to look for complex geological tasks. It was able to identify something as subtle as an oolitic textures in thin sections. Oolites can be hard to spot owing to their low relief in some sections. But they can indicate a hydraulic shallow water depositional environment, which helps piece together reservoir history.

The image below shows a large number of ooids that are distinct from the background.

Ooids image with correct detection results. CC By 4.0. Source: Wang et al (2023).

The image below shows some misclassifications at the edges and corners. Adjusting the model can help it disregard overlapping corners and identify the ooid. The model showed mostly satisfactory results, with complete ooid forms in the centre of the image correctly pointed out.

Misclassification of ooids at the image’s edges. CC By 4.0. Source: Wang et al (2023).

AI approaches may not be perfect. However, this model can learn, refine and adjust its accuracy with training data.

Better data, better answers

These examples show training samples have an impact on accuracy. The ooids below are less clear than the previous images, with less contrast from the background. A low number of training samples caused this, with only 14% resembling the image.

The first image had 54% training samples showing similarity. Clearly, reducing the amount of training samples leads to poorer detection performance.

Poor ooids detection results. CC By 4.0. Source: Wang et al (2023).

These comparisons show how training samples help avoid model misjudgement. This explains why the model performed better on simple images than complex ones.

In deep learning, data augmentation can artificially increase a dataset’s size and variability. This can strengthen a model’s ability to detect varied and subtle geological features.

However, too much data can also be a problem, albeit one that deep learning can resolve.

There may be ways to improve accuracy and efficiency of thin section analysis. More testing and refinement of models are needed for complex cases with varied sample quality and textures.

Industry interest

Technology and energy companies have noticed the potential of AI and deep learning models. This year, numerous professionals attended a Zeiss online seminar about how their AI machine learning model can be tailored.

The company claims the right training can give its model the insight to “know” what to look for in petrographic thin sections for the energy industry.

Zeiss attributed this to its deep learning-powered “object-based” classification. The team said it was able to tailor the model to improve its analysis of sample features. This could include detection of a specific texture, microfossil, or mineral grain.

In the lab, Kelly Hattori editing a photo of a microfossil taken through a microscope at Florissant Fossil Beds National Monument, Colorado. (NPS Photo by Kelly Hattori, GIP). Source: WikiCommons.

The ability to learn and adapt is an advantage when it comes to varying rock types and geological questions. Even if the model makes a mistake, a deep learning approach, like the Zeiss model, allows it to avoid making the same mistake again, the team explained.

They mentioned their model avoids human bias and errors. Geologists are able to fine-tune the specifics of what they want, from research to exploration.

The Zeiss webinar focused on petrographic thin section analysis in particular. This is an area where technology may help most, given the painstaking nature of manual data collection.

AI advantages and caveats

Petrography has evolved with the computing age to become digitised. Now, AI is advancing this further.

When set up and trained right, and with sufficient training data, this AI assistance could do most of the petrography handiwork, but faster and with less bias.

This could allow a geologist to focus on “bigger picture” problems during energy transition projects. Avoiding bias or human error may also be a major advantage of AI models, although this needs to be tested through new projects.

This month, the BGS DSG event is a positive step. The Alan Turing Institute is showing the benefits of harnessing AI in industries, like the energy transition.

Meanwhile, the UK government has voiced its support for new AI skills and applications through its National AI Strategy. There is a global race on to develop new concepts around AI, but progress has the ability to drive global change.

It would be encouraging to see if new AI developments, such as at Zeiss, speed the delivery of the energy transition. Hopefully, as the DSG workshop demonstrates this month, digital innovations and geology skills can combine for industrial and societal benefit.

Related Content

Between gas risks and green needs

Storing carbon in “stacked pancakes”