Skip to main content

Is the growth of AI in geoscience industries being held back by unrealistic expectations of “the magic” it can perform? Are we missing out on a valuable technology because we’re feeding it the wrong data diet?

Our Seequent Evo webinar on The application of AI and ML in Geoscience explores a range of current Artificial Intelligence and Machine Learning questions.

Six panellists touch on subjects as broad as the simplest definition – “an enabler of things” – through to the power of human versus computer judgement in the geoscience field. But one tenet they all return to is AI’s dependency on good quality data to do its job.

The combination of AI and ML holds huge potential for any sector that has geoscience at its heart, and the technologies have advanced dramatically across the last decade. For example, this year Google declared it had trained its open-source Switch Transformer AI model to scale up to a record 1.6 trillion parameters, whereas nine years earlier, a ‘mere’ 60 million had been considered a remarkable achievement.

AI has already proved its mettle in solving scientific questions beyond the ability of humans, and doing so at remarkable speed. So it’s not surprising that with the appropriate tools, approach, and data to work on, AI and ML promise to increase productivity, reduce waste, enhance the accuracy and speed of exploration, and lift the drudge of dull, repetitive tasks from geoscientists so they can focus on what they do best and add value to the business.

But to date both Al and ML have received only cautious welcomes from geoscience-based industries, and penetration in areas where they could make real contributions remains limited. Why?

Cool tools still need grounded targets

“There is a sense of hyper excitement around these technologies”, says Sebastian Goodfellow, Assistant Professor, Civil and Mineral Engineering, University of Toronto. “People want to use cool tools, but at their heart Al and Machine Learning are just mathematics, they’re not magic!” And like any other mathematical tool, they need reliable, consistent, quality data to function effectively.

Interpretation vs contradiction

For better or worse, much geoscience data is wrapped up in interpretation, and this poses an issue for AI. Even at the basic level of, say, labelling rock types for drill core data, the results can vary from geo to geo, situation to situation, with data gatherers having to synthesise all the inputs while striving to maintain consistent judgements.

“Human beings are very capable of balancing contradictions in our heads,” says Sam Cantor, Section head, Economic Geology at Minerva.

“We can log a region and call it a certain rock type, then go away, do something else, come back later and find that our present self now disagrees with our past self. We can handle that contradiction. Unfortunately machines cannot…”

Sometimes more data is not the answer

Working with good quality, unbiased data from the outset is an imperative for successful AI and ML applications, agreed all panellists on our webinar. “There is also a common wisdom in Machine Learning that the more data the better”, adds Sebastian Goodfellow. “But often what we actually see flies in the face of that.

“A smaller, well-labelled data set is, from my experience in this domain, much more useful than a big data set of noisy labels.”

Where can it work best?

So where did our panellists believe geoscience-based industries could find the most successful and profitable inroads to AI and ML usage?

Natalie Caciagli, Practice Lead Geochemistry at BHP, “Look for areas where you are currently collecting robust data, and by that I mean things like numerical geochemical data, something that’s not dependant on human judgement or observation. It’s data we already know is unbiased – a lot easier than working with visual logging comments for example – and is therefore the easiest to deal with.”

Lucy Potter, Geology and Mineral Resources Professional, “As an industry we tend to place a lot of value or ‘glamour’ on the people processing the data, when it’s the people collecting the data and doing it with consistency and the right controls who are important for AI and Machine Learning. If the data is collected properly from the outset, it gives you a real headstart.”

Sam Cantor, Section head, Economic Geology, Minerva, “For exploration projects with small teams, where it may just be one geologist, I think the more narrow AI tasks that automate a repetitive, time consuming job can be really valuable. Tasks that don’t need an expert geologist’s attention are great places to apply AI solutions, and are really implementable at these scales.”

Mike Martos, Chief Geological Data Scientist, Newmont, “As humans, we can only think probably in three, four, or five different dimensions. Computers can help us look at hundreds of different layers or columns, and try to make sense of it, or find patterns… Machine Learning is useful for reducing the dimensionality of all that data and feeding it into a more interpretable pipeline. But ultimately, to Sam’s point, that judgement decision is left to the humans.”

Mikael Arthursson, CTO at Minalyze, “Remember that it doesn’t always have to be too complex. You don’t necessarily have to throw a full blown, deep-learning algorithm at every problem out there. Sometimes a simple decision tree or a set of very clear rules might get you there fast, efficiently and simply – and still fall under the definition of AI even if it’s not deep learning.”

To watch our full webinar, follow the link below

Watch webinar

Panelists

Sebastian Goodfellow

Assistant Professor, Civil and Mineral Engineering, University of Toronto

Mike Martos

Chief Geological Data Scientist, Newmont

Natalie Caciagli

Practice Lead – Geochemistry, BHP

Lucy Potter

Geology and Mineral Resources Professional

Samuel Cantor

Section head, Economic Geology, Minerva

Mikael Arthursson

CTO, Minalyze