Steve Jobs once described the computer as the bicycle of the mind. Here at Datarock, we see AI as a bicycle of the mine site.
My best friend Robot
If you grew up in Australia in the early 1990’s, you may have sat around the TV in your lounge room and watched ‘Beyond 2000’ with your family on a Saturday night. The show talked about technological innovations that were supposed to improve our lives and be widely available to the mass market after the year 2000. You may remember the introductory scene where the music plays and a golden robot pushes a kid on the swing in a grim-looking park? You may have imagined a future where childhood would be played side by side with your best friend Robot. Always someone to muck around with, as long as he brought the snacks. The years beyond 2000 seemed so thrilling and fantastical and honestly, a little scary. Even today, most articles written about Artificial Intelligence are accompanied by a graphic of a faceless head with a blue beam of light for a brain. Creepy stuff really. Perhaps it feels again like Artificial Intelligence (AI) has reached into our households, and this time the reach is not futuristic; it’s now.
Can I talk for a moment about Artificial Intelligence and Machine Learning?
The graphic of a faceless head doesn’t exactly feel warm or embracing. Its power is in its mental projection, and that’s exactly what AI algorithms are designed to do – to approximate the way a human brain thinks. When my daughter was a toddler, she had a terrifying habit of running in a car park, so I taught her to recognise our white Subaru Outback so that she would always stop when she saw it (and I made sure I parked close to the lift!). She first had to recognise that a car is fundamentally different to a bus or truck or bike, and with time she recognised the car to be the colour white, and then recognised its shape. When she later learned her letters and numbers, she memorised the number plate. This is a rudimentary example of the formation of a neural network in a growing human brain, and in essence, how an artificial neural network algorithm also finds patterns within a dataset.
The term ‘Machine Learning’ is used more often in a geological context. It’s a branch of AI, but instead of knowing a set of defined rules to answer a problem, it uses algorithms that have been trained on data to produce predictive models. For example, if you know that a quartz vein in a porphyry deposit looks mostly white in a stick of drilled core, you could provide a bunch of core photos of that quartz vein to the algorithm. You could train the algorithm to recognise different versions of that vein by showing it examples of quartz veins with different thickness, in different rocks, at different angles to the core and through different depths down the hole. The geoscientist needs to know some of this information first, and then a machine learning algorithm can find other examples of where those quartz veins exist in the entire deposit, or in historical photos where the core has disintegrated.
Machine learning can deal with things that are historically poorly audited, not collected well and do it better.
The mining industry is facing simultaneous issues:
- the demand for decarbonisation to feed the global transition to renewable energy, and
- the push for digitalisation.
There’s a need to revolutionise the industry itself. To decarbonise first, before it can then supply the rest of the world. Perhaps one of the key ways to operate under these conditions is the acquisition and processing of high-quality foundational datasets; datasets that are acquired or generated everywhere across a deposit, are of high quality and consistency and have a suitable resolution. These datasets underpin all of the future decisions around how to find, and extract, an orebody.
The problem is that the geoscientists responsible for generating and interpreting this data are under increasing pressure. Over the last decade especially, companies have significantly increased the volume of collected quantitative sensor data. A geoscientist may be drowning in data, while also having an increased responsibility to ensure site health and safety, compliance and stakeholder management. Compounding this problem is that the number of graduate geoscientists coming out of university has dropped significantly, and the number of universities offering geoscience classes has dropped with, or because of, it. We cannot just throw more people at this problem. (https://eos.org/opinions/australias-unfolding-geoscience-malady).
The number of undergraduate geoscientists enrolled at university in the past decade is dropping in Australia, United Kingdom and the United States of America (https://eos.org/opinions/australias-unfolding-geoscience-malady)
I’m paraphrasing a Maya Angelou quote here – ”you can forget what people say or do, but not how people make you feel”.
There are a subset of tasks in everyone’s job that need to be performed with auditable meticulousness, because of technical or business imperative. There are often another subset of tasks that are hardly mentally stimulating or rewarding and are often done begrudgingly or half-heartedly. When these subsets of tasks intersect, we risk facing problems of consistency and rigour, potentially impacting the certainty of downstream business decisions. Amazingly, it is those exact tasks, the dull, time-consuming, repetitive ones, that are prime for machine learning augmentation. So how about we leave those to AI? And we humans retain creativity, problem solving, communication and emotional intelligence? We keep hold of the gut feel. We keep hold of the interpretation and contextual knowledge that a geoscientist has obtained through years of study and experience working in the field, picking up rocks. We make space for our geoscientists to be experts in their geological field first, and then ‘experts in the loop’ of the algorithm. To build the training models – really scrutinise them for accuracy and quality as they are being generated – and provide feedback for model updates as the understanding of a deposit changes over time.
It’s hard, but try not to be distracted by the existential risks of AI, or at least, place them in context.
We’re often asked, will AI or machine learning take my geoscience job? The short, and truthful answer; mostly no, but some aspects of it will change. Even for the most robust generic machine learning models, for example, the generation of machine-learned RQD or fractures, it is critical to have a geoscientist reviewing, comparing and ensuring the accuracy of the output prior to accepting it into a database. Ideally and practically, machine learning used well, will provide the geoscientist with a better understanding of a model, have the ability to correct errors, and have a high level of confidence that the generated data reflects their geological understanding and ‘gut feel’. A machine learning algorithm should be a trusted tool, providing useful, cool insights. Better than useful and cool, it will help the geoscientist to log more efficiently, and more consistently. To provide spatial evidence of a vein set or fault structure that was previously only suspected. To understand a regional geological dataset or the detail of a geochemical or geophysical dataset in a different, more holistic way.
I don’t see a future where a golden robot replaces the role of a human. I see the present-day, existing algorithms offering a helpful hand to the geoscientist. To give reprieve from some aspects of a dull and repetitive workload, and provide the chance to learn more about the rocks beneath our feet. I see a future where the sensible and appropriate use of AI, and machine learning algorithms, provide actionable, geoscientific insights, from the drill to the mill.