How Geologists Can Outwit Artificial Intelligence

AI is not the same as big data; AI doesn't really exist no matter how many companies want to tell you that they are the experts

– Roger Schank (pioneer of artificial intelligence and cognitive psychology)

ABSTRACT

...for those who don’t want to read this very LONG post!

There is growing sentiment that artificial intelligence (AI) and machine learning will take the exploration and mining industry to the next level, beyond its somewhat ‘uncool’ public reputation. Does that mean that field geologists are no longer relevant? In this post, I argue that this is far from the truth. If you examine AI just below the surface to find out what it really is and how it operates, it’s clear that true AI doesn’t exist in 2018. Real AI—the sort that can replace the sparse dot-connecting ability of trained geologists—simply doesn’t exist, and if that sort of AI is possible, it may be decades away from becoming a reality. Using a simple geological example, I argue that geologists operate similarly to Columbo, a 1970s TV detective—inferring, continuously thinking, and asking tangential and often seemingly irrelevant and stupid questions about a clue that doesn’t quite fit into the whole picture. An AI machine, not even the powerful IBM Watson, would ask these questions because one or two clues are deemed statistical outliers. No amount of computer processing power, machine learning, or AI can fill the boots and brains of geologists trained in appropriate relevant skills who interpret disconnected data. At least, not in the foreseeable future. Where AI shines, and where geologists can’t compete, is in the realm of dense multidimensional datasets, but this domain is largely limited to 2D datasets that are irrelevant for hidden ore bodies. Geologists will do well if they focus on learning the basic core geological skills required and continuously invent new and novel ways to analyse sparse drilling data. I believe that the most abundant sources of hidden deposits aren’t locked up in multidimensional datasets; instead, they are sitting in unpublished and publicly unavailable drilling databases, which are scattered across separate company archives. Geologists with no programming or AI skills can interpret these hidden datasets immediately, but it’s impossible for AI to analyse them because it’s not ‘big data’—the area where AI shines. 

The end of the world, and our careers?

If the latest mining industry news is to be believed, my golden years of analysing drilling data are nearly over. My expertise, which I love and which I have devoted my professional life to, will eventually be replaced by an artificial intelligence (AI) app you can download for an introductory price of $1.99.

In the big scheme of things I shouldn’t worry because humans are doomed anyway. According to Elon Musk, Sam Harris, and the late Stephen Hawking, it’s only a matter of time before AI will take over our world and eventually destroy Earth, and all of us along with it. Then it dawned on me—SpaceX’s shuttle to Mars would be the only path to escape Armageddon! Very clever, Mr Musk. Very clever.

But, seriously, should I really be worrying about this threat to my ability to have fun interpreting geological data? Is our ability to invent new methods in mineral exploration and solving geological problems coming to an end? That is the question I explore in this post—can geologists outwit the coming AI disruption to our careers? In what specific situations can geologists still contribute to society in an AI-dominated landscape?

The predicted AI disruption

Over the past year or two, there has been a lot of publicity about AI, and it’s touted by some as the greatest ‘disruptive’ answer to the mineral exploration industry. Implicit in this predicted revolutionary change is an assumption that humans aren’t capable of analysing the unquantifiable masses of data we’ve accumulated and continue to gather; therefore, we need intelligent computer programs to identify patterns we can’t spot. Many companies are predicting that AI will heavily assist, or even take over, our human analytical functions. But is this something we should really be concerned about?

Although there is merit to some of this argument, I don’t think it will affect the expertise in which geologists really shine—an expertise that has very similar characteristics to those my TV hero, Lieutenant Columbo, excelled in. I’ll come back to Columbo later, but first I’ll give you an example where we cannot compete with AI and contrast this with where we clearly can compete with AI. I’ll explain by discussing our perception of a simple and understandable geological feature—the morphology of meandering river deposits.

Anyone can spot meandering rivers, right?

Anyone with a sense of wonder can appreciate the time, effort, and thinking that went into the production of the ‘Fisk Maps’—the Mississippi River maps by Harold Fisk (1944). It’s an amazing work of geological investigation that can also be appreciated as pure art (Figure 1). But you don’t have to be educated in geology to identify meandering river deposits. You only need to see enough of them and you can identify them—it’s something we’ve done unconsciously all our lives. 

Figure 1. One of many Mississippi River geological maps by Fisk (1944). The variously coloured channel fills are older channels that have been abandoned.

The end result of passively taking in the shapes of meandering rivers, through observation, images, and movies, means that just about anyone can spot meandering rivers and abandoned channel fills in the field, on a map, in Google Earth, or even on Mars. Similarly, machine learning algorithms can recognise meandering rivers just as they can identify cats, dogs, and humans.

Assuming we can all recognise meandering rivers, I have a challenge for you. Watch my 3-minute summer holiday video taken at Avalon Beach, just north of Sydney, Australia. The challenge is simple—see if you can spot any meandering rivers.

Did you find any?

There are literally stacks of meandering rivers shown in this video, but you may have missed them.

Okay, there aren’t any active meandering rivers in the video, but what you see is the same as the abandoned channel fills that are colourfully coded in the Fisk Maps. However, you’re seeing a vertical section through the channels, not a view from above. The sedimentary rocks seen in the cliffs and on the wave-cut platform south of Avalon Beach were deposited by meandering rivers that formed about 240 million years ago in the Triassic era. Avalon has some of the very best examples of low-energy, meandering fluvial deposits that I’ve seen anywhere in the world, so I’ll explain the features you can see in the video.

Most people would be aware that meandering rivers move from side to side and change course over time, as shown so well in Fisk’s maps of the Mississippi. As sediments build up along an active river course, the flow of water eventually seeks another location on the floodplain, simply because it is lower in elevation than the active river channel. But before such a breach occurs, the point bars of meandering rivers migrate laterally, leaving an accumulation of inclined silts and sand beds. To fully appreciate the sideways migration of meandering channels and point bars, watch this time-lapse movie: 

Point bars dip toward the active channel and the sediments accumulate (accrete) laterally, as I’ve labelled below in a screenshot from the video (Figure 2).

Figure 2. The anatomy of point bars

In the opening sequence of my Avalon video, there are four stacked meandering river deposits, but only one is obvious in the centre of the field of view above the swimming pool. The meandering river migrated from right to left (which is south), and this is indicated by the apparent dip of the low-angle cross strata that lies between two erosion surfaces marked on the photo below (labelled ‘point bar 2’; Figure 3). The inclined surfaces represent the dipping surface of the point bar that migrated with time. The other inclined point bars are not readily identifiable because of poor exposure at the pool level, and higher up, the cliff face runs parallel to the strike of the point bars. However, the easterly dip of ‘point bar 3’ in Figure 3 becomes evident when the camera rotates to the south later in the video. The reason these point bars are stacked on top of each other is that they were preserved from sedimentary basin subsidence. If the basin subsidence had been slower, these point bars from the channel base up to the overbank fines (the horizontal fine-grained silts and muds) wouldn’t have been so completely preserved. Those preserved at Avalon Beach are very special and we are lucky to have them preserved in such an easily accessible location. 

Figure 3. Meandering point bar deposits in the opening scene.

Figure 4. Meandering point bar deposits towards the end of my video. On the wave-cut platform, you can find ripples that demonstrate that the flow of the current was along the strike of the inclined beds, as expected in a meandering river.

Figure 5. A meandering river deposit with preserved stacked point bars (modified from Miall 1985). The meandering rivers at what is now Avalon Beach carried a mixture of sand, silt, and mud and the scale shown here is appropriate for the Avalon examples.

Judging from the scale of the point bar thicknesses, the rivers were around 15 m deep and ranged from 50 to 100 m wide, so they are quite small (Figure 5). If these deposits resulted from a river system as wide as the Mississippi, it would have been very difficult to identify the dipping point bar surfaces, as the slope would have been far gentler—a much larger exposure is required to spot this.

Now that you’ve seen the Avalon examples, would you be able to spot similarly small meandering river deposits if they were preserved elsewhere? Possibly, now that you’ve seen these features. But I expect most of you, especially those in the metal mining industry, would find it difficult to spot these features in the rock record. To test my expectation, I sent the Figure 4 image to 24 professional geologists from the mining industry who were on LinkedIn and asked them to identify the depositional environment of the sedimentary rocks exposed in the cliff face. The replies ranged widely and included marine, tidal, fluvial, aeolian, volcanic, and lacustrine depositional environments. Only three people identified that the inclined strata represented point bars of meandering streams. That’s a 13% identification rate, compared to the expected 100% identification rate of meandering river deposits when seen in plan view (Figure 5). This confirmed to me that only a few would be able to identify the same geological feature when it is seen from an unfamiliar viewpoint.

Now, I could have conducted a similar exercise using a completely different depositional environment, say for example, a volcanic terrain. I would have failed miserably in identifying specific volcanic facies seen in sectional views, but others, who are familiar with volcanic environments, would have excelled. However, I expect that the success rate of the geological community as a whole would not be much different to that seen in the meandering rivers example—only a minority would be able to be very specific about identifying particular environments from sectional views.

How are meandering river deposits and AI connected?

Millions of personal and aerial photos of meandering rivers and abandoned channels are available on the internet. A simple AI algorithm could trawl through and automatically identify the main features from this vast collection of photos. Humans would be equally adept at identifying and recognising meandering river deposits from the same aerial photos, although we’d be much slower.

However, this is not the case with sectional views of meandering river deposits. Meandering river deposits were not fully identified in sectional view until 1964, so the information about their features is limited. The general public wouldn’t recognise them without training, and most geologists would be unfamiliar with them in that view too. The lack of sectional view images of meandering river deposits means that AI would find it difficult to identify these features automatically. Identifying these features—and other poorly documented geological features—is an expertise geologists will continue to provide for many years to come.

So, is that it? Is our future to focus on poorly documented geological features and make a career out of that expertise?

Well, that’s a good start, but there’s a limitation to that strategy, as most of you will immediately recognise. With time, more and more photos and descriptions of point bar deposits will be available online. It’s only a matter of time before the AI bots catch up to our ability to identify meandering river deposits from just about all angles, and even from video footage.

If we want to keep working as geologists, how do we combat ever-improving AI systems? Although seems like a hopeless situation, there’s a clear solution, but we need to understand how AI works in the real world to identify its weaknesses. If you delve into this subject, you’ll eventually realise that AI, as portrayed in the media (a formidable intelligent machine that will take over the world), does not exist.

Practical AI does not exist

Most of the media publicity relating to AI refers to processing vast amounts of data that humans cannot process in one or more lifetimes.

As a result, many of us have the impression that AI is synonymous with big data. But AI isn’t just about big data processing. The ultimate AI that researchers are trying to develop is a computerised version of human cognition, including being able to excel in what humans, even babies, can do instinctively. The cognitive ability we use to connect seemingly unrelated dots to each other is a uniquely human trait, and it will be a long time before computers can do this. In reality, true AI does not actually exist—yet. What we refer to as ‘AI’ in 2018, at least in reference to data density, is illustrated in Figure 6.

Figure 6. How data density is related to artificial cognition (AI) and human cognition. Humans are adept at interpreting sparse disconnected data; AI is more adept at processing large amounts of networked data.

The only reason why AI is equated to big data in 2018 is because that’s the only realm in which AI can compete with human cognition (Figure 6). Even then, it does it in a way that is very simplistic and almost unintelligent to a human. If you want a convincing explanation of this, watch this very interesting TED talk about how an AI robot, Todai Robot, was developed to try to pass the Tokyo University entrance exam. Todai Robot failed the exam several years in a row, but it did manage to beat 80% of the students who sat the exam. You could assume from this that Todai Robot is a smart machine, almost as capable as humans, but that is far from the truth when you look at how Todai Robot operates.

As explained in the talk by the developer Noriko Arai, Todai Robot doesn’t understand any of the questions it tries to answer; instead, it makes a statistical guess at what each question is about by analysing the words in the question and searching for related words in a vast database.

Clearly, the rapid analytical capability of such a large database is not something humans can compete with, but humans can understand the question clearly in about the same time Todai Robot can analyse the entire database and take a guess what the question means. This is not remotely close to human cognition, and as a result, this AI strategy for answering a question is not analogous to real human understanding.

Todai Robot essentially works like a Google search. When you type in a question in a Google search, it ignores the question and only analyses the words. Google search doesn’t understand the question, but it searches through the network and finds the associated words, then relies on common connections between words to provide a statistically viable result. This inability to understand the question is where Todai Robot failed. It didn’t fail because it didn’t answer the question completely—instead, it was guessing what the question meant, and it did this for every question in the exam.

Does this scenario sound familiar? It reminds me of a particular geomorphology exam in the second year of university. I had failed to study enough for the subject, so I didn’t understand some of the exam questions. However, I recognised some words, so I padded the answers with anything related to these key words without having any understanding of the questions. The result? I received big fat ZEROs for my answers. Strategically, I did the right thing because at least I had some chance by guessing, but clearly this is not the smartest strategy to take in any exam. This isn’t an intelligent approach; in fact, it’s really stupid, but surprisingly this is the strategy of Todai Robot, and it seems to be the case with many AI systems if you dig down deep enough to find out how they operate.

From this perspective what the Todai Robot does is actually a level down from most students. Most students sitting the Tokyo University entrance exam would have understood the question, but probably didn’t study enough to answer the question properly. A demonstration of how unintelligent AI’s approach to answering exam questions is illustrated in Figure 7 from another TED talk by Gary Marcus. 

Figure 7. A question answered by AI gets it wrong. Source: TED talk by Gary Marcus.

Geologists would pick B as the answer to this question, but most AI would statistically guess the answer to be C because of the common word association of ‘dinosaurs’ with the key words ‘earthquake’, ‘history’, and ‘planet’. Statistical guesses show that AI is not inherently intelligent—at least not in the way humans are intelligent—because it had no capacity to reason.

The realm of less data

As imperfect as AI might be in 2018, we can’t compete with it for processing large datasets (far left in Figure 6). But given how primitive AI actually is and its limited ability to deal with disconnected data points, I would suggest that focusing on the far right of Figure 6 would serve you well for the rest of your career. You’re probably thinking that not working with large datasets goes completely against keeping your job as a geologist, and that it would be impossible because everyone is gathering and processing more and more data.

There is one way that I am aware to shift your career into the right side of Figure 6, and I’ll explain this by using the meandering river identification challenge discussed earlier.

Figure 8 shows where we are in terms of identifying meandering stream deposits in 2018. The plan view of point bars are where AI and human cognition overlap, but, at present, the section view of point bars can only be identified by experienced geologists. 

Figure 8. Plan and Section views of point bars in terms of data density in 2018. 

But with time, the situation of 2018 will change to that shown in Figure 9. This is an inevitable process of increasingly cost-effective digitisation. 

Figure 9. Plan and Section views of point bars in terms of data density in the future. 

In the future, AI will be able to automatically and easily identify more meandering channel systems on Earth, as well as on other planets. As data becomes more available, AI will be able to identify sectional views of point bars and learn how meandering rivers look in various sections and scale. Humans cannot compete with this ability to process massive datasets.

This is true for any type of data analysis. Over time, more data will be gathered, which leads to more time for AI systems to improve and replace the analytical capabilities of humans.

If the left side of Figure 9 is a ‘no-go’ zone, lost to AI, then how do we geologists get to the right side? To examine this question, in the context of meandering river deposits, we need to go back to 1937 when one person recognised what a point bar deposit looked like in section. At that time there was only one data point, not connected to anything—a scenario where current AI systems can’t operate. This is illustrated in Figure 10.

It isn’t known when and who identified the meandering river channels seen from above (labelled ‘long ago’ in Figure 10). However, the sectional view of point bars is a recently recognised feature that can traced back in the scientific literature to 1937.

Figure 10. Plan and Section views of point bars in terms of data density in the past, present, and future.

The seed of knowledge was provided by Mackin (1937), who was the first person to illustrate how fluvial channel bars migrated laterally in a river (Figure 11). However, Mackin’s description lay dormant for nearly 30 years as other fluvial research continued, including Fisk’s (1944) detailed and much quoted work on the Mississippi River. It wasn’t until 1964 that geologists were fully able to appreciate what laterally migrated point bars looked like in the ancient record (Allan, 1964). If you’re interested, Miall (1978) provides a fascinating history of the various discoveries and developments in understanding fluvial deposits in the early years of fluvial sedimentology research.

Figure 11. Laterally migrating fluvial channel deposit as observed by MacKin (1937) during (A) low water stage and (B) high water stage. The fluctuating flow energy results in alternating beds of fine and coarse grained sediments, similar to that preserved at Avalon (Figures 3 and 4). 

Surprisingly, from 1937 to 1964 identifying point bar deposits in the ancient record was wide open for anyone to pursue and develop. If I’d run the point bar quiz between these dates, perhaps even well into the 1970s, very few geologists in the world would have correctly identified the structures seen in Figures 3 and 4. Even though very detailed and excellent research work was done by many geologists during those three decades, it didn’t specifically address what point bars looked like preserved in the ancient record.

Human capacity for imagination and invention

As outlined in the TED talk by Gary Marcus, the most significant discoveries are found where little data exists (Figure 12)—this is the area where AI can’t operate but is where inventors thrive. Perhaps we might pick up only one or two clues, if we are looking for them; humans are very good at this, especially exploration geologists who are adept at working with only a few clues. These clues come from identifying odd outliers that don’t fit into our human experience. Inventors come up with one idea after another—ideas that no one has thought of before—and often these clues are right under our noses just waiting for someone to notice. 

While AI is best at coming up with answers using large amounts of data, humans are adept at formulating questions from just a few data points (Case, 2018).

Figure 12. On the left is the common information that we know the most about—this is the area in which AI can excel, leading to new discoveries that we wouldn’t have seen due to the high density of data. However, unusually important things, and new inventions, reside in the right side of this graph where there is little data to examine and where AI can’t operate.

When I was a teenager, I was obsessed with the TV detective, Lieutenant Columbo. I believe that how I operate as a geologist was influenced by how Columbo made observations and noticed small things that bothered him but didn’t fit the overall picture. It didn’t bother Columbo that everyone else thought him rather odd, which is unlike how most of us operate in the real world. Most people don’t express opinions that conflict with the majority opinion because it makes them stand out as an odd-ball. When we express a contrarian view, we are often told to keep our opinion to ourselves, or told that this is not the view of the majority; therefore, it must be wrong. Even clever AI systems. such as IBM Watson, operate like this—but that’s because Watson is trained by people who likely reflect the popular opinion of the time. A company spending millions of dollars on training IBM Watson would not take a chance with training the system using an lone thinker, or an inventor who ignores popular opinion. Therefore, as clever as it may be, Watson is unlikely to suddenly discover an outlier of any significance; instead, it will deliver outcomes consistent with the training that it received—the opinion of the majority. Perhaps Watson can point out more expected observations that were missed by the experts, but it is unlikely to identify an entirely new and novel observation—something only human inventors and investigators can imagine.

Ironically, the existence of this societal peer pressure is why some inventors are able to make a decent living—they live under the radar of most of the population, and we don’t even notice them even if we use their products every day. We’re happy to use their products, but rarely wonder how they were invented. Invariably, they were invented in that solitary place of Figure 6 where few people choose to operate, yet this is the safest place to be to escape from the cannibalisation by AI in the future. 

Innovations rarely come from public consensus and popular opinion, but from thinkers who don’t take for granted what is considered ‘common sense’ by the majority.

These inventors observe and act on a few data points—the statistical outliers. I believe that process of identifying small anomalies is the path to true influential scientific discoveries and disruptive industrial solutions. Right now (2018), AI can’t do this, so I believe this is where geologists can concentrate their efforts.

The exploration industry can only focus on Big Data

With the rising popularity of AI and machine learning in other industries, the exploration industry is also applying these methods to unearth as yet undiscovered mineral deposits. Their first port of call, of course, is 2D regional datasets, as that is where the most amount of dense information is available. With the sheer processing and analysing power that AI brings to vast arrays of multi-dimensional datasets, no doubt there will be significant discoveries that we would have never contemplated. Companies using AI exploration methods (such as Goldspot, and Quantum Pacific Exploration) will be well positioned to potentially discover significant deposits. The question is, will this application of AI methods in exploration lead to a ‘disruption’ in the exploration industry as touted by some? To answer this question, I highlight one serious issue that is of concern with the currently available AI technology—unexplainable AI.

Unexplainable AI vs Explainable AI

One of the most problematic characteristic of current AI methods is that the processed outputs cannot be explained as to how they were derived by the AI. That is, if AI has identified exploration targets to drill, we don’t know why the AI processing selected these targets—current systems of AI can’t be interrogated for an explanation. Identifying cat images on the Internet shouldn’t concern us if done by unexplainable AI, but we should be very concerned if we’re looking to spend millions of dollars on drilling that’s based on AI outputs that can’t be explained. Are exploration and mining companies likely to go ahead with drilling targets highlighted by such an unexplainable AI? I doubt it, unless the exploration manager understands the justifiable geological reasons.

AI technology, as we currently know it, is a black box, so outputs generated from dense data on the left side of Figure 6 effectively become a reduced set of data points that shift to the right side of Figure 6. These sparse points must be explained by a knowledgeable geologist, so this is another way geologists can contribute to the AI analysis workflow and remain relevant in the years to come. Even if AI methods are used to process multidimensional data, geologists would still be responsible for working out whether the targets suggested by AI are geologically reasonable.

Ultimately, the current system of throwing all our data into one AI black box and hoping for an exploration miracle is unlikely to prevail long-term.

Instead, research is being undertaken to develop a transparent ‘glass box’ system of AI (as opposed to an opaque black box)—this is known as Explainable Artificial Intelligence (XAI). In the scientific application of AI, and for mineral exploration, transparent XAI is likely to succeed over the unexplainable AI systems being used today.

Brownfield data—the hidden gem

AI may identify new deposits and perhaps reveal new classes of deposits from multidimensional 2D datasets that couldn’t be identified previously through human analysis. However, even if we discover all of the surficial deposits, it is doubtful using AI in this way will significantly change our ability to discover hidden deposits.

The reason is obvious from the meandering river example I shared at the beginning of this post. Most people, including current AI, can easily identify meandering river deposits in plan view. However, only a fraction of people can recognise sectional views of meandering river deposits. Therefore, we can conclude that recognising some geological object in plan view does not automatically lead to an understanding of the 3D continuity of this same object in the subsurface. I specifically used the meandering river example to drive home this point because it’s simple and easily recognised by everyone, including non-geologists.

Recognising hidden deposits requires a completely different skill set—an ability to understand the patterns seen in sparse drilling data below ground, and analogous to the ability of the small number of geologists who could identify sectional views of the meandering river deposits, but vastly more difficult because we are dealing with structurally deformed and metamorphosed rocks. Unlike plan views of meandering rivers, which every geologist understands, geometries that are produced from structural deformations are not understood by the majority, even in plan view. This makes the process of exploration far more difficult and challenging, especially when the basics of structural geology are rarely taught at universities.

Discovery of multiple deposits cheaply and with high confidence will unlikely to come from the AI-assisted analysis of dense 2D data acquired from the Earth’s surface. Instead, the real opportunities lie in studying the sparse historical drilling data scattered all over the place in separate company databases.

Individually, these datasets aren’t striking because most of them only contain single assay data that can’t be processed with AI. Collectively, however, they are a vast store of scattered information can be used teach us about the geometry and structural control of mineral deposits.

The current unwritten policy with these datasets is to hide them from competitors for fear of them potentially discovering secrets they may contain. However, this closed-data policy is keeping the exploration industry as whole in the equivalent of the Dark Ages.

Not only are we not learning from these databases, we’re also not discovering the vast amounts of mineral deposits hidden in these databases that only have been analysed by a few, if any. If such data were made available to geologists experienced at interpreting these datasets using modern analytical methods, many previously hidden opportunities may be identified. This would require nothing more than loading the data into a 3D viewer to start analysing the data with human eyes—no new technology needs to be invented. These opportunities are just lying dormant, waiting to be discovered.

But how about the Sigma-Lamaque gold deposit?

One of the reasons why AI has recently become prominent in mining industry news is due to Integra Gold’s Gold Rush Challenge competition in 2016. Integra Gold released a vast amount of digitised data from the Sigma-Lamaque gold deposit in late 2015. Several of the top five contestants used machine learning methods to analyse this large database and identify exploration targets in 3D (the presentation by the winning team, SGS Geostat, can be found here). Although these success stories would make most people conclude that AI is here to stay to analyse drill hole data, this is not quite an accurate representation of reality.

The Sigma-Lamaque mining dataset was unusual for two reasons—it was very large, and it had been ‘cleaned’ over many months before being released to the public. Despite this, the total number of data points and data layers were quite small compared to datasets commonly used for AI analysis in other industries. This is shown in Figure 13—the Sigma-Lamaque dataset (red dot) is at the far left end of most historical mining datasets that are saved in company archives (green dot and extent). Most mining companies’ historical datasets are very small, sparse, and not located in the data-dense field where AI methods can be applied. They lie well within the low data density area where human analysis is still only the viable option for analysis. This is changing as new analytical methods are introduced into the mining industry (purple arrow), but even with the acquisition of such high-density multidimensional data, aspects of the data still fall in the sparse data category, which requires human interpretation.

Figure 13. Sigma-Lamaque dataset lies in the far left side of most drilling datasets of historical prospects and deposits. Modern multi-dimensional data obtained from drilling extend further into the realm of big data, which require an AI approach to processing.

Contrary to the popular opinion that AI methods will revolutionise and disrupt the exploration and mining industry, if you look at the detail, the only narrow area for which that comment is relevant is for 2D surface data and multi-attribute drilling data being produced now and into the future. It doesn’t apply to most historical drilling data, which are sparse and only contain a single column of assay values. I would argue that most hidden opportunities are buried in archived and locked-up sparse datasets—these are well within the realm of human cognition and interpretation.

Conclusion

In this post, I discussed the logic behind how sparse data and human logical thinking are where geologists would be able to operate for many decades to come, unencumbered by AI. This is the arena of the inventive mind, and humans are the only ones who can identify a few outliers in the domain of few data and come up with convincing testable ideas.

This post demonstrates the process of connecting these disconnected and unrelated dots into one (hopefully) coherent argument, and these dots were:

• AI may end up destroying the world, according to some intelligent people.
• Maps of rivers of the Mississippi River from 1944.
• A holiday video of cliff exposures of Triassic fluvial rocks in Sydney, Australia.
• Practical AI does not exist.
• AI does poorly in the realm of sparse data.
• Unusual outliers and inventing something new is a human ability.
• An obscure geomorphology publication from 1937.
• Attention to small details, as demonstrated by the TV detective Columbo from the 1970s.
• Analysis of vast amounts of 2D data by AI cannot address the task of identifying hidden deposits below the earth’s surface.
• The low-cost and easy opportunities lie in brownfield exploration, but policies that hide away drilling data is hurting the exploration and mining industry.

I could only put all these seemingly unrelated subjects together in a coherent narrative because I happen to be human, and connecting the dots also just happens to be what every other geologist (and human) is particularly good at.

So, am I worried about AI destroying our livelihood as geologists?

Well, let’s just say that I won’t signing up as a SpaceX customer any time soon.

I’m sorry, Mr Musk.

---

Jun Cowan, PhD, is a director and principal structural geologist of consulting firm Orefind, and the conceptual founder of Leapfrog geological modelling software. He holds an Adjunct Senior Research Fellow position at the School of Geosciences, Monash University.

Based in Fremantle, Western Australia, Orefind is a geological consulting company founded by structural geologists Brett Davis and Jun Cowan. Visit www.orefind.com for more information. This post, and many more like this, can be found on the Orefind website. Constructive feedback is always appreciated.

References

Allan, J.R.L., 1964, Studies in fluviatile sedimentation: six cyclothems from the Lower Old Red Sandstone, Anglo-Welsh Basin. Sedimentology 3, 163–198.

Case, N., 2018, How To Become A Centaur. The Journal of Design and Science (JoDS). MIT Press.

Fisk, H.N., 1944, Geological investigation of the alluvial valley of the Lower Mississippi River activity: Mississippi River Commission, 82p.

Mackin, J.H., 1937, Erosional history of the Big Horn Basin, Wyoming. GSA Bulletin, 48, 813–894.

Miall, A.D., 1978, Fluvial sedimentology: an historical review. In Fluvial Sedimentology (A.D. Miall, Ed). Canadian Society of Petroleum Geologists, Memoir 5. 859pp.

Miall, A.D., 1985, Architectural-Element Analysis: A new method of facies analysis applied to fluvial deposits. Earth-Science Reviews 22, 261–308.