There has been a lot more interest in artificial intelligence (AI) from a range of industries, including the mining industry, in recent years - there has been a consistent uptake of AI-powered technologies across all stages in the industry, from early exploration to resource estimation and production.
Roman Teslyuk, founder and CEO of Earth AI, says: "The general opinion has certainly changed from the fear of being replaced by robots, to the fear of missing out on the benefits AI can bring."
AI has also been a major topic at mining conferences worldwide, and every major firm has made some sort of step into exploring and embracing AI.
"Embracing AI has the potential to impact every part of the industry, not just mineral exploration," comments Denis Laviolette, CEO, president and founder of GoldSpot Discoveries. "From our end, we focus on the application of AI to mineral exploration and resource investing. However, strides have been made to streamline operations, health and safety, risk reduction, mapping and data collection, automate vehicles and drill equipment, and move to more data-driven decision-making."
He adds that the industry is very much at the beginning of its journey to embrace AI but sees it as inevitable as companies move towards creating competitive advantages through data. Laviolette says: "In the next few years, you'll start to see more use cases at conferences and the talk shift from the ‘what and why' of AI, to the ‘how and what's next' as more success stories circulate."
Martin Blouin, CEO and lead AI developer at Geolearn, cites the Integra Gold Rush Challenge of 2015-16 as an important concrete example of this increase in interest in the mining industry, marking a turning point.
It was an incentive prize challenge run by Integra Gold Corp (part of Eldorado Gold since 2017) and powered by the HeroX crowdsourcing platform. After registering for the challenge, competitors from any background or country were given access to a database built from 6TB and 75 years of historical mining and exploration data from Integra Gold's formerly producing Sigma/Lamaque mines in Val-d'Or, Quebec, Canada.
The winners, the SGS Geostat team from Quebec, utilised a combination of machine learning and traditional geological methods to produce targets across the Lamaque project and was awarded a C$500,000 (US$380,000) prize.
Blouin adds: "More recently, the Unearthed community has become invested in creating tools for the mining industry, with most of them including artificial intelligence."
John McGaughey, founder and president of Mira Geoscience, agrees that AI has been a growing subject of conversation in the mining industry over the last five years or so, but cautions: "There has been lots of noise and confusion as to how to apply it correctly and generate useful outcomes for the mining industry.
"Its practical application in the context of image filtering and recognition, geochemical interpretation, remote sensing, and predictive exploration models are already having an impact on standard practices in numerous companies."
As the mining industry learns to use AI in more applications, it will become a standard tool, in the same way that geographic information systems (GIS) or geostatistics have in the last 20 years.
However, in the opinion of Réjean Girard, CEO and general manager of IOS Services Géoscientifiques, the capabilities of AI have been misunderstood and not pushed enough.
"To have a sustainable interest from the mining industry, it must get out of the floodlights and be integrated into daily applications," he says. "The perception of AI by the mining community must shift from sensationalistic to realistic i.e. what it can actually do."
There are some significant benefits of using AI over more traditional methods of mineral exploration. Traditional approaches to data integration applied to geology typically focus on statistical classification models, employing numerous assumptions that are generally not met in practice.
"Geology is a natural science that involves large amount of heterogeneous data, much of which is not parametric or stochastic," explains Girard. "The human brain is wired to detect patterns it has been trained for but is quite inefficient at discovering new patterns. How can we recognise the meaning of a rock if we have never seen such rock before? With exclusion of geophysics and geochemistry, most geological data cannot be processed mathematically, as it is subjective. How green are basalts?"
This is where AI comes in, as a new set of processing tools capable of using heterogeneous data to detect known or even unknown patterns. Girard says: "It is not taking over from traditional data analysis or decision-making tools; it complements them in a fast and replicable manner and displaces the role of the geologist from interpreting data to interpreting results."
Many companies use AI for replicating traditional thinking in a reliable, fast and thorough manner. Girard notes: "For example, AI can be used to classify features such as minerals from photographs, to detect discrepancies between text description and coded form, to highlight unusual associations between chemical elements, to provide proxies when a critical data is missing for a model, etc. And it works, as long as that the training data set is based on reliable data."
Laviolette agrees that mining is a big data game. He says: "On any single project, there are decades of maps and surveys spanning geophysics, geochemistry, satellite data and geology."
In addition, a single drill hole can produce 200MB of data, and exploration on a single property can consist of hundreds of drill holes - this adds up to terabytes of data. The sheer amount of data that can be collected is quickly outpacing the industry's ability to process all this information.
Laviolette suggests: "This is where the power of AI comes in - taking large quantities of data spanning different data types and deriving insights from it in collaboration with earth scientists and data scientists. It's not about replacing the geologist, rather it's about supplementing traditional methods, adding a powerful tool in the toolbox to spend less, minimise error, reduce guesswork, and increase rates of discoveries, all while exploring less.
According to Teslyuk, the key benefits of exploration AI systems can be summarised in three main points:
- They can discover something that has been overlooked for decades. Traditional methods usually follow some particular logic or methodology and tend to build up on previous mistakes or false facts. Large training databases and objective data sources usually eliminate the bias-related issues and produce objective results.
- AI prediction performance can be monitored, measured and improved, unlike traditional exploration where success metrics are rarely measured.
- With modern parallel computing, it is fast to learn from data sets that contain billions of datapoints. For example, Earth AI's worldwide data set is 600TB - there is so much knowledge there, no single person will ever be able to look through it and learn the relationships. In contrast, traditional exploration methods are usually limited to a particular project area.
AI is also a lot more flexible than computer vision methods. Blouin says: "Because artificial intelligence algorithms train on data labelled by professionals, such algorithms focus directly on the expected results, instead of a translation of the expected results by a programmer, which guarantees results comparable to those of professionals.
"Moreover, artificial intelligence is able to tackle subjective tasks or tasks that require abstractness, such as recognising objects. Naive computer vision cannot tackle such tasks."
In addition, AI is attractive for its ability to quantify uncertainty in classification tasks. "Most algorithms output naturally their degree of confidence," explains Blouin. "For instance, an algorithm capable of detecting the rock type in a photo, such as Predikor, will assign a degree of confidence to every defined rock type and will select the most trusted one."
However, AI - and its subset, machine learning - are often seen as hard to interpret and trust because the workflow, decision factors and algorithms can seem obscure. "It is important to note that there is no unique AI solution to all problems, but it may lead to opening the door to new ideas," says Jean-Philippe Paiement, director of global consulting at Mira Geoscience.
"Thinking outside of the box is necessary to make new discoveries as exploration is becoming harder, with greater focus at depth or under cover. AI application to interpretative problems can remove cognitive biases from interpretation and provide in-depth insights into multivariate environments."
He points out that recent, successful advances have been made in stochastic, non-Euclidean approaches to the problem of understanding complex data relationships. "Predictive models are constructed from the integration of complex data sets without the limiting assumptions of traditional statistical approaches," observes Paiement. "These new approaches can easily handle continuous, discrete, noisy and missing data with limited impacts from interpretative assumptions."
Michel Fontaine, president and CEO of Albert Mining, notes the expression: ‘The best place to find a mine is beside a mine'.
"This is exactly what our CARDS [Computer Aided Resource Detection System] system is able to do," he says. "The algorithms are able to extract the digital signature of a region that has already discovered gold, copper, nickel, etc in the past. We are using the discoveries of the past to predict the future."
Factors to consider
There are several things that mining companies should consider if they are interested in using AI for mineral exploration. First, they should decide whether they plan to develop a data capability in-house or seek a partnership with an experienced external firm.
"There are pros and cons to both approaches," suggests GoldSpot's Laviolette. "Building in-house capabilities can provide the long-term competitive advantage and expertise in data that sets you apart from your competitors. However, this route is capital intensive at first and can be riddled with organisational and operational challenges.
"On the other hand, working with a partner that has proven expertise can allow you to receive results quickly, and at the fraction of the cost of building a new team. Many firms prefer starting with this route to assess how AI can truly make a difference to their business and show results within a few months of an engagement."
Next, the company needs to do a data assessment and understand if they have the right type, quality and quantity of data necessary for AI to work on mineral exploration. "For example, for specific algorithms to work, there needs to be a high quantity of high-quality images," says Laviolette. "That being said, there are ways to work around some of these constraints, and GoldSpot has gained experience working in different environments, whether it be on a regional or a local scale.
"Many different types of AI techniques need a minimum amount of good quality data in order to be effective, and it will depend on the problem you are trying to solve. Is your data of the right resolution? Do you have enough data? Do you have the right data types? If not, you can then take steps to rectify any gaps in your data assessment."
AI is about inferring the abstract reasoning linking data and its labels, i.e. the classes or values associated with it, for instance rock type. "This task is not easy for a machine," comments Geolearn's Blouin. "They require lots of examples and a wide variety of them, in order for the algorithms to have a general idea of the phenomenon. Therefore, mining companies should absolutely keep all sorts of data they acquire and keep it stored in an easily manipulated numerical format, such as tables."
He recommends that companies should also monitor the quality of their data and labels, as the quality of the algorithm depends directly on data and its quality. He says: "This includes keeping track of parts of data sets they trust more, so that data scientists have smaller, quality data sets on which to focus."
Most mining companies are under the wrong impression that AI requires sophisticated machinery. Blouin observes: "However, little to no workflow modification is ever needed and AI tools can be run on computers."
Mira Geoscience's Paiement thinks that the mining industry has not entered the real realm of the big data problems yet, as data is typically still scarce, and quality is not always the focus of the collection campaigns.
"Setting up the problem correctly and identifying the relevant data to apply machine learning techniques are the key steps in the workflow," he says. "Data quality is also crucial. In the future, companies should focus on proper data acquisition from all types of sampling sources (such as drilling or fieldwork) and at least include systematic high-resolution pictures (including hyperspectral), physical rock properties measurements (magnetic susceptibility, density, conductivity), and geochemical assaying for every sample.
"It would be very interesting for systematic wireline logging to gain more traction in our side of the business."
Paiement adds that the potential benefits of AI in mineral exploration are staggeringly large, yet its application is far from simple. "In the example of mineral deposit targeting, explorers are trying to identify the location of ore deposits at the core of a very complex, natural system - the result of millions of years of the geological processes including structural reworking and hydrothermal alteration," he says.
"Evidence of the deposit footprint must be assembled from interpretation of subtle alteration effects extending kilometres from the target. This is a very different, and much more complex, circumstance than can be found in traditional applications of AI.
"The route to solving these challenges includes taking the focus off the methods of AI as a discipline unto itself and putting the focus on how the mineral exploration problem is set up for AI predictive models to work coherently. This is where deep domain knowledge and a mining industry-specific, supporting computational framework is required."
Earth AI's Teslyuk recommends that companies should start with measuring their current cost efficiencies and success rates. He says: "This way, [they] can benchmark and see if AI-powered systems improve these metrics."
It is also advisable to build a constant feedback loop for testing new approaches and dealing with problems. "Just give it a good shot," advises Teslyuk. "Adopting something new is never easy and it requires both enthusiasm and persistence."
Albert Mining's Fontaine says that if companies want to save time and money, AI is the only way to work in mineral exploration right now. He suggests: "If you have already found mineralisation on your project, we can extract the digital signature of this discovery and find it by using technologies that have already been proven several times to work.
"The most common problem is the coverage of large exploration areas by overburden. You have rock specialists (geologists) who do not see the rock because it is covered with sand, earth, forests, rocks and others."
However, IOS Services Géoscientifiques' Girard says that one of the failure factors of AI in mineral exploration is to perceive it as a panacea. "There is a learning curve to go through," he warns.
"Pouring money to play with the software is not sustainable! AI cannot compensate for poor decision-making. Cut your teeth on small projects that have potential for high success rate, using mature, available and applicable AI technologies. Their immediate payback will be your improved soundness."
Mining Magazine spoke to several providers of AI technology used in mineral exploration to find out more about the solutions that they offer.
Albert Mining is a services company using AI that has an extensive portfolio of gold, copper and zinc properties in Quebec, Canada. The company uses its proprietary Computer Aided Resources Detection System (CARDS) to help mineral exploration professionals identify areas with a high statistical probability of similarity to known areas of mineralisation.
"The backbone of CARDS is the MCubiX-KE data mining engine," says Fontaine. "MCubiX-KE uses powerful pattern recognition algorithms to learn the ‘signatures' or ‘fingerprints' of known mineralised sites, uses these as a training data set, and identifies points (targets) with a high statistical probability of similarity to known areas of mineralisation across less explored regions."
Data is entered into CARDS in the form of geo-referenced data points and images. Each point in the database is linked to its own set of characteristics that are extracted from a variety of sources, such as:
- Proximity to lithological contacts / specific intrusive suites; or
- Satellite imagery;
- Geochemical surveys: rock, soil, lake bottom, drill hole assays;
- Geological maps: rock type, alteration;
- Geophysical surveys: magnetic fields, derivative fields, gravity, radiometry;
- Digital elevation models;
- Proximity to mineral occurrences / mineralised drill holes;
- Proximity to interpreted lineaments / mapped faults and shear zones.
"You don't need to hold all this variety of data or variables," explains Fontaine. "More than 80% of all our discoveries were done just using geophysical surveys and digital elevation models.
"In addition, in the analysis of each point in the database, and the characteristics of all points within a specified distance (neighbourhood) are weighed into the evaluation of that point. In this manner, points lacking data can still be highlighted by CARDS if the combination of their limited characteristics and their proximity to points with other significant characteristics is similar to that of known positive points."
Albert Mining recommends that targets generated by CARDS should be evaluated in conjunction with all readily available geological data in the evaluation of the economic potential of a property as well as in the outlining of exploration targets.
There is a roster of well-known mining companies that have used CARDS technology to augment their mineral exploration programmes. Fontaine says the technology has already proven beyond all doubt that it works, with 30 client discoveries.
For example, in March this year, Spruce Ridge Resources announced a significant nickel, cobalt, palladium and platinum discovery near Timmins in Ontario, Canada. The company namechecked Albert Mining as providing an AI review of data that identified the area as being prospective for nickel.
In September 2018, Falco Resources announced that fieldwork had resulted in the discovery of new gold showings at Four Corners, approximately 20km from its Horne 5 project in Quebec, Canada. It had appointed Albert Mining the previous December to analyse historical data in the Rouyn-Noranda mining camp using its CARDS technology.
The AI software learned the signatures of positive and negative gold and base metals targets and identified new targets with a high discovery potential. Of the 50 anomalies identified, 15 were gold, 13 were copper, 11 were zinc and 11 were silver; these anomalies were grouped into 11 exploration areas.
Earth AI's mission is to fundamentally improve the efficiency of mineral exploration in order to provide enough metals and minerals for the current and future generations, and it has evolved into a vertically integrated mineral exploration company.
"We have a benchmarked greenfields success rate of 26%, meaning every fourth AI target turns into a prospect that can be visually recognised and has a chemical enrichment," notes Teslyuk. "We have our geological teams, geophysical drones and recently we have built and started in-house drilling operations."
Earth AI's drill rig is very light, requires no track or site preparation and makes minimal environmental impact. Teslyuk says: "The rig has a system of sensors and we are automating the controls."
The company holds nine exploration licences, containing 18 projects in greenfield Northern Territory of Australia.
"Our Elkedra North project is probably the world's first AI-discovered mineral deposit," states Teslyuk. "[It was] predicted by our AI system as a lead target, and sampled by me personally last year; it turned out to be a 1.5km-long gossanous mineralised zone with 2.1% lead and 0.5% vanadium pentoxide and minor silver and rare earths enrichments in surface rock samples."
Earth AI will be releasing the drilling results from the Elkedra North project soon.
Teslyuk says: "We hope this precedent will stand as solid proof that the times have come when AI-powered exploration methods have become better and more cost-efficient than traditional methods."
Geolearn currently provides three products for mineral exploration using AI, as well as custom solutions and AI courses. Blouin explains: "All these products - Predikor, UnBOX and BEARD - rely heavily on deep learning, a branch of AI."
Predikor provides logs of multiple rock properties, such as veins, gold content or alteration, directly from core photos. "Geologists can thereafter counter-verify their work or use Predikor as a draft log," says Blouin. "Predikor yields uncertainties, which allows geologists to focus on parts of the cores that cannot be described easily from pictures."
UnBOX allows users to linearise their core box photos, i.e. extract the cores from their boxes and combine them into a single core. This allows better and friendlier modelling.
Finally, BEARD [Borehole Enhanced and Automated Realtime Description] automatically describes rock chips contents from photos.
Geolearn has been working with mineral exploration companies but could not provide further details. It has also been working with the Geological Survey of Canada on advanced processing of geophysical data using AI.
GoldSpot Discoveries' AI solutions target big-data problems, making full use of historically unutilised data to better comprehend resource property potential.
The company has developed a monetisation strategy into multiple verticals of the mining and investment industry, including service offerings, staking and royalty acquisition, and the development of its own AI-driven trading platform.
"The idea is simple - we believe that mineral deposits form in the earth's crust for a reason," comments Laviolette. "Then we use our secret weapon - a truly cross-functional team of earth and data scientists to assess the data of a given property, in order to understand what the recipe is for mineral deposits."
The beginning of every project begins with some of GoldSpot's team on the ground with the clients on the property itself. Laviolette says: "This step is crucial to understanding the data and nuances of the property that only the clients will have.
"We then take this knowledge away, assess, process, transform and stitch together all the data for the property in a way that our algorithms understand. Our data scientists then collaborate extensively with our earth scientists and build models to understand what the recipe is for mineral deposits on the property.
"Once we have perfected, we then apply our models to the entire property to understand which areas have the highest potential for mineralisation."
Targets are formed based on these models, and are shared with the client, along with all the newly cleaned data processed by the GoldSpot team.
"Every time we have worked with a major, upon receipt of the deliverables they have invested in us," notes Laviolette. "This is an incredible testimony to our ability to drive results with AI in mineral exploration."
GoldSpot's investors include McEwen Mining, Hochschild Mining and Sprott Mining, and the company has done multiple projects for its clients.
IOS Services Géoscientifiques
IOS Services Géoscientifiques seldom offers plain AI services to the industry, but the company uses AI in a wide array of projects it does for clients. "AI is not a goal, not a purpose, not a panacea, not a fashion, not a product!" says Girard. "It is a tool to be used in daily operations to solve a multitude of problems, big and small."
The company uses AI in a range of tasks, such as:
- Gold grain detection and counting from an automated optical microscope to control recoveries in mills; and
- Resource classification involving arbitrary ore qualities in operating mine-smelter complexes;
- Exploration targeting at the deposit scale, property scale or provincial scale;
- Geochemical fingerprinting in complex environment such as peat or claybelts;
- Ore type classification based on drill core photographs;
- Mineral classification in sands.
In the last few years, one of the company's AI applications has been embedded in ARTGold, a successful product that automates the detrital gold grain counting in glacial sediments. Girard explains: "The techniques generate gold grain images by the thousands with the use of automated SEM [scanning electron microscopes], that all need to be classified in regard of shape and texture.
"While the conventional technique requires the skill of a trained geologist to classify [the images] underneath the microscope, with horrendous rate of misclassification, we trained a convolutional neuron network to do the job, with 90% accuracy, in minutes.
"We pushed the technique so we can now tell if a gold grain has been derived from a sulphide-rich rock, a quartz vein, a sheared rock or else, to better target the source. This is predictive geology!"
Mira Geoscience's staff have developed an in-depth knowledge of the machine learning world and are able to exploit the advantages of both supervised and unsupervised learning. The company offers tailored solutions to the industry to solve specific problems such as:
- Targeting of mineral systems; and
- Multivariate interpretation;
- Data filtering and clustering of geochemistry;
- Burden modelling in the coal industry;
- Geohazard forecasting.
"At Mira Geoscience, we have expertise in applying machine learning algorithms as custom solutions to complex geological and geotechnical problems," says McGaughey. "Our team is proficient at identifying the right algorithm to solve a given problem and evaluate its performance in prediction outcomes. We apply our advanced geological modelling and geophysical inversion capability to create a fully integrated interpretation that captures all the relevant features in a single earth model."
The company has also developed a supporting data management and computational framework through intensive, industry-collaborative R&D over the past five years, which directly addresses the problem of quantitatively integrating 3-D and 4-D mineral exploration data sets and interpretation. The system, Geoscience INTEGRATOR, brings together structured and unstructured data and interpretation, from drill hole data to all types of geological, geophysical and geochemical data.
"It is fully 4-D, tracking both space and time," explains Paiement. "Documents and files can be stored, managed, and linked to data and interpretation to provide relevant metadata and contextual links. It is the industry's first multi-disciplinary, 4-D data management framework and, as such, delivers the platform required by computational systems such as AI, that aim to answer questions that only quantitative data integration can answer."
Most importantly, the system provides a ‘data fusion' capability specifically aimed at mining industry problems. Paiement says: "Thus, the fundamental missing link standing in the way of AI success in the mining industry is overcome. This is game-changing technology for mineral exploration, providing a sound, robust solution to the once-intractable problem of integrating highly disparate data across space and time."
In sparse data or grassroot environments, Paiement notes that it is useful to be able to quickly identify alteration footprints of iron oxide copper gold (IOCG) mineralisation. "In the case of the Mont Dore project area in Australia, we tested whether it would be possible to identify the alteration footprint of IOCG deposits using airborne magnetic data as a starting point," he says. "We had access to magnetics and radiometric data. Using the four features and a hierarchical clustering (unsupervised learning) approach, we tested the ability to generate proxies for the mass balance of potassium and uranium. The clustering method enabled the generation of groups of data points with similar magnetic, potassium, thorium and uranium signatures, which exhibits remarkable correspondence to the geological map.
"Once the groups were established, a 3-D regression was used to estimate the expected potassium and uranium signatures for each group of points. Then, a residual value was calculated and used as a proxy for mass balance of both elements.
"For example, areas with excess potassium compared to the expected, calculated value are considered as being enriched in potassium by alteration processes. Both the potassium and uranium residuals are then combined to produce an alteration map for IOCG deposit footprints; areas of both potassium and uranium enrichment are considered prospective. The approach performed quite nicely at finding already known deposits in the Mont Dore area."
Since 2015, Mira Geoscience has been applying Bearing Point's predictive analytics approach (HyperCube) to exploration targeting. "This approach can be applied wherever conventional weights-ofevidence, logistic regression, neural networks or other data-driven approaches would be appropriate," notes McGaughey.
"Hypercube analyses relationships amongst many variables simultaneously in multi-dimensional data space rather than criteria by criteria. It removes the difficulties of determining ‘cut-offs' or thresholds for individual exploration criteria by replacing them with more interpretively useful multi-parameter ‘rules' driven by geological reasoning."
The company also used this predictive modelling for mineral system targeting at the Mont Dore project, producing a 3-D model and predictive exploration map using the weights-of-evidence (WofE) approach
"Using the same data sets, we tested the power of predictive analytics," says McGaughey. "The results were much more useful and performed better at finding known deposits."