Artificial intelligence (AI) and machine learning are so ubiquitous in the media these days that they have garnered a healthy dose of skepticism from the public, in many cases deservedly so. Machine learning comprises computer programs that are capable of solving classification or prediction problems by making inferences and decisions from a dataset without human intervention. Many such algorithms have been around for decades; their recent successes are in part due to increased computer power but also a new generation of algorithms that are more effective than their soft-brained human creators at the tasks they were designed to complete. The application of machine learning to mineral exploration, geometallurgy, and human operator replacement is accelerating, but that does not necessarily mean that blind application of AI will result in increased productivity. The real-world achievements are expected to vary wildly depending on the potential gains, the quality of data and the expertise of the humans behind the scenes.
Google and Facebook have been using machine learning for years to curate their ads. Investment banks have been using these algorithms to parse through news to make bets on market trends milliseconds before their competitors. The newest generation of machine learning algorithms, such as neural networks and decision trees, have been enhanced with extra layers of analysis and decision-making that significantly increase prediction success. A negative byproduct of these extra layers is that the algorithms usually become black boxes from which we cannot know what features rely upon which parameters of data, or how each prediction is made. This lack of transparency often results in a lack of trust in the process and the results. These new powerful machine learning algorithms are by and large in the public domain and are even available to every Joe and Jane through cloud-based applications, such as Microsoft’s Azure platform. Development has also accelerated from a hardware standpoint; for example, the newest generation of smartphones contain chips dedicated to machine learning, enabling facial recognition and augmented reality.
What it means for mining
One of the strengths of machine learning is the efficient identification of patterns in data that enable classification. Autonomous driving relies heavily on machine learning algorithms to delimit and re-adjust to the center of the lane several times per second based primarily on photos of the road ahead. To accomplish this feat, the system needs to learn from millions of images from similar sets of cameras and distance sensors in similar conditions. This implies that exceptional circumstances such as snow in Morocco or fog in the Abitibi will cause stoppages or accidents; thus reinforcing the need for human intervention to ensure safety and productivity. The application to open-pit or underground mining is actually easier than on public roads because the extent of road networks and the number of possible collision types are very small. The potential benefit of removing the driver from mining equipment is massive from safety, to efficiency and cost reduction. This is particularly true for fly-in fly-out operations where the cost of keeping a human driver on site, fed, warm and happy are increasingly prohibitive. A few equipment manufacturers already offer autonomous fleets, and several others have working prototypes.
Related: Rio Tinto to expand autonomous truck fleet by 50 per cent by 2019
This same adeptness to classification makes the application of machine learning particularly useful in the identification of exploration targets to replace our mineral inventory. Exploration geologists have a limited capability to digest rich data provided by the new generation of exploration tools including geophysics, hyperspectral imaging, geochemistry, and mineralogy. Machine learning techniques, if properly coded and calibrated, can see through all of the layers of data to identify the combination of the features in the data that can predict the presence of ore.
The caveat
Unfortunately, these techniques are fussy about the type of data we feed them. Above all, a lot of data are required to properly train a classifier; the more data, the better. Further, the coverage and resolution of each data type should be consistent. This is rarely the case for exploration projects with various generations of surveys and drilling. The clustering of the most reliable information (drilling) around what we already know is a huge limitation of this method because we are trying to make predictions where we have no data. The application of machine learning to exploration was a central theme at the recent decennial exploration conference in Toronto. There was a consensus that these powerful techniques have huge potential but require close human supervision due to their propensity to produce spurious results. When analyzing satellite images to find ore deposits, for example, the algorithm will handily identify the repeating pattern of colourful lakes (tailings ponds) and point to sites such as the Timmins golf course and Rouyn waste-water treatment plant as great exploration targets.
Related: Guy Desharnais of SGS led his team to win Integra's Gold Rush Challenge by combining machine learning and virtual reality
We can also apply these algorithms to entire mining operations to help predict and react to different ore types and optimize operations. This concept is essentially an upgrade to geometallurgy: each block of rock within a deposit is tagged with all the information that can affect its economic viability. This should include grade, recovery, hardness, mining recovery/dilution as well as the costs to mine, process and reclaim those blocks. All of these parameters are essential to effectively optimize an operation and enable short- and medium-term planning, but they are difficult to estimate locally. The quality of machine learning predictions is highly dependent on data quality but more so on the quantity and wide distribution of data. Using carefully planned data collection procedures on all rock samples from exploration through to plant entrance using a combination of cost-effective analytical tools (e.g. handheld XRF, photographs and hyperspectral images) and a calibration survey of comprehensive metallurgical testing, we can infer the local rock properties as predicted by machine learning. The result is a baseline of rock performance that enables exponentially better optimization and planning. Mining and processing performance data are fed back into the model to improve future predictions, reduce risk and increase profits. The actual cost of such a process is tiny compared to any unexpected shut-downs or loss of efficiencies.
As with the advent of any burgeoning technology, there will be a transition period where decision-makers will be wary of technologies that have yet to show tangible results. The companies that embrace AI and machine learning to improve their profitability should expect some false starts until the human teams implementing them understand the amazing power and limitations of their application to our industry.
Guy Desharnais is the director of mineral resources evaluation at Osisko Gold Royalties and one of CIM’s current Distinguished Lecturers.
Got an opinion on one of our columns? Send your comments to editor@cim.org.
