Data analysis and decision making in the mining sector

Resumo

Mining is a traditional economic activity that has historically resisted to change. The digital transformation of mining and metals companies presents a unique opportunity to boost production, streamline processes, improve metal recovery and yield, reduce costs, minimize supply chain complexity, and ultimately, modernize their image. In 2021, at the moment of finalizing this doctoral work, mining was still far behind other sectors in terms of digitizing its processes and implementing Artificial Intelligence (AI). Specifically, Backer Hughes (2021) reports that across the industry, 1 million TB of data is generated per minute from drills, trucks, processing plants and other equipment, but less than 1% of these data is analyzed. Against this unfavorable backdrop, mining companies will have to address massive carbon reductions over the next 10 to 20 years to achieve the policy goal of net-zero carbon emissions by 2050. Data, digitalization and AI are perceived as a solution to these problems, but they are also becoming buzzwords that do not find their place in the mining industry. In this context, it is necessary to ask: how will data-driven technology change tomorrow's mining and energy work? This doctoral thesis seeks to answer that question, focusing on how to optimize decision making by leveraging the technological advances of AI-based digital platforms. To do so, it merges years of theoretical developments from world experts in probabilistic reasoning with Bayesian networks (Pearl, 1988), decision analysis (Howard, 1968) and behavioral economics (Kahneman & Tversky, 1979) into a single decision framework for evaluating complex business scenarios. In consequence, the main objective of this research work is to develop advanced decision tools using Bayesian machine learning (BayesianML) to model complex business scenarios in the mining and energy sector. As an application, multiple scenarios dominated by risks arising from soil contamination, occupational accidents and operational activities are evaluated. As a cornerstone of this thesis, information theory is explored as a differential tool to be incorporated in future automatic machine learning (AutoML) processes to understand how uncertainty is propagated in decisions conditioned by different attitudes towards risk. Overall, there is a strong need to provide the scientific community and the business world with the know-how to successfully implement AI techniques, identifying common procedural errors and extended cognitive biases. More importantly, there is still a huge need to explain how uncertainty influences decision making under risk. When talking about risk, it is inevitable to acknowledge the mining industry's exposure to black swans or highly improbable events that produce large-scale consequences. The author is aligned with Taleb's (2007) interpretation on the impossibility of predicting black swans. However, exposure to these events can be measured, and there are multiple ways to do so. In this dissertation, Bayesian networks are considered to be an ideal computational technique and data visualization tool for this purpose. In particular, the potential to create Bayesian networks that automatically learn from data (i.e., BayesianML) opens a new horizon. More specifically, BayesianML together with information theory (Shannon, 1948), could help to understand the fine line between black swans and dragon kings. Precisely, the dragon king theory (Sornette, 2009) argues that many extreme events are predictable to some extent, especially if the structure and dynamics of the complex system are analyzed. In conclusion, this doctoral thesis demonstrates how the value of information (VoI) and flexibility (VoF) can be used when interpreting uncertainty and how to take advantage of it. In this context, the results allow to conclude the usefulness of a large number of techniques such as influence diagrams (IDs), probabilistic structural equations models (PSEMs) or the creation of hybrid models with geostatistical techniques. As a final product of this doctoral thesis, the author collects the main conclusions obtained in the articles published as part of this thesis. With this knowledge, an application framework based on 4 phases (structuring, modeling, reasoning and updating) and 10 actions is proposed as an instrument to guide mining and energy professionals on how to use BayesianML, and AI in general, to model complex business scenarios and make informed decisions.