Algorithmic Geology: Tackling Methodological Challenges in Applying Machine Learning to Rock Engineering

Technological advancements have made rock engineering more data-driven, leading to increased use of machine learning (ML). While the use of ML in rock engineering has the potential to transform the industry, several methodological issues should first be addressed: (i) rock engineering’s use of biased (poor quality) data, resulting in biased ML models and (ii) limited rock mass classification and characterization data. If these issues are not addressed, rock engineering risks using unreliable ML models that can have potential real-life adverse impacts. This paper aims to provide an overview of these methodological issues and demonstrate their impact on the reliability of ML models using surrogate models. To take full advantage of the benefits of ML, rock engineers should make sure that their ML models are reliable by ensuring that there are sufficient unbiased data to develop reliable ML models. In the context of this paper, the term sufficient retains a relative meaning since the amount of data that is sufficient to develop reliable a ML models depends on the problem under consideration and the application of the ML model (e.g., pre-feasibility, feasibility, design stage).