On behalf of BHP, SRK Consulting and DeepIQ have partnered in an initiative to tailor a combined approach to improve prospectivity modelling. By combining advanced data processing, knowledge compilation, geological interpretation, and machine learning (ML), the team investigated multiple approaches to support the discovery of large porphyry copper deposits in a well-explored region in the northern Chilean Andes.
The study utilised a mineral system approach to address the available data. Based on global and local geological knowledge, the main exploration vectors and criteria for porphyry copper deposits were tabulated over the Area of Interest. All valuable information was extracted from available regional datasets using a wide range of interpretational and processing methods and algorithms. Where regional datasets were considered lacking sufficient robustness, these were reworked to ensure they were fit for ingestion into an ML-based workflow. This information was translated into evidential layers and the multi-modal datasets were ingested into a cloud data platform and integrated to form an analytics-ready dataset. These data were analysed using a combination of knowledge-driven algorithms and advanced ML and deep learning algorithms, with various strengths and weaknesses of each approach incorporated into an iterative feedback loop.
The benefits of combining geological knowledge with ML include that ML promises a more objective approach, minimisation of human bias, and the rapid interrogation of large amounts of data. However, ML can be subject to biases (e.g., the data used and/or the relationship between data and the geological processes and/or features of interest). Inappropriate data selection, poor quality data, or a tenuous or complicated link between the data and the geological feature or deposit of interest can all lead to poor outcomes. A nuanced combination of geoscience and data science can help minimise these biases and lead to better outcomes.
The data analysis outputs needed to identify known porphyry copper deposit locations with a high level of accuracy were the first step in demonstrating a credible and potentially robust approach to exploration.
Hundreds of individual datasets were generated during the study, in addition to those originally provided to the study team. How the datasets related back to the mineral system model was always factored into the feedback loop process between the geoscientists and data scientists when deciding which datasets to use in the mineral potential models. This feedback process improved understanding of the model outputs/results and subsequent refinement of the outputs.
Geological data sets included solid geology interpretation, structure reinterpretation, geochemical data analysis, mineral/deposit occurrences (including the redefinition of footprints), and 2D alteration models. Geophysical and remotely sensed data sets included magnetics, electromagnetics, gravity, minor radiometrics, digital terrain models and satellite imagery (ASTER, Sentinel 2).
Sparse and non-uniform data inputs such as geochemistry were generally used as confirmatory rather than mandatory data elements in mineral potential models, reducing the reliance on imputation (the process of replacing missing data with substituted values).
A 3D litho-structural model of the Area of Interest was also generated focusing on developing the structural and intrusive framework in providing a 3D structural and geological framework of the Area of Interest to assist with a better understanding of the architecture of the region and assess the underlying porphyry copper deposit controls in 3D – and to provide an extension of the 2D targeting framework into 3D to test the applicability of 3D targeting at a regional scale.
An extensive input dataset and inference network were used to generate the final mineral potential models via several iterations using well-established fuzzy logic workflows and advanced data analytics. The data analysis outputs needed to identify known porphyry copper deposit locations with a high level of accuracy as the first step in demonstrating a credible and potentially robust approach to exploration.
The ML-based mineral potential was estimated in 2D and 3D using distributed and scalable cloud computing frameworks. The best results were merged with the geological knowledge-driven fuzzy logic-based outputs using a Fusion approach to delineate focus areas.
Key outcomes included:
- A few key evidential layers were the main drivers.
- Most of these resulted from enhancing and reinterpreting geological datasets including creating new ones.
- Several of the new maps were also important tools in the review process and iterative analysis (including specialist peer review).
- Intrusion layers (near the surface and at depth) and major structures were important evidential layers in the final mineral potential models derived from both geological knowledge – and ML-based modelling.
- Magnetic alteration mapping was noted as an additional driver in ML modelling.
Conclusions:
- A productive, collaborative approach allowed for powerful cross-fertilisation of ideas with no one specialist or specialist team driving outcomes. A nuanced application of data analytics within a mineral system framework is optimal for regional greenfield applications.
- Enhanced and new datasets, based on geological knowledge, were very important to the outcomes achieved by this study.
- Introduction of an ML-based Fusion model was instrumental in generating optimal mineral potential models. Fusion is an effective means of combining a suite of complementary models, including knowledge-based models, that capture expert knowledge and deep learning models that are data-driven.
- Sparse and non-uniform data inputs can be used as confirmatory rather than mandatory data elements. Synthesis of sparse data inputs into geologic proxies is an effective strategy in poor data environments.
- Several high prospectivity areas were identified for further detailed review in a well-explored region.
