SIGÉOM: 30 Years of Innovation — From Data Integration to Machine Learning

The Ministère des Ressources naturelles et de la Faune (MRNF) continues to be a global leader in geoscientific and spatial data management, thanks to the Québec Geomining Information System (SIGÉOM). This innovative geoscientific and mining tool is available to all free of charge and has been in constant evolution for more than 30 years.

This system brings together data from the MRNF's geoscientific surveys (bedrock and sedimentary cover mapping, geophysics, etc.), research work, mining industry exploration work filed for renewal of their exclusive exploration rights and mining activities disclosed by companies.

This presentation will highlight SIGÉOM's 30 years of existence and innovation, emphasizing its essential role in the knowledge of Québec's territory, the valorization of geological data and the discovery of mineral resources in Québec. The major stages in its development, marked by a number of technological advances, will be presented. The presentation will discuss the value of the system and its impact on the mineral exploration industry in Québec.

Populating geochemical data into SIGÉOM is a particularly tedious and time-consuming task. Delegating this work to artificial intelligence enables automation and significant time savings—but it also raises concerns about the quality of the results.

So, what kind of process can help resolve this dilemma?

This presentation illustrates a human-guided AI approach that enables automation while maintaining strong control over the quality of data extracted from exploration reports. We’ll explain how, over a 16-month period, 2,000 reports totalling 200,000 pages were processed to generate 500,000 geolocated analyses, an equal number of non-geolocated analyses, and around 200,000 sample descriptions without associated analyses.

Using this case study, we’ll explore how AI can support the updating and use of a rich GIS like SIGÉOM.

This presentation is based on a paper published in EarthDoc: From Québec’s Assessment Reports to the SIGEOM GIS with a Human-in-the-Loop AI.

Structured and organized databases are essential for the deployment of machine learning (ML) tools. SIGÉOM has enabled Québec to quickly distinguish itself in the development of machine learning applied to mineral exploration. In this presentation, we will highlight the progress made since 2010 in our research projects related to the use of ML in exploration using SIGÉOM data. These advances relate to predictive geology and targeting, and to the increased resolution of geophysical data.

Geology is fundamentally a naturalist science, based on the observation and detailed description of complex objects. With the digitization of data, these descriptions have often been simplified into categories or lithological codes, leading to a loss of information. Yet, a precise and well-documented description remains a stable and rich source of data, even as interpretations and coding systems evolve over time.

Large Language Models (LLMs) represent a technological bridge that reconnects detailed geological descriptions with modern digital tools. Capable of automatically processing large volumes of text, they enable the systematic extraction, structuring, and analysis of information contained in geological descriptions.

To illustrate this approach, we present an application on the Menarik Lake property in Quebec. Several outcrops have been described and mapped there. From these textual descriptions, an LLM automatically extracts structured information such as lithology, alteration, or mineralization. This information can then be spatially projected to produce an interpretive map based solely on the content of the original descriptions.

This technology opens the way to various applications, including the reclassification of historical data, automated search for geological objects, and database harmonization. By facilitating the integration of textual descriptions into digital workflows, it may help revalue field observations as a direct and usable source of geoscientific information.

We present a fully automated, data-centric workflow for mineral system exploration that transforms heterogeneous geoscientific datasets into high-resolution probabilistic prospectivity maps within a unified framework. Inputs—including geochemistry, geophysics, structural, and 3D models—are ingested through schema-driven parsers and undergo rigorous quality control for outliers and spatial-temporal consistency. Domain-aware feature data augmentation extract exploration-relevant signals via computer vision decompositions of potential fields, tensor-based fabric maps, automated lineament detection, and probabilistic lithological maps derived from geochemical ratios and spectral signatures. Generative models reconstruct sparse legacy data to augment under-sampled regions. A multi-task deep learning model, trained on over 170,000 ground-truth samples across orogenic Au, porphyry Cu-Au-Mo, and LCT pegmatite systems, performs mineral system predictions and pixel-level property reconstructions. Model interpretability is achieved through SHAP values, enabling feature importance analysis. Application to the Abitibi Greenstone Belt demonstrates accurate recovery of known gold corridors, discovery of blind targets beneath cover, and insights into multi-scale controls on mineralization. Executable in under one hour on cloud GPUs, this reproducible and transparent workflow provides rapid, scalable subsurface insights to guide exploration targeting.

Since 2003, Azimut's exploration strategy has been based on the systematic exploitation and valorization of the MRNF's SIGÉOM digital database. The purpose of processing this data is to increase the probability of discovery by identifying new major targets with signatures comparable to those of known deposits.

The data used throughout Québec, by province or geological subprovince, are mainly multi-element geochemistry of lake-bottom sediments, geophysical data (magnetism, gravimetry) and data on showings and deposits. The AZtechMineTM expert system provides an innovative way of extracting the statistical signature of mineralization and, at the same time, identifying comparable signatures in under-explored or unexplored contexts. The approach can be applied by metal or by deposit type. Targets are represented cartographically with their statistical signature converted into a relative probability of discovery. The results are then ranked empirically, taking into account the constituent parameters of the targets, their intensity, shape, size and exploration history.

In 20 years of exploration, the implementation of this province-wide predictive approach for several metals (gold, copper, nickel, uranium, lithium) has demonstrated the considerable value of the data acquired and made available by the Governement of Québec. Predictive approaches are a competitive advantage provided they can be validated by tangible results in the field, where prospecting still plays a fundamental role. Several examples of the prediction-validation relationship will be reviewed.