"At the Frotet project in Quebec, CGI's 3D inversion models were integral in imaging gold-bearing structures within sulphidized intrusive rocks that are completely concealed by glacial deposits and lakes. From their models, we were able to target small chargeable zones which ultimately led to the Regnault discovery in March, 2020."
Executive Vice President of Exploration
Given all the laws of physics, the precise deterministic relationship between acquired data and actionable insight isn’t available in many industrial applications. AI techniques offer a powerful data-driven approach to learn unknown relationships between the acquired data and the actionable insight desired. Once an AI system learns a relationship, it can be used to confidently predict future scenarios.
Leveraging its extensive experience solving inverse problems for a multitude of sparse, incomplete and low signal-to-noise data, Computational Geosciences has pioneered VNet deep learning convolutional neural networks (CNNs). This multi-resolution architecture integrates across varied length scales and data types. This allows for not only geo-spatial relationships across multiple length scales to be learned, but also subtle intra-dataset spatial dependencies.
Based on decades of experience working with-real world geoscience data, Computational Geosciences’ AI solutions are founded on data quality control and a unique combination of domain expertise and deep knowledge of machine learning techniques. Additional capabilities in data augmentation, semi-supervised learning, and graph neural networks round out a portfolio of solutions that can be practically leveraged for any application in the discovery-to-closure lifecycle.
Many prospectivity machine learning algorithms used in the past (such as weights of evidence, random forests and self-organizing maps) processed data in a point-wise manner. That is, they take geoscience data at a particular spatial point (i.e., the magnetic value at a certain spot) and they discern relationships based on data values alone, without factoring in any geospatial patterns. In simple terms this means that whether a point has a high magnetic or low magnetic reading is the only thing that is important, while ignoring whether this point sits in the middle of a large anomaly, on the edge, or in a flat region with little change.
To make use of the advancing field of CNNs, CGI developed the VNet CNN architecture with the goal of using it specifically for geoscience applications. The VNet is a multi-resolution architecture that can integrate across varied length scales and multiple data types. Different geoscientific datasets can be combined by forming a multi-channel image, where each image channel (or layer) represents a dataset.
Within the architecture, the convolution operators operate on both the spatial and intra-channel dimensions. Hence the VNet network can detect both subtle spatial patterns within a single dataset, and between the datasets. This approach is significantly more powerful than point-wise algorithms since it captures the complexity of geoscientific data. The VNet architecture is the best approach to ensuring that all spatial patterns from the various datasets are analyzed, leading to the best possible output.