Jessell, Mark1, Guo, Jiateng2, Li, Yunqiang 2, Lindsay, Mark1
1Mineral Exploration Cooperative Research Centre, Centre for Exploration Targeting, School of Earth Sciences, UWA, Perth, Australia; 2College of Resources and Civil Engineering, Northeastern University, Shenyang, China
Although both machine learning (ML) and geophysical inversion techniques have been applied to the challenge of characterising 3D geology from available data, they each suffer from many of the common challenges facing the application of ML to geoscientific models: a spatiotemporal structure; heterogeneity of information in space and time; interest in rare phenomena (such as ore deposits or earthquakes); uncertainty in the data and a lack of ground truth (Karpatne et al., 2017). We present an open resource consisting of 1 million 3D geological models and corresponding gravity and magnetic fields; all labelled by the simplified geological history that defines the model and use realistic density and magnetic susceptibility properties. The models were constructed by randomly perturbing the model parameters (including fold wavelength, fault offset, dyke thickness etc.) and order of predefined kinematic ‘events’ using the Open Source Noddy modelling engine (Jessell, 1981; Jessell & Valenta, 1996). The first two events are predefined as a stratigraphy and a TILT of the model, and the subsequent three events can be any from the following list of seven events: TILT, FOLD, SHEAR-ZONE, FAULT, PLUG, DYKE, UNCONFORMITY with a total of 343 possible model label combinations (e.g. FOLD FAULT UNCONFORMITY). A model label such as TILT->FOLD->DYKE, has many parameters associated with the event, such as tilt angle (TILT), fold wavelength (FOLD) and dyke position (DYKE) which all modify the resulting geology. The density and magnetic susceptibility of each unit in the model are perturbed around idealised distributions for specified lithologies. Each model represents a 4km cube in the Earth divided into 20m voxels, making 8 million voxels per model. These models, which can be accessed programmatically via a small python juypter notebook, together with the full description of the parameters needed to reconstruct the model, and the gravity and magnetic fields, forms the basis for training a ML model, or used to benchmark new geophysical inversion schemes. This poster presentation will explain the methodology used, and will provide access to the models via a dedicated website.
We acknowledge the support of the MinEx CRC and the Loop: Enabling Stochastic 3D Geological Modelling (LP170100985) consortia. The work has been supported by the Mineral Exploration Cooperative Research Centre whose activities are funded by the Australian Government’s Cooperative Research Centre Programme. This is MinEx CRC Document 2020/47.
Mark Jessell is a Professor at the Centre for Exploration Targeting at The University of Western Australia. His current scientific interests revolve around integration of geology and geophysics in 3D (the Loop project), and the tectonics and metallogenesis of the West African and Guyanese Cratons (WAXI & SAXI).