**Gillfeather-clark, Tasman ^{1},** Horrocks, Dr Tom

^{1}, Wedge, Dr Daniel

^{1}, Holden, Prof Eun-Jung

^{1}

^{1}Center of Exploration Targeting, School of Earth Sciences, University Of Western Australia, , Australia

A common problem in geoscience is spatially extrapolating sparse data, such as drill core logs or geochemical assays, in the presence of larger encompassing datasets such as inversion volumes. For example, a seismic-derived basement surface may be refined using a collocated physical model in which the basement and regolith are in high contrast, for example, an airborne electromagnetic (AEM) inversion volume. What makes this integration challenging for an interpreter or a machine learning approach is relating different datasets in a way that is both robust and scalable. While co-kriging is a scalable approach for this problem it requires variables to be consistently correlated, where graphs can address changing relationships. We propose the usage of graphs to understand the relationships between multiple datasets in 3D space.

A graph is a data structure comprised of entities (nodes) and connections between them (edges). Nodes can have attributes like values and labels, while edges can be directional or bidirectional, and weighted or unweighted. Graphs are underutilized in geosciences due to a lack of inherently graph structured data and relatively recent development of Graph Neural Networks. Our work shows the potential for graphs and graph neural networks to be applied to geoscience problems, as well as a general workflow to successfully implement a graph structure on 3D geoscience datasets.

We focus on basement delineation over a detrital iron deposit of the Fortescue valley in the Pilbara. Detrital iron deposits are accumulations of detritus eroded from a primary deposit (Koodaideri Iron) and trapped in basement depressions. Thus, the paleo-basement surface is crucial to the formation of the deposits. We use two key datasets: an AEM laterally constrained inversion volume (LCI) and a basement surface. This surface is produced using interpretation of: AEM inversion data, drilling, and seismic data.

Graph Neural Networks (GNNs) are powerful neural network architectures designed for graph data. The central paradigm of GNNs is called message passing, where prediction is made by ‘passing’ the embedded state of a node to its neighbours. This means that classification is controlled by the edges, we define instead of the attributes of the nodes. Further this allows us to connect nodes from different datasets that do not have common attributes, but that occupy the same space.

The nodes of our network come from the AEM LCI point cloud, where conductivity and uncertainty are discrete points in 3D space. We use the basement surface to label our nodes as either cover (above) or basement (below). The edges of the graph are generated using an adjacency matrix calculated by thresholding a cosine-similarity matrix, based on the z-score of the similarity value. Cosine similarity is a measure used to compare any two vectors. These vectors are generated based on relationships we consider important within our data such as spatial relationships (X, Y, Z), or conductivity depth relationships (COND, Z). We can combine and weight these matrices using the dot product for any number of similarity matrices.

Our preliminary results predicted with 93.5% accuracy if an inversion point was basement or cover in the 4200-point line using only 200 labelled points. The points that were misclassified, were mainly along the basement interface where the greatest uncertainty of label quality exists, based on conflicting interpretations. This represents a preliminary examination of our research problem and future work includes increasing the number of survey lines in the graph, different labelling regimes, the integration of additional datasets (drilling), and the propagation of more challenging labels (lithology).

**Biography**

Tasman is a PhD candidate within the Center of Exploration Targeting at UWA. He is studying the integration of Machine Learning and Geophysics under Professor Eun-Jung Holden, Dr Tom Horrocks and Dr Daniel Wedge.