Interpretation of multi-parameter crustal tomographic models with unsupervised machine learning

Interpretation of tomographic images of the continental crust is complicated by the non-linear interaction of different factors including composition, porosity, pore space geometry, pore pressure, pore fluid properties, temperature, and melt content. Multi-parameter interpretation can be used to mitigate some of the trade-offs between the above factors but can suffer from strong non-uniqueness. To reduce the complexity of the problem we subdivide the domain into classes using unsupervised machine learning. We explored the performance of different clustering algorithms including self-organized maps (SOMs), DBSCAN, fuzzy C-means, fuzzy SOM, fuzzy DBSCAN and HDBSCAN. For testing of the algorithms, we built 2D synthetic P-wave velocity (Vp), S-wave velocity (Vs), Vp/Vs, density, P-wave attenuation (Qp) and S-wave attenuation (Qs) models. We tested the robustness of the methods with respect to noisy, smooth and incomplete crustal models. We tested the performance with two common combinations of physical properties: (a) Vp, Vs, Vp/Vs and (b) Vp, Qp, density. The results of the clustering can then be used to reduce the degrees of freedom of rock physics interpretation and petrophysical inversion to obtain more robust estimates of porosity, temperature and melt fraction.