Combining Machine Learning and Geophysical Inversion for Applied Geophysics

Abstract

Machine learning and geophysical inversion both represent ways that the applied geophysicist might gain knowledge from field observations and remote sensed data. The two approaches represent contrasting philosophies based respectively on statistics and physics. Both potentially add insights which might help constrain 3D geology by geophysical means. Machine learning uses patterns in data to provide statistically controlled predictions, e.g. of lithology. In contrast, geophysical inversion relies on modelling the physical response of 3D geological block geometry in a deterministic manner. Although both approaches are widely used, it is not currently commonplace in applied geosciences to make use of a combined approach. We present an example which aims to refine the 3D geology in a prospective region of west Tasmania. Although the region is geologically well-mapped, thick vegetation and significant topography present a challenging set of conditions under which to refine the lithology and block geometry to a level of detail which will support the next generation of exploration. We use multiple layers of remote sensed geophysical data to provide probabilistic information on near-surface lithology extent using the Random Forests classifier. We show how the statistical, robust, output from the machine learning exercise can be used to guide the construction of improved volume geometry within a 3D GOCAD geological and geophysical modelling environment. This enables better constraints to be supplied to the geophysical inversion with resulting improvements in the detail of the 3D geology.