Building uncertainty-aware subsurface models with 3D magnetotelluric inversion

Magnetotellurics (MT) images electrical conductivity, a property influenced by fluids, alteration, graphite, and partial melt. Because many different subsurface configurations can explain the MT responses observed in the field, the preferred subsurface conductivity model estimated through inversion is often ambiguous. A central challenge is that a robust stochastic model exploration in 3D, while conceptually the best way to expose such ambiguity, is too computationally demanding to be used as a standard practice. Consequently, most studies report a single deterministic model, which obscures the range of alternative geological structures that are equally consistent with the data.

In this work we propose a practical approach to stochastic model exploration in 3D MT inversion. Drawing inspiration from annealing methods, we adopt a Very Fast Simulated Annealing (VFSA) framework, with sparse parameterization that makes the optimisation feasible for large-scale problems. Our inversion algorithm is built around the widely used ModEM forward solver, ensuring compatibility with existing workflows. Instead of producing one definitive model, the workflow can generate a set of plausible models that explain the data equally well. From this ensemble one can compute statistical summaries: a mean model that captures the most consistent structures and quantitative measures of variability that highlight where geometry, depth, or connectivity remain uncertain. This representation enables geologists to make interpretations while being explicitly aware of the uncertainty inherent in the inversion.  We demonstrate that the approach works at regional scale using a large-area subset of USArray MT data from Cascadia. This dataset has been extensively studied and previously modelled with deterministic 3D inversion, which allows us to benchmark our results.