Dual-Uncertainty Hybrid Inversion (DUHI): A Physics-Constrained, Uncertainty-Aware Framework for Magnetotelluric Inversion — Reference Implementation, Benchmark, and Field Validation Data

Le résumé fourni par la source

Magnetotelluric (MT) inversion is nonlinear and non-unique, and field data further complicate the problem through heterogeneous noise, static shift, and local violations of the two-dimensional assumption used in profile-based inversion. A learned inverse mapping can be fast and locally accurate, but a point estimate alone carries no independent measure of physical consistency and can be confidently wrong outside its training distribution. Dual-Uncertainty Hybrid Inversion (DUHI) addresses this by separating two distinct uncertainty sources—station- and frequency-dependent observation reliability in data space and predictive uncertainty from a deep ensemble in model space—and combining them inside a full electromagnetic inversion, rather than averaging independently produced models. The result is not a simple blend of a classical and a learned model: reliable observations and confident, well-supported learned structure are allowed to guide the solution, while unreliable data and uncertain learned predictions are overridden by the forward physics. On a frozen synthetic benchmark (six in-distribution and six structurally out-of-distribution geological families, n = 600 realizations each), DUHI does not exceed a direct learned point estimate in raw point accuracy, but it is the only configuration that is simultaneously reasonably accurate and independently verifiable against the physical data — substantially and consistently improving on conventional Occam inversion and AI-initialized-but-unconstrained inversion whenever the data contain heterogeneous observational corruption, while that advantage collapses to statistical indistinguishability in the noise-free limit. This confirms the benefit is specifically a consequence of heterogeneous data unreliability rather than a universal property of the formulation. A residual accuracy gap to the direct learned estimate is traced to a specific, mechanistic limitation of the classical solver—a single global roughness-misfit trade-off that cannot adapt locally—and a first, solver-compatible fix is shown to narrow it substantially without closing it entirely. Applied to real audio-magnetotelluric field data from the Baohuashan Cu-Mo prospect (profiles L26 and L30), DUHI reduces the data misfit substantially relative to conventional inversion while remaining within a markedly narrower, more geologically plausible resistivity range than either the AI-initialized-but-unconstrained or the conventional inversion. A formal, synthetically calibrated acceptance gate is also introduced and applied to both field lines, illustrating that verifiability is a property that must be actively checked against physical criteria rather than assumed from a plausible-looking image.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.