Confronting Plate Models with the Deep Mantle

Abstract

Mantle circulation models, numerical models of whole-mantle convection driven by tectonic reconstructions, are widely used to predict present-day mantle structure and, in turn, to evaluate the reconstructions themselves. Yet many studies compare modelled temperature fields directly with seismic tomography, overlooking nonlinearity in the temperature-velocity relationship, the strongly heterogeneous resolution of seismic imaging, and misfit metrics that saturate when anomalies do not overlap. This thesis develops a robust, physically consistent approach to evaluating competing tectonic reconstructions against tomography. I first develop a transferable comparison methodology, demonstrated with the G-ADOPT finite-element modelling library, comprising: (i) physically consistent conversion of predicted temperatures into seismic velocities; (ii) tomographic resolution operators to filter converted structures prior to comparison; and (iii) the Wasserstein metric to quantify misfit informatively even without anomaly overlap. After benchmarking G-ADOPT models across well-studied subduction regimes, I apply these tools to the collision of the Ontong-Java plateau with the Melanesian arc, where collision timing remains uncertain (~25 vs ~12 Ma).

I show that physically consistent post-processing can flip qualitative inferences of slab morphology, shift inferred slab depths by tens of kilometres, change inferred sinking rates by up to 1.2 mm/yr, and laterally displace slab anomalies by up to 639 km. Benchmarking shows G-ADOPT reproduces mantle structure well where slab rollback dominates, but predicts incorrect dip polarity where trenches advance. The case study strongly supports the older collision scenario: only the ~25 Ma model reproduces slab material beneath the Melanesian arc, with 10-28% lower misfit, supporting a role for the collision in regional plate reorganisation. These results highlight the value of publishing tomography models alongside resolution operators.