A 2026 arXiv preprint reports a continually learning neural-operator surrogate that inverted 2,131,667 in-range airborne electromagnetic soundings in 24.9 seconds on one GPU. Posterior similarity was tested separately on matched comparisons involving 15 soundings, not across the survey-scale application.
A laboratory study found that growing and decaying rupture pulses approximately followed a common size–speed relation predicted by a two-dimensional friction theory.
GeoFormer, a geometry-aware Transformer for prestack seismic data, recorded the lowest reported RMSE on all four field datasets, but the study's noise examples were qualitative.
A numerical model of subduction initiation showed heat and uplift associated with an arc-like structure, even though no clearly developed cold slab appeared during the modeled period.