A preprint posted on arXiv on 9 October by statistics researchers at Ohio State University and the University of Wisconsin–Madison aims to scale inference for spatial fields modeled with Gaussian processes but non-Gaussian likelihoods. The proposed framework combines a sparse nearest-neighbor autoregressive structure with normalizing-flow conditionals and amortizes variational parameters through a shared neural network. This lets the same learned structure serve many observations instead of fitting separate parameters for each one.

The authors also demonstrated the method on binary ice-coverage observations from NOAA’s Arctic Ocean mapping data. For a subset of 21,855 observations, the amortized model produced posterior maps described in the paper as accurate using 1,398 variational parameters; the non-amortized comparison used about 200,000 parameters but did not converge reliably in initial runs. This is a methodological demonstration: its results do not measure forecast changes in ice or operational ice-monitoring accuracy. The large-data example alone also does not establish generalization to every region or date. The work is a preprint that has not yet undergone peer review.