Trans-dimensional gravity and magnetic inversion
Billion-parameter Bayesian inversion of potential-field data on a single consumer GPU, from one vendor-portable source tree.
This page is a stub. Fuller writeup, figures, and code links to follow.
Potential-field inversion is badly non-unique: many subsurface density and susceptibility distributions produce the same measurements at the surface. The usual response is to fix a parameterisation in advance and regularise toward it, which produces a tidy answer and hides the question of how much of that answer the data actually supports.
Trans-dimensional Bayesian inversion treats the number of parameters as unknown and infers it alongside everything else, so the resolution of the result is a posterior quantity rather than a modelling choice. The cost is sampling, and the sampling is expensive.
The computational side of this is the part I can say most about right now. Kernels are written once in Julia against KernelAbstractions.jl and execute unchanged on NVIDIA, AMD, and Intel hardware, with the forward operator verified to produce identical results across all three. That makes billion-parameter (1024³) inference tractable on a single consumer card rather than a cluster allocation.
Presented at the DOE CSGF Annual Program Review 2026, and at AGU 2022–2024.
No comments here, by design. If you have something to say — a correction, a different edition of the same source, results of your own — I would like to read it: hello@bmoyer.net.