Electromagnetic Subsurface Imaging of Continent-Scale Airborne Data Using Parallel Inference Algorithms

Much of Australia has been surveyed with low-flying airborne electromagnetic (AEM) instrumentation under Geoscience Australia’s AusAEM program. Acquired AEM data allow for imaging the earth's buried geology down to depths of 300-500 m. Such imaging is crucial for managing Australia’s subsurface minerals, energy and groundwater resources, by allowing geoscientists to build a 3D framework of the shallow geological architecture. However, individual AEM lines can be up to 500 km long, data are acquired every 10-12 m, and conventional electromagnetic conductivity imaging methods based on optimisation are unable to accurately characterise the subsurface imaging resolution. Bayesian probabilistic methods can do so, but at significant computational cost if naively used. Efficient Markov chain sampling strategies with parameter dimension reduction, which leverage the high-performance distributed computing capabilities inherent in the Julia language, have now made large scale Bayesian AEM imaging possible. In this work we show the results of imaging using the Julia-based, open-source, High Quality Geophysical Analysis (HiQGA) package, on continent-wide data using Bayesian probabilistic methods. We are unaware of any similar analysis at this scale, routinely using 41,600 cpu-cores for up to three hours in semi-embarrassingly parallel fashion on the National Computational Infrastructure’s Gadi cluster at the Australian National University. Consequently, deeper geology can be mapped, and subsurface 3D geology can be rapidly demarcated using posterior percentiles of conductivity, when contrasted with deterministic methods. Compared to the cost of AEM acquisition, extraction of subsurface information with computation at scale greatly increases the economic and social return on public AEM data acquisition.

Abstract presented at the 2024 Supercomputing Asia Conference, Sydney NSW (SAC2024)