Probabilistic Deep Learning Inversion of AusAEM Data: Bayesian Resistivity Models and Trained Invertible Neural Networks

This dataset includes (1) continental-scale electrical resistivity models of Australia derived from AusAEM airborne electromagnetic (AEM) data using invertible neural networks (INNs), (2) all the training data (both synthetic 1D electrical conductivity models and the corresponding AEM responses based on AusAEM survey configurations), (3) PyTorch code for training the networks, (4) the final trained networks, and (5) example workflows. Main products Complete inversion results for 29,118,139 soundings in 3,584 line records across 23 survey blocks, including all publicly available AusAEM data as of April 2025 and supplementary results from the Frome survey area. Twelve final trained INNs: six INNs for TEMPEST measurements and six INNs for SkyTEM measurements, with model configurations and saved signal-standardization and height-normalization parameters. TEMPEST measurements are modeled using 1D Earth models with 31 layers, and SkyTEM measurements using models with 27 layers. Regional HDF5 files containing posterior means and population standard deviations of log10 resistivity, observed signals before z-score standardization, flight heights, coordinates, layer-top depths and depth of investigation (DOI). Supporting materials include a separate HDF5 file containing national depth-of-investigation estimates, synthetic training data in TXT format, geological province boundaries, and nine Jupyter notebooks demonstrating network training, inversion of synthetic and field AEM data, reading and extracting archived inversion results, and visualization of resistivity, uncertainty and DOI maps. Notebooks 03–09 include saved results; training is disabled by default. The DOI is the shallowest model layer-top depth at which the posterior standard deviation of log10 resistivity strictly exceeds 0.6. Posterior sample ensembles for the full national dataset are omitted to limit file size; users can generate posterior samples directly using the provided notebooks and saved, pretrained networks. Related article: Wu, S., Sun, J., and Chen, J. (2026). Continental-scale probabilistic resistivity imaging of Australia using deep learning: Implications for geology, groundwater, and critical minerals. arXiv:2609.19516. Method reference: Wu, S., Huang, Q., and Zhao, L. (2023). Fast Bayesian Inversion of Airborne Electromagnetic Data Based on the Invertible Neural Network. IEEE Transactions on Geoscience and Remote Sensing, 61, Article 5907211. doi:10.1109/TGRS.2023.3264777. Licensing and sources: Original synthetic data, model weights and inversion results are licensed under CC BY 4.0; original notebook code is licensed under MIT. Third-party observations and geometry are from processed datasets published by Geoscience Australia and retain their source licences. In particular, Capricorn observations and associated survey geometry retain CC BY-NC 4.0 and are excluded from the original-products CC BY 4.0 grant. Geological province boundaries are from Blake and Kilgour (1998), Geological Regions of Australia, 1:5,000,000 scale (Geoscience Australia), CC BY 4.0. README provides source citations and licence coverage for all 23 regional H5 files. Use: Download the 27 independent ZIP archives and both West ZIP parts (.001 and .002) for the complete dataset, or select the packages listed in the standalone README.md for a specific workflow. First join the two West parts in order to restore inversion_results_West.zip, then extract all selected ZIPs into the same parent directory to create one shared AusAEM_upload folder. The standalone README.md provides the current title, data descriptions, and download and joining instructions and supersedes the corresponding text inside code_and_networks.zip. ZIP_SHA256SUMS.txt provides SHA256 checksums for all downloadable archives, both parts and the standalone README. Corresponding author: Sihong Wu, [e-mail]; ORCID 0000-0002-5584-664X.