HPC / Machine Learning 2021

UncoverML Geoscience Pipeline

Deployed an ML pipeline on Australia's Gadi supercomputer that reduced continental-scale geological analysis from 3 weeks to 11 hours, distributing workloads across 480 cores and 20 compute nodes.

Client

Geoscience Australia

UncoverML Geoscience Pipeline - Data analytics visualization

Key Results

480

Cores Utilized

3wk→11hr

Analysis Time

6 Teams

Using Nationally

2.8K

Lines Migrated to Py3

The Challenge

Geoscience Australia's researchers had continental-scale raster datasets (100+ GB per analysis) but no way to run ML models on them at scale. Single-machine runs took 3 weeks and frequently crashed, blocking mineral exploration and geological mapping projects.

Key challenges included:

  • 100+ GB raster datasets crashing on single-machine runs after days of processing
  • NCI Gadi environment requiring PBS job scheduling and MPI configuration
  • No existing tooling to distribute geoscience ML across 20+ compute nodes
  • Researchers needing results integrated with existing GDAL/Rasterio workflows
  • Legacy Python 2 codebase (2,800 lines) blocking compatibility with modern ML libraries

National Computational Infrastructure

NCI's Gadi supercomputer delivers 9 petaflops of computing power to Australian researchers. Making geoscience ML workflows run efficiently on Gadi unlocked analyses that were simply impossible on desktop hardware.

Gadi Supercomputer

Multi-petaflop computing power

Massive Datasets

Petabytes of geoscience data

Our Solution

A continental-scale analysis that used to take 3 weeks now completes in 11 hours across 20 nodes. Six research teams use the pipeline nationally, and the Python 3 migration unlocked access to modern scikit-learn, PyTorch, and XGBoost models.

MPI Distribution

Workloads distributed across 20 nodes and 480 cores via MPI, achieving near-linear scaling for feature extraction.

Feature Extraction

Scalable raster pipeline processing 100+ GB datasets in parallel, extracting 47 geophysical features per grid cell.

Hyperparameter Optimization

Grid search across 480 cores evaluated 1,200+ parameter combinations in hours instead of weeks.

Prediction Mapping

Continental-scale prediction maps generated in 11 hours, producing GeoTIFF outputs compatible with standard GIS tools.

Python 3 Migration

Full port of 2,800 lines from Python 2 to 3, unlocking compatibility with modern ML libraries and long-term support.

scikit-learn Integration

Pluggable model interface supporting Random Forest, Gradient Boosting, and SVR, selectable via config file.

Project Impact

Research Acceleration

  • Continental-scale analysis reduced from 3 weeks to 11 hours
  • Enabled 100+ GB analyses that previously crashed on desktop hardware
  • Adopted by 6 research teams across Geoscience Australia and universities
  • Directly supports mineral prospectivity mapping and geological surveys

Open Source Contribution

  • All improvements contributed upstream to the open-source UncoverML project
  • Python 3 migration adopted by the broader research community
  • Full documentation and PBS job templates enabling reproducible analyses
  • Pipeline now used as the foundation for 3 published geoscience papers