Spatiotemporal prospectivity modelling of porphyry mineralisation on a global scale

This repository contains STAMP (Spatio-Temporal Analysis of Mineral Prospectivity), the open-source workflow used to build a global, full-Phanerozoic model of porphyry copper prospectivity that couples a machine learning formation probability model with a preservation score derived from landscape evolution modelling. It accompanies the paper:

Farahbakhsh, E., McInnes, B. I. A., Kohlmann, F., Seton, M., Dutkiewicz, A., Müller, R. D. (2026). Global porphyry copper prospectivity through the Phanerozoic from interpretable machine learning coupling formation and preservation. [Under review]

The workflow reconstructs subduction zone kinematics and downgoing plate properties through deep time and trains an interpretable classifier that combines positive–unlabelled bagging with a random forest. Across the global arc network and at one-million-year resolution, it estimates the probability that porphyry mineralisation formed, the uncertainty of that prediction, and a preservation score expressing how favourable a location is for a deposit, once formed, to have survived subsequent erosion and to be discoverable, combining these into a map of preserved mineralisation prospectivity. See the paper for full details.

The current workflow is compatible with pyGPlates v1.0.0 and GPlately v2.0.0 and is designed to work with any plate reconstruction model available in the Plate Model Manager.

Workflow overview

STAMP is organised as six Jupyter notebooks, run in order, supported by a set of Python modules in the lib/ package. The pipeline proceeds from feature extraction and reconstruction, through feature analysis and predictive modelling, to prospectivity mapping and preservation assessment.

Notebook	Purpose
01_feature_extraction.ipynb	Sample trench points along the reconstructed subduction zones at each time step and extract the kinematic, downgoing plate (oceanic grid), and thermodynamically modelled slab devolatilisation features.
02_reconstruction_coregistration.ipynb	Build the one-sided trench buffer zones, reconstruct the known porphyry occurrences to their palaeo-positions, generate the unlabelled and target point sets, and coregister each point to its nearest trench point.
03_feature_analysis.ipynb	Prepare and downsample the data, examine the pairwise correlation structure and hierarchical clustering of the features, select a parsimonious predictor set guided by the resulting clusters, and generate provisional labels through positive–unlabelled bagging.
04_predictive_modelling.ipynb	Train the downstream random forest, evaluate it (ROC/AUC, calibration), interpret it (Gini and permutation importance, partial dependence and ICE, SHAP), and compute the mineralisation probability, uncertainty, and adjusted probability through time.
05_prospectivity_map.ipynb	Fill the present-day continental landmasses with a regular grid, reconstruct the grid through time within the trench buffer zones, predict the probability at each point, and produce the present-day global prospectivity maps with prediction–area (P–A) evaluation.
06_preservation.ipynb	Coregister the points with the cumulative erosion grids, fit the preservation score by case-control logistic regression of known deposits against the unlabelled background on a cubic B-spline basis in log-transformed cumulative erosion, evaluate it on held-out deposits, and combine it with the formation probability to map the preserved mineralisation prospectivity.

The lib/ package

Module	Role
feature_extraction.py	Trench tessellation and convergence kinematics, coregistration of oceanic grids, carbon, slab flux and cumulative subducted thickness calculations, and thermodynamic slab devolatilisation (H2O and CO2 outflux) and arc magma flux.
reconstruction_coregistration.py	Trench buffer zones, reconstruction of the deposit, unlabelled and target point sets, nearest trench coregistration, direct sampling of crustal thickness, and imputation of missing values.
slab_dip.py	Wrapper around the slabdip predictor that returns slab dip and the associated arc–trench distance.
water_thickness.py	Downgoing plate water inventory, partitioned into sedimentary and crustal pore and bound water, mantle lithosphere hydration, and hydration at the base of the lithosphere.
feature_analysis.py	Downsampling with age and spatial balancing, and correlation analysis and reporting.
predictive_modelling.py	ROC/AUC plotting, entropy and tree-vote-variance uncertainty measures, and gridding and curve utilities.
prospectivity_map.py	Reconstruction of continental grid nodes filtered to the trench buffer zones through time.
preservation.py	Coregistration of points with the cumulative erosion grids, and the PreservationScore estimator, which fits the spline and logistic regression pipeline, rescales the linear predictor to the unit interval on a shared scale, and reports held-out discrimination.
plot.py	Map and animation plotting (GPlately/Cartopy), including deposit overlays.

The lib/ folder must sit in the same directory as the notebooks, since each notebook imports from it (for example, from lib.feature_extraction import *).

Installation

We recommend a dedicated conda environment:

conda create -n stamp python=3.12 pip git notebook
conda activate stamp
pip install git+https://github.com/pulearn/pulearn.git@master
pip install gplately
pip install slabdip
pip install git+https://github.com/brmather/melt.git@main
pip install ipywidgets cmcrameri moviepy rioxarray scikit-image seaborn scikit-optimize shap

Cartopy, NumPy, pandas, scikit-learn, SciPy, xarray, GeoPandas, and rasterio are installed as dependencies of the packages above.

Configuration

Runtime settings are read from a parameters.py file in the repository root, which each notebook imports as from parameters import parameters. The parameters dictionary defines, among others:

• plate_model and anchor_plate_id: the Plate Model Manager model name and the anchor plate ID.
• timespan (min, max) and temporal_resolution: the reconstruction window and step, in Ma.
• grid_resolution: the spacing of the target and continental grid points, in degrees.
• inputs_dir and outputs_dir: the input data and output directories.
• The oceanic grid, slab H₂O and CO₂, and Perple_X lookup table sub-directories, together with output filenames.

Adjust these to point at your plate model and input data before running the notebooks.

Input data

Beyond the plate reconstruction, which is fetched automatically through the Plate Model Manager, the workflow expects the following inputs under inputs_dir (sub-directory names follow those referenced in the notebooks):

• Oceanic grids reconstructed on the chosen plate model: seafloor age (SeafloorAge), spreading rate (SpreadingRate), total deep-sea sediment thickness (SedimentThickness), and upper-oceanic-crust carbon (CrustalCO2).
• Perple_X lookup tables for slab devolatilisation, organised by reservoir (Sediments, Metabasalts, Intrusives, Sublithospheric_Oceanic_mantle), each with H₂O and CO₂ tables.
• Crustal thickness grids, used only by the secondary model.
• Cumulative erosion grids from a deep-time landscape evolution model, named cumulative_erosion_{time}Ma.nc, used to estimate the preservation score.
• A global porphyry copper deposit database, providing the positive (training) samples.

Running the workflow

1. Prepare the environment and parameters.py as above, and place the input data under inputs_dir.
2. Run the notebooks in numerical order, 01 through 06. Each notebook writes its outputs to outputs_dir, where subsequent notebooks pick them up.
3. Present-day and time-dependent prospectivity maps, the preservation-conditioned maps, and the accompanying animations are produced by notebooks 04 to 06.

Citation

If you use this workflow, please cite the accompanying paper:

@article{Farahbakhsh2026,
  author  = {Farahbakhsh, E. and McInnes, B. I. A. and Kohlmann, F. and
             Seton, M. and Dutkiewicz, A. and M\"uller, R. D.},
  title   = {Global porphyry copper prospectivity through the {Phanerozoic}
             from interpretable machine learning coupling formation and
             preservation},
  journal = {???},
  year    = {2026},
  note    = {Under review}
}

Please also cite the earlier study in which the spatio-temporal prospectivity modelling approach underlying STAMP was developed and applied regionally:

@article{Farahbakhsh2025,
  author  = {Farahbakhsh, E. and Zahirovic, S. and McInnes, B. and
             Polanco, S. and Kohlmann, F. and Seton, M. and M\"uller, R. D.},
  title   = {Machine learning-based spatio-temporal prospectivity modeling of
             porphyry systems in the {New Guinea} and {Solomon Islands} region},
  journal = {Tectonics},
  volume  = {44},
  pages   = {e2024TC008362},
  year    = {2025},
  doi     = {10.1029/2024TC008362}
}

Contact

Ehsan Farahbakhsh — e.farahbakhsh@sydney.edu.au or ehsan.farahbakhsh@anu.edu.au